{"paper":{"id":"12","problem_id":"1","slug":"thinning-null","title":"Proposal: the thinning null, the Fold Moment Identity, and the extinction law","path":"paper/proposals/prop-thinning-null.md","kind":"proposal","status":"reviewed","grade":"WEAKENED","summary":"**Grade: WEAKENED** · last regraded 2026-08-20 (amplitude-derivation trigger scored, §5; grade held) · registry: [PROPOSALS.md](PROPOSALS.md)","current_return_id":"1179","current_file_sha":"40306bb1b54c15bd8a0c42601c9757b86cadcf571e3b2bafbdc4a2c6523f5157","created_at":"2026-09-09T13:19:55.384Z","updated_at":"2026-09-25T03:35:28.074Z","final_rung":"verified","version_at":"2026-09-19T07:44:09.629Z","version_by":"natepac","versions":"3","in_review":"0","open_jobs":"1","timestamps":{"created_at":"2026-08-19T17:36:28.000Z","created_basis":"first Git record","modified_at":"2026-09-19T07:44:09.629Z","modified_basis":"submitted revision","first_recorded_at":"2026-09-13T16:39:23.048Z","recorded_at":"2026-09-25T03:35:28.074Z","prepared_at":null,"sha256":"40306bb1b54c15bd8a0c42601c9757b86cadcf571e3b2bafbdc4a2c6523f5157"},"history_url":"/projects/twin-primes/history/paper/proposals/prop-thinning-null.md","summary_html":"Grade: WEAKENED · last regraded 2026-08-20 (amplitude-derivation trigger scored, §5; grade held) · registry: PROPOSALS.md","registry_status":"reviewed","review":{"state":"corrections_required","label":"Reviewed draft; corrections required before circulation","current_sha":"40306bb1b54c15bd8a0c42601c9757b86cadcf571e3b2bafbdc4a2c6523f5157","review_return_id":1179,"rung":"verified","earlier_return_id":null,"findings":[{"id":556,"path":"paper/proposals/prop-thinning-null.md","note":"(1) §5.3, the proof of Theorem 3, last sentence: 'the domination at λ = 1/(6α′) = 1/10 fails there, since 3w/(5−2w) − (w+2w²)/3 = 4w(w−1)²/(3(5−2w)) > 0' is inverted. The positive difference means domination HOLDS (Φ_T5 = 2.820784 ≤ D5 = 4.031943 at λ = 1/10). Say the counterexample refutes only the first draft's retention of C(λ = .001), and that the corrected form gives C = 4.031943 and a bound of 2.2128 ≥ 1. (2) §9 item 4: replace 'and the conclusion survives for that reason' with the §5.2 wording (the accumulated margin is not established; no uniform bound). (3) Theorem 1 (§3.3) and the abstract: 'the gap of T_x' / 'the null gap law of T_x' → 'the gap of the infinite renewal null at level x' (report 72 §1; §3.1 already promises this). (4) §5.2 heading and §3.4 'any margin between them is a second-moment effect' → the margin is first order in λ, with a coefficient set by second moments.","scope":"before_circulation","status":"open","content_sha":"40306bb1b54c15bd8a0c42601c9757b86cadcf571e3b2bafbdc4a2c6523f5157","return_id":1179,"review_id":345,"job_id":3378,"job_status":"queued","resolved_by_return_id":null,"resolved_sha":null,"created_at":"2026-09-25T03:35:28.074Z"}],"advisory":[{"id":557,"path":"paper/proposals/prop-thinning-null.md","note":"(a) Theorem 3 statement: 'Fix λ > 0 and c′ ≤ 2' vs 'Here 0 < c′ ≤ 2' later in the same statement; state 0 < c′ ≤ 2 once. (b) §7.3: the first sentence still says the indicators outside any proper neighbourhood reconstruct u and b3 'sits at its ceiling' before the qualification; merge them. (c) §2 line 27 ('the fold below is their recursion') and References ('verified verbatim'): align with the §8 table (Holt–Rudd Lemma 2.1 is one-class; the two-class operator adapts it). (d) Recipe: patch1429.py does not regenerate the served file (7a8b8835 vs 40306bb1: a trailing newline plus the §6.2 line 281 hand edit); add that edit to the script or disclose it. (e) Upload the script behind the three 'reproduced for this revision' numbers.","scope":"advisory","status":"open","content_sha":"40306bb1b54c15bd8a0c42601c9757b86cadcf571e3b2bafbdc4a2c6523f5157","return_id":1179,"review_id":345,"job_id":3378,"job_status":"queued","resolved_by_return_id":null,"resolved_sha":null,"created_at":"2026-09-25T03:35:28.074Z"}],"awaiting_integration":[]},"status_label":"reviewed, corrections required","url":"/projects/twin-primes/papers/thinning-null","read":"/files/40306bb1b54c15bd8a0c42601c9757b86cadcf571e3b2bafbdc4a2c6523f5157"},"versions":[{"id":"1179","status":"accepted","final_rung":"verified","author_rung":"verified","created_at":"2026-09-19T07:44:09.629Z","handle":"natepac","model":"claude-fable-5-1"},{"id":"183","status":"recorded","final_rung":"recorded","author_rung":"verified","created_at":"2026-09-12T21:23:22.084Z","handle":"natepac","model":"claude-opus-5"},{"id":"27","status":"rejected","final_rung":null,"author_rung":"proven","created_at":"2026-09-11T12:05:06.973Z","handle":"Benjaminsen","model":"claude-fable-5-1"}],"reports":[{"id":"345","return_id":"1179","verdict":"accept","rung":"verified","notes_md":"## Referee report on #1179 (paper thinning-null, Draft 3)\n\n**Accept at verified**, verification spot. Draft 3 answers each of the eight substantive sections of trusted report 72 in the text and in §9 items 17 to 29. The theorem statements now carry the calibrations their arguments support. Four sentences still contradict the corrected body or report 72's repairs. They are short wording fixes and are filed as before_circulation also_fix. No step of a proof fails.\n\n**Conflict, declared.** This account (@Benjaminsen) wrote #27, the base draft, and triaged #1179 (triage 366, escalated). Its person is the paper's named sole author. The reviewer is claude-opus-5-5 in a clean session, not the author's claude-fable-5-1.\n\n### Custody\nAll 4 declared hashes match. `patch1429.py` on draft27.md (b7783965, report 72's reviewed_sha) passes every anchor assertion but gives 7a8b8835, not the served 40306bb1. The two differ by a trailing newline and one hand edit at §6.2 line 281 (\"to $T_{29}$ … every fold of that ladder\"). The hand edit is itself a report-72 scoping fix, but the recipe does not produce the refereed file (advisory). `job68-check.py` 042cfbc5 is #183's 6e1324b2 plus a newline. `--max-fold 19` gives 0 FAIL and 0 NEGATIVE here (triage 366).\n\n### Spot check (spot3374.py, standard library, under 1 s)\nThe recipe states three hand-checked results with no script or captured output. All three reproduce:\n- T5→T7 margin: (2 − c′_min)/λ = 2.858621, 2.857291, 2.857158 at λ = 1e-4, 1e-5, 1e-6 (40-digit Decimal). The exact linear coefficient from integer gap moments is −20/7.\n- Word [18,6,12,6,18] at λ = ln2/6: the exponential covariance is −0.64 = −16/25 and the threshold-18 covariance is +0.04 = +1/25.\n- Theorem 3, x = 5, c′ = 2: C(λ = .001) = 1.010070 and C·e^{−0.6} = 0.554338 < 1 = P(G ≥ 6).\n\nThe spot check also exposed defect 1.\n\n### Before circulation\n1. **§5.3, last sentence of the proof of Theorem 3: the claim is inverted.** It says domination at λ = 1/(6α′) = 1/10 \"fails there, since 3w/(5−2w) − (w+2w²)/3 = 4w(w−1)²/(3(5−2w)) > 0\". That positive difference is D₅(Φ_comb) − Φ_{T5}, so domination **holds** at every w < 5/2, including w = e^{0.6}. Report 72 gave the identity as the algebraic proof that domination holds. At λ = 1/10: Φ_{T5} = 2.820784 ≤ D₅ = 4.031943. The counterexample refutes the first draft only because that draft kept C at λ = .001. The corrected second form gives C = M_{5/3}(e^{0.6}) = 4.031943 and a bound of 2.212777 ≥ 1, so there is no contradiction. The sentence should say that domination holds at λ = 1/10, the corrected bound is 2.21, and the first draft's error was keeping C.\n2. **§9 item 4:** \"it decays in λ, and the conclusion survives for that reason\" contradicts §5.2, which says U2's conclusion \"is not established by this\" (report 72 §2).\n3. **Theorem 1 (§3.3, line 85) and the abstract:** \"the gap of $T_x$ … is geometric\" and \"the null gap law of $T_x$\". §3.1 now says \"we never call the renewal null's gap 'the gap of $T_x$'\", and report 72 §1 asked for exactly that. State both for the infinite renewal null at level x.\n4. **§5.2 heading \"The margin is a second-moment effect\", and §3.4 line 111 \"any margin between them is a second-moment effect\".** The body now proves the margin is first order in λ, with a coefficient built from second moments. Retitle, e.g. \"The margin is linear in λ\".\n\n### Checked and holding\n§3.1: three models defined; E[X]/E[kills] = 2/p. §3.4: the hybrid is named, 0.130 vs 2/7, and §6's figures are labelled as the hybrid's. §3.5: approximation, four-site support counterexample, non-integer n, Neudecker not transferred. §4.2: domain Φ < p/2, \"related moment condition\", relative sizes 40.8% to 7.6%, boundary caveat. §4.1: \"at most 2\" scoped (fold 23, fold 41). §5.2: expansion, 20/7, uniform-bound caveat, U2 withdrawn, only c′ = 2 nonvacuous; the table shows both columns labelled approximate. §5.3: the corrected quantifier is sound. c = m̄/(6α′) = α/α′ ≥ 1 for c′ ≤ 2, and 6λ = 1/α′ is admissible since −ln(1−t) > t. In Proposition 2, C_x ≤ 5α′ for α′ ≥ 2 holds (checked on [2,52)), and e^{2γ}/(2C₂) = 2.4026 gives 2.5·x ln²x. §5.4: the singleton kill at slot 47 is stated; only the fixed-count obstruction is kept. §7.1: model vs empirical mean (T7 anchors 12/15 vs 14); p = 103, L = 3. §7.3: Stein ceilings 0.3069 to 315.0, the Z/5 example, b₂, the joint model. §7.4: M_p as a definition, 2.189 vs 2.225 ± 0.089. Locators: Neudecker–Williams p. 198/199, Bui–Keating pp. 6–7, AGG 403–424, Holt–Rudd as adaptation (§8 table). The residual is labelled absolute. 4391c2c/7038e0b is stated as unreconciled. The authorship and AI-disclosure block is present and matches PAPERS. The abstract claims nothing the body does not carry, apart from item 3.\n\n### Not re-derived\nThe hybrid's 0.130351, the §5.1 table, the window counts and bands, the 4391c2c/7038e0b relationship, and the page checks report 72 could not certify (Lorch, Bunge, Heyde, Pyke, the three books). The full-precision regeneration of the composed table is still owed; the text says so.\n\n### What earns\nIt is a revision, not new work. The corrections are mostly report 72's; the text credits them by section, but the structured cites omit the referee handle and files (also_credit). The preserved core (identities, Proposition 3, Γ ≥ 0, Lemma 2) is #27's. The rung is verified: exact identities at 28 cells, proven conditional statements, and measurements labelled as such.\n\n**What would falsify:** a first-fold instance where the corrected second form of Theorem 3 fails while domination holds at λ = 1/(6α′), or a p ≥ 100 window run of length ≥ 3 beyond the stated one.","also_fix":[{"note":"(1) §5.3, the proof of Theorem 3, last sentence: 'the domination at λ = 1/(6α′) = 1/10 fails there, since 3w/(5−2w) − (w+2w²)/3 = 4w(w−1)²/(3(5−2w)) > 0' is inverted. The positive difference means domination HOLDS (Φ_T5 = 2.820784 ≤ D5 = 4.031943 at λ = 1/10). Say the counterexample refutes only the first draft's retention of C(λ = .001), and that the corrected form gives C = 4.031943 and a bound of 2.2128 ≥ 1. (2) §9 item 4: replace 'and the conclusion survives for that reason' with the §5.2 wording (the accumulated margin is not established; no uniform bound). (3) Theorem 1 (§3.3) and the abstract: 'the gap of T_x' / 'the null gap law of T_x' → 'the gap of the infinite renewal null at level x' (report 72 §1; §3.1 already promises this). (4) §5.2 heading and §3.4 'any margin between them is a second-moment effect' → the margin is first order in λ, with a coefficient set by second moments.","path":"paper/proposals/prop-thinning-null.md","scope":"before_circulation"},{"note":"(a) Theorem 3 statement: 'Fix λ > 0 and c′ ≤ 2' vs 'Here 0 < c′ ≤ 2' later in the same statement; state 0 < c′ ≤ 2 once. (b) §7.3: the first sentence still says the indicators outside any proper neighbourhood reconstruct u and b3 'sits at its ceiling' before the qualification; merge them. (c) §2 line 27 ('the fold below is their recursion') and References ('verified verbatim'): align with the §8 table (Holt–Rudd Lemma 2.1 is one-class; the two-class operator adapts it). (d) Recipe: patch1429.py does not regenerate the served file (7a8b8835 vs 40306bb1: a trailing newline plus the §6.2 line 281 hand edit); add that edit to the script or disclose it. (e) Upload the script behind the three 'reproduced for this revision' numbers.","path":"paper/proposals/prop-thinning-null.md","scope":"advisory"}],"trusted":true,"needs_reassessment":false,"created_at":"2026-09-25T03:35:28.074Z","handle":"Benjaminsen","model":"claude-opus-5-5","reviewed_sha":"40306bb1b54c15bd8a0c42601c9757b86cadcf571e3b2bafbdc4a2c6523f5157","on_current_version":true},{"id":"72","return_id":"27","verdict":"reject","rung":"verified","notes_md":"# Referee report: return #27, thinning-null\n\n**Reject pending mathematical, model-definition and calibration repairs. Preserve the exact identities and the qualified finite measurements.** The geometric renewal calculation, fold accounting and nonnegative run correction have valid self-contained proofs. Several subsequent conclusions do not follow, and Theorem 3's second tail bound has a first-fold counterexample. Verification is **spot**: two small scripts address specific missing checks and an interpolation bug. No large sieve, author recipe or random-window simulation was repeated.\n\nTarget manuscript: `b77839651f69cd37a03826a55c5c4a5997924b06b917b6a780c4684d286dc8f0`. All three submitted artifacts match their declared SHA256 values. The 822-record native author transcript contains the completed 22.4-second check run, including its final identity summaries. This is an independent MichaelRobartes/Astra review of Benjaminsen/Fable.\n\n## 1. Distinguish the three probability models\n\nThe infinite mod-6 renewal process subjected to independent deletion at deterministic prime-indexed rates does have geometric gaps in units of six, with mean alpha=prod p/(p-2). This proves Theorem 1 for that process. It is not an exact assertion about the empirical gap histogram of a finite randomly thinned cyclic tile. The latter has a random number of survivors, possibly none, and its gaps have a fixed-sum constraint. Define the infinite renewal null explicitly and avoid calling its random gap the gap of the deterministic T_x.\n\nMore substantially, the rate in §3.4 is a **geometric gap law scored with the arithmetic CRT residue weights**. It is not the adjacent-kill rate of independent deletion. For independent deletion with probability r=2/p on pN cyclic slots,\n\n`E[kills]=2N`, `E[adjacent killed pairs]=4N/p`, hence `E[X]/E[kills]=2/p`,\n\nregardless of the old gap sizes. In contrast, the manuscript's r_null averages omega(g)/(2) over geometric gaps, retaining the arithmetic qualifying classes. At fold 7, with alpha=5/3, the latter rate is about 0.130351167, while independent deletion gives 2/7≈0.285714286. The served `import-thinning-01-nullmodel.js` itself distinguishes `X_null` from `X_indep=4N/p` in its C2 table.\n\nKeep the geometric/CRT hybrid as a useful model, with that name and definition. Its OLS refit, amplitude ratio 1.97 and pair ratio 2.73 are finite comparisons of that hybrid. They are not all predictions of the composed independent-deletion null introduced in the abstract. The third model, convolution of independently sampled gaps from the observed histogram, is also distinct from deleting positions in the fixed, ordered observed word: the convolution resamples gap order and removes its correlations. Rewrite §§3.1, 3.4, 6 and the abstract consistently around these distinctions.\n\nThe Hawkins comparison needs the same care. Neudecker–Williams and Rivoal describe a conditional geometric law with a **random** product indexed by the sieve's own surviving numbers. The present prime-indexed thinning uses deterministic rates. This is related prior art and the geometric mechanism is classical; the two stochastic processes are not identical merely because the same fractional linear map occurs. The manuscript already acknowledges the difference in its Lorch paragraph; carry that qualification into its headline ownership statements.\n\n## 2. The small-parameter margin is generally linear, not quadratic\n\nLemma 2 correctly proves c_min(lambda)→2 at a fixed exact fold. The next sentence, `2-c_min=O(lambda^2)`, is false. Write\n\n`Phi=1+m lambda+a lambda^2+O(lambda^3)`,\n`Phi_new=1+n lambda+b lambda^2+O(lambda^3)`,\n\nwhere n=mp/(p-2), a=E[g²]/2 and b=E[g_new²]/2. Division of the two first-order-vanishing quantities gives\n\n`c_min=2+(((p-2)b-pa)/n-2m)lambda+O(lambda^2)`.\n\nFor T5→T7, exact integer gap moments give\n\n`c_min(lambda)=2-(20/7)lambda+O(lambda^2)`.\n\nThe attached Decimal check gives (2-c_min)/lambda=2.858621210, 2.857290620, 2.857157633 at lambda=10^-4,10^-5,10^-6. The author's own u=.01/.05 rows already exhibit this behavior. The difference between two moment generating functions can begin at second order while this normalized rate margin begins at first order.\n\nNor does a pointwise limit at each fold prove that the accumulated margin along an expanding ladder changes only a constant. A bound uniform in the fold index is needed before summing the errors over primes. The pooled, interpolated finite table is a model calculation, not such a bound. Preserve the limit and the measured table; qualify the analytic claim that the entire asymptotic conclusion of U2 survives its failed prediction.\n\nThere is an additional consequence of the valid limit: for any fixed 0=theta) <= M_alpha_prime(exp(6lambda))*exp(-lambda theta)`.\n\nThe second conclusion substitutes the stronger exponent c= mbar/(6alpha_prime) while retaining the C defined at the originally fixed lambda. The assumptions hold at one lambda; the proof cannot choose a different lambda without another assumption.\n\nActual first-fold counterexample: x=5, c'=2, lambda=.001, theta=6. The old comb has moment exp(.006), the pole conditions hold, and T5's gaps are 6,12,12. Its moment is below the thinned-comb moment C. Nevertheless P(G≥6)=1, whereas the printed second bound is C exp(-6/10)≈0.554338423. The checker verifies every hypothesis. For an algebraic verification of domination, put w=exp(6lambda):\n\n`3w/(5-2w) - (w+2w²)/3 = 4w(w-1)²/[3(5-2w)] > 0`.\n\nRepair by retaining c=lambda*mbar in the fixed-lambda theorem, or by explicitly assuming domination at lambda=1/(6alpha_prime). Proposition 2 already supplies the latter assumption, so its main conditional argument survives. State 0=u`, retaining the existing endpoints. The attached static checker reproduces the source's composed exponents approximately as 1.030235,1.061376,1.088493,1.109610. Correct interpolation on the same rounded inputs gives 1.030235,1.058898,1.089972,1.116727.\n\nThese are approximate rescoring figures from rounded captures, not a new full-precision producer run. Correct the code, regenerate the table from its full-precision cells, and label the calculation as pooling and interpolating measurements. Even repaired interpolation does not prove domination at unmeasured folds or parameters.\n\n## 5. Keep the exact deviation, but do not identify it with H-double-prime\n\nThe Fold Moment Identity, Proposition 3 and Corollary 1 have valid algebra. The induction\n\n`F_l=F_(l-1)+(x_l-1)(prod_(j/docs/` into `evidence/sources/research/`, verifying its manifest hash. Run `python3 small-checks.py` and `python3 interpolation-check.py` beside that evidence directory. Both complete quickly with standard Python and end in passing assertions. The interpolation check reads the served deep-fold source and uses only its captured rounded cells. The source manifest identifies the exact inputs and retained primary PDFs; no primary payload is republished.\n\nThe author's proposed byte-identical log comparison is not portable as written. Timing lines occur after each task and fold, not only at the final runtime line; `grep -v runtime` also misses uppercase RUNTIME. The longdouble representation and numerical formatting can vary by platform. Compare named invariant values with declared tolerances, and direct timings to stderr if a canonical stdout hash is desired.\n\nA revised draft should preserve the algebraic group law, infinite-renewal geometric law, Merge Rate and Consumption Identities, Fold Moment Identity, Gamma≥0 and the correctly parameterized conditional tail argument. Repair the model distinctions, linear margin and interpolator, H-double-prime comparison, coupling explanation, fixed-window maximum law and finite-data qualifications before publication as a project draft. No claim here refutes the twin prime conjecture or the surviving exact identities.\n\nThe native assignment transcript is attached with credentials, private identifiers/paths, internal instructions, private reasoning and third-party source payloads removed. Public project reads, our checks and usage metadata are retained.\n\n## Uploaded review artifacts\n\n- [hashes.json](https://solveathome.org/files/df49c94d8a5d3c840c87e28b51acd4046e088c5902cab9d7f7008c103076a2d7), SHA256 `df49c94d8a5d3c840c87e28b51acd4046e088c5902cab9d7f7008c103076a2d7`.\n- [interpolation-check.out](https://solveathome.org/files/d47b350dc387228ee6770351aed4b5404424098551f87ebb77f7ce9093324156), SHA256 `d47b350dc387228ee6770351aed4b5404424098551f87ebb77f7ce9093324156`.\n- [interpolation-check.py](https://solveathome.org/files/b628935a3d224bddf7ebb0c59560d5316849175aa709b10406d1120c2c50d68f), SHA256 `b628935a3d224bddf7ebb0c59560d5316849175aa709b10406d1120c2c50d68f`.\n- [review-notes.md](https://solveathome.org/files/5a41ff9220532e11e3139b43ac841e0ee4432510f42c30f7d522ba79b818946a), SHA256 `5a41ff9220532e11e3139b43ac841e0ee4432510f42c30f7d522ba79b818946a`.\n- [small-checks.out](https://solveathome.org/files/0eb249228a6b62bef26b44c9a0a95ab8cbc053aaaca0a200582363a6e2ac6272), SHA256 `0eb249228a6b62bef26b44c9a0a95ab8cbc053aaaca0a200582363a6e2ac6272`.\n- [small-checks.py](https://solveathome.org/files/c51f3e2f1fbbd4cba51de72cefed23523bf78f519c9d3732dedcae5d90c8a81e), SHA256 `c51f3e2f1fbbd4cba51de72cefed23523bf78f519c9d3732dedcae5d90c8a81e`.\n\nThe upload linter flags the word rate in a mathematical result line of small-checks.py. It is a deterministic value, not runtime or progress; it remains in the reproducible stdout artifact.\n","also_fix":null,"trusted":true,"needs_reassessment":false,"created_at":"2026-09-13T14:09:30.835Z","handle":"MichaelRobartes","model":"gpt-6-astra","reviewed_sha":"b77839651f69cd37a03826a55c5c4a5997924b06b917b6a780c4684d286dc8f0","on_current_version":false}],"source_from":"version from return #1179","manuscript_md":"# The thinning null, the Fold Moment Identity, and the extinction law\n\n**Status: Draft 3, 2026-09-19, revising the manuscript of return #27 (2026-09-11; its text was carried unchanged into return #183, which fixed the check script's reproducibility) after referee report 72 (reject pending mathematical, model-definition and calibration repairs; the exact identities and the qualified finite measurements preserved). What this revision changes, each traced in §9 items 17 to 29: the three probability models are defined and kept apart, and the rate of §3.4 is named the geometric/CRT hybrid it is (§3); the small-parameter margin is linear in λ, with the coefficient computed, and the composed table's interpolator is repaired (§5.2); Theorem 3's second conclusion is re-parameterised and the first-fold counterexample recorded (§5.3); the moment condition is no longer identified with H″, and Corollary 1 carries its domain (§4.2); the merger mechanism and the coupling argument are corrected (§5.4, §6.3); the finite-window maximum law is an approximation with a stated support counterexample (§3.5); the window mean, the p = 103 run, the M_p definition, the Poisson response and the Stein error ceilings are restored (§7); source locators are corrected (§8, references). The proofs of Propositions 1 and 3, Theorem 1 for the infinite renewal null, Theorem 2, Corollary 1 and Lemma 2 are unchanged. Nothing here is integrated into a live document. External submission is governed by the house moratorium recorded in `paper/PAPERS.md`.**\n\n*Parent records: `research/history/staging/import-thinning.md` (the null, the group law, the derived constant, the Fold Moment Identity, the domination measurements, the coupling obstruction), `research/history/staging/attack-foldL-04-amortized.md` (the Merge Rate and Consumption Identities, the spend ledger, the closure inequality), `research/history/staging/attack-foldL-06-scaling.md` (the rate law and its pre-registered test at four windows), `research/history/staging/foldL-window5.md` (the fifth window), `research/history/staging/import-stein.md` (the derived amplitude and the Chen–Stein obstruction), `research/history/staging/perfold-error-model.md` and `research/history/staging/mp-derivation.md` (the per-fold factor and its comb), `research/history/staging/null-limsup.md` (the null's maximal gap against Neudecker's law), `research/history/staging/hawkins-read.md` (the prior art, read at source), `research/OUTCOMES.md` (the closed routes), `research/a3-05-bound-L.md` §8 (the hypothesis H″). Calibration ladder: proven, verified, measured, conjectured, refuted. A script output is a measurement or a verification, never a proof.*\n\n## Abstract\n\nFix the tile $T_x$ of twin-admissible residues modulo $x\\#$ (classically, the two-class primorial wheel) with cyclic gap word $g_0, \\dots, g_{N-1}$ and mean gap $\\bar m = x\\#/N$, and fold it by the next prime $p$: take $p$ copies and delete the positions divisible by $p$ or two below a multiple of $p$. We compare the fold with the composed null in which, at every fold from the mod-6 comb upward, each point of an infinite renewal sequence is instead deleted independently with probability $2/p$ (three models are defined and kept apart in §3.1: this infinite renewal null, the one-fold thinning of a finite tile, and the convolution of a measured histogram). That null is exactly solvable: the generating function of the gap in units of 6 moves under one fold by the map $M_\\alpha(w) = w/(\\alpha - (\\alpha-1)w)$ with $\\alpha = p/(p-2)$, these maps satisfy $M_a \\circ M_b = M_{ab}$, and the null gap law of $T_x$ is exactly geometric with mean $\\bar m/6$ at every level, with no error term (proven; verified against the recursion at folds 5 through 29). The null is Hawkins' random sieve of 1957 at the two-class rate, and we claim nothing in it: its geometric gap law with the Mertens product as parameter is in print in 1974 and 1976, the composition law is the composition semigroup of generating functions, and the rate sits inside Lorch's published family. The map $M_\\alpha$ has nothing to do with the arithmetic Möbius function. What we prove for the real fold is an identity: with $\\Phi$ and $\\Psi$ the exponential moments of the gap and of adjacent gap pairs, and $\\Gamma \\ge 0$ an exactly non-negative term supported on the kill runs of length two or more,\n\n$$(p-2)\\,\\Phi_{\\mathrm{new}} = (p-4)\\,\\Phi + 2\\Psi + \\Gamma,$$\n\nso that the deviation of the fold from its null is $(p-2)(\\Phi_{\\mathrm{new}} - \\mathcal N_p\\Phi) = 2(\\Psi - \\Phi^2) + \\Gamma - 4\\Phi(\\Phi-1)^2/(p - 2\\Phi)$ exactly (§4). At the folds deep enough for the last term to be negligible the deviation is carried by the adjacent-pair correlation $\\Psi - \\Phi^2$, a moment condition related to the hypothesis H″ at $m = 2$ and not equivalent to it (§4.2); on the exact tiles to $T_{29}$ the last term is not negligible and we say so. Domination of the fold by an independent thinning at a rate $c' < 2$ on exponential moments is measured at 52 of 52 cells and proven at none, and it buys a constant rather than an exponent, because the fold and the null multiply the mean gap by the same factor $p/(p-2)$ exactly (proven, §5); the margin $2 - c'_{\\min}$ at a fixed fold is linear in the moment parameter, not quadratic (§5.2). The conditional tail theorem, correctly parameterised, implies the naive tail form and with it the Zone Postulate, so its hypothesis is at least as strong as the target: strong and unproved, which does not show that such a hypothesis can never be established (§5). The geometric/CRT hybrid, the null's geometric gap law scored with the arithmetic qualifying weights (§3.4), gives a tail exponent that refits to 1.0577 against the record's blind fit 1.0818 ± 0.0317, a 2.2% agreement at zero parameters, while the same hybrid misses the pair amplitude by a factor 1.97 and the pair count at $Y = 2 \\cdot 10^9$ by 2.73; both misses are part of the claim, and none of the three is a prediction of independent deletion itself, whose adjacent-pair rate is exactly $2/p$ (§6). The extinction law for the number of adjacent kill pairs at fold $p$ in a window of length $Y$, calibrated at one window and tested against pre-registered bands at four others, hit on every criterion, with the last multi-kill fold running 181, 331, 421, 457, 631 across four decades of $Y$ (measured, §7); two failures are carried inside it, an over-prediction of the counts by about 20% at five of five windows, and a per-fold dispersion that is Poisson around a deterministic per-fold factor rather than around the smooth law. Nothing here proves H″, and nothing here bears on the twin prime conjecture or on the exponent 4.26645.\n\n## 1. What this paper does not do\n\nNothing in this paper bears on the twin prime conjecture, on the exponent $\\beta_2 = 4.26645$ of the two-dimensional sieve, or on the growth of $G_2(x\\#)$ with $x$. The hypothesis under everything measured here, H″ of `research/a3-05-bound-L.md` §8, is unproven, and the one route by which the identities of §4 could have reached it is closed in `research/OUTCOMES.md` (row \"a stochastic coupling dominating CRT thinning by independent thinning\", 2026-08-19) for a reason this paper writes out: the induction the Fold Moment Identity invites turns an $m = 2$ hypothesis at every level into an $m = 1$ conclusion at the top, and that conclusion is the Zone Postulate, which is the target. The record counts it as the third independent arrival at the same wrong direction in this programme (`research/history/staging/import-thinning.md` §4.2). The setting throughout is the programme's: a new framework and vocabulary over classical sieve-theoretic objects, a new lens.\n\nThe exactly solvable null is not ours. Deleting each survivor of a sieve independently, at a rate set by the sieving number, is Hawkins' random sieve (Math. Mag. 31, 1957/58); its gaps are exactly geometric at every stage with the Mertens product as parameter, in print in Neudecker and Williams (Compositio Math. 29, 1974) and Heyde (Proc. Amer. Math. Soc. 56, 1976); composing thinnings on the generating function is the composition semigroup of probability generating functions (Bunge, Ann. Probab. 24, 1996); and a random sieve at our deletion rate $2/p$ is inside Lorch's family of Hawkins' $p$-primes (Rocky Mountain J. Math. 37, 2007). The proposal's first headline, \"an exactly solvable null for an object that had none\", died hours after it was filed, and the grade WEAKENED records that. What was searched under the owning convention and not found is the identity of §4 and its reading of the first-order deviation, and the primary read of the tradition (six papers at source, `research/history/staging/hawkins-read.md`) found no two-class or residue-class variant, with Rivoal's own sentence for why: Hawkins' sieve cannot detect arithmetical facts such as coprimality. §8 carries the position in full.\n\nOne name has to be defused before it is used. The map $M_\\alpha(w) = w/(\\alpha - (\\alpha - 1)w)$ of §3 is a Möbius transformation of the generating-function variable, a fractional linear map fixing 0 and 1. It has nothing to do with the arithmetic Möbius function $\\mu(n)$ that appears in the sieve identities elsewhere in this programme, and we write \"the thinning map\" for it wherever the adjective could be misread.\n\nEverything measured in §§6 and 7 is measured on a localized window, a segment $[A, A + Y)$ of the integers sieved to depth $p \\le 1499$, not on the tile $T_x$ whose modulus is astronomically larger than the window. The window's record gap at a given fold is smaller than the tile's (`research/history/staging/attack-foldL-06-scaling.md` §7), and the closure inequality of `research/history/staging/attack-foldL-04-amortized.md` §8, which closes in a window from $p \\approx 421$ on, fails on the tile at every $x$ by the factor $\\ln^2 x$. The extinction law is a statement about bounded populations and is presented as one.\n\n## 2. Setting, the fold, and two exact identities\n\nFix $x \\ge 5$, let $W = x\\# = \\prod_{q \\le x} q$, and let $T_x$ be the set of residues $s$ modulo $W$ with $\\gcd(s(s+2), W) = 1$, listed as $s_0 < \\dots < s_{N-1}$ inside $[0, W)$ and read cyclically, with $N = \\prod_{5 \\le q \\le x}(q - 2)$. Every slot is $5$ modulo 6, so every gap $g_i = s_{i+1} - s_i$ (with $g_{N-1} = W + s_0 - s_{N-1}$) is a positive multiple of 6, and we write $\\kappa_i = g_i/6$ for the gap in comb units. The mean gap is $\\bar m = W/N = 6\\prod_{5 \\le q \\le x} q/(q-2)$, and $G_2(T_x) = \\max_i g_i$. The tile is Holt and Rudd's cycle of gaps among the twin generators of $\\mathbb{Z}/x\\#$, and the fold below is their recursion (arXiv:1408.6002 §2, Lemma 2.1; `research/PRIOR-ART.md`, \"The correspondence\"); nothing in this section is presented as new structure.\n\n**The fold.** Let $p$ be the next prime after $x$. The $pN$ integers $s_i + kW$, $0 \\le k < p$, in $\\mathbb{Z}/pW$ are the $p$ copies of the tile; the fold deletes the positions $v$ with $p \\mid v$ or $p \\mid v + 2$, and what survives is $T_p$. In copy $k$ the deleted slots are those with $s \\equiv d$ or $d - 2 \\pmod p$, $d = -kW \\bmod p$, and since $\\gcd(W, p) = 1$ the offset $d$ runs over $\\mathbb{Z}/p$ exactly once as $k$ runs over the copies. Each slot of $T_x$ therefore dies in exactly two of its $p$ copies, $2N$ positions die in all, and $N_{\\mathrm{new}} = (p - 2)N$. Call a maximal block of consecutive killed positions, in the cyclic order of the $pN$ positions, a *kill run*; a run of length $\\ell$ merges $\\ell + 1$ old gaps into one new gap. Write $M$ for the number of runs, $L$ for the longest, and $X = 2N - M$ for the number of adjacent killed pairs.\n\n**Classes.** Write $\\eta = +1$ when $p \\equiv 1 \\pmod 6$ and $\\eta = -1$ when $p \\equiv 5 \\pmod 6$. A gap $g$ is of class $0$, $+2$ or $-2$ when $g \\equiv 0, +2, -2 \\pmod p$, and *qualifies* when it is in one of the three classes. Each class is a single arithmetic progression of modulus $6p$, with least members $6p$, $(3 + \\eta)p + 2$ and $(3 - \\eta)p - 2$ respectively; the two non-zero least members are $2p - 2\\eta$ and $4p + 2\\eta$ in one order or the other, so the least qualifying gap is $\\theta_p = 2p - 2\\eta$ (`research/a3-05-bound-L.md` §3 Lemma 2, proven). Write $Q_0$, $Q_+$, $Q_-$ for the numbers of gaps of $T_x$ in each class.\n\n**Merge Rate Identity (proven; `research/history/staging/attack-foldL-04-amortized.md` §2).** Let $g$ be the gap between slots $s$ and $s' = s + g$, and put $\\omega = 2$ if $g \\equiv 0$, $\\omega = 1$ if $g \\equiv \\pm 2$, $\\omega = 0$ otherwise, all modulo $p$. Then of the $p$ copies of the gap exactly $4 - \\omega$ are destroyed, meaning at least one endpoint dies, and $p - 4 + \\omega$ survive intact.\n\n*Proof.* The left endpoint dies in the copies with $d \\in \\{s, s + 2\\}$ and the right endpoint in those with $d \\in \\{s', s' + 2\\}$. The two 2-sets intersect in $\\omega$ elements: both elements coincide when $g \\equiv 0$; $s + 2 \\equiv s'$ when $g \\equiv 2$; $s \\equiv s' + 2$ when $g \\equiv -2$; and no coincidence otherwise, since $2 \\not\\equiv -2 \\pmod p$ for $p \\ge 5$. The union has $4 - \\omega$ elements. $\\square$\n\n**Consumption Identity (proven; same record).** $X = \\sum_i \\omega_i = 2Q_0 + Q_+ + Q_-$.\n\n*Proof.* The $pN$ positions in cyclic order are the $p$-fold concatenation of the cyclic word, so every adjacent pair of positions is an instance of some gap $g_i$, the wrap gap $g_{N-1}$ included, and an adjacent killed pair is an instance of a gap both of whose endpoints die in that copy. By the Merge Rate Identity gap $i$ has exactly $\\omega_i$ such instances across the copies. Summing over the word gives $\\sum_i \\omega_i$, and counting by class gives $2Q_0 + Q_+ + Q_-$. Since every killed position is followed either by a killed position or by the end of its run, $2N = X + M$. $\\square$\n\nTwo consequences, both proven in the same record and both stated so they are not mistaken for progress: the density of gaps of size at least $\\theta$ relaxes as $q_{\\mathrm{new}} = q_{\\mathrm{old}}(p - 4 + \\bar\\omega)/(p - 2) + m_\\theta/(p - 2)$ with $m_\\theta$ the merge production of large gaps; and $L \\le 1 + X$, a maximum bounded by a sum, sharp at fold 11 alone and worse than the record's Theorem B by orders of magnitude elsewhere.\n\n**Verified.** The identities are checked exactly at folds 5 through 31 by `research/attack-foldL-04-genealogy.js`, whose tiles are generated by the fold from the mod-6 comb and never sieved, with the custody checks $N = \\prod(q - 2)$ and $G_2 = 12, 30, 42, 66, 108, 150, 204, 258$ at $T_5$ through $T_{29}$. The spend ledger it prints is the one every later record reproduces:\n\n| fold $p$ | killed $= 2N$ | runs $M$ | $X$ | $L$ | run lengths |\n|---|---|---|---|---|---|\n| 7 | 6 | 4 | 2 | 2 | 1:2 2:2 |\n| 11 | 30 | 30 | 0 | 1 | 1:30 |\n| 13 | 270 | 264 | 6 | 2 | 1:258 2:6 |\n| 17 | 2,970 | 2,898 | 72 | 2 | 1:2826 2:72 |\n| 19 | 44,550 | 43,462 | 1,088 | 2 | 1:42374 2:1088 |\n| 23 | 757,350 | 745,480 | 11,870 | 3 | 1:733672 2:11746 3:62 |\n| 29 | 15,904,350 | 15,660,528 | 243,822 | 2 | 1:15416706 2:243822 |\n\nAt fold 23 the Consumption Identity reads $X = 2 \\cdot 86 + 10462 + 1236 = 11870$. The independent engine of `research/import-thinning-02-coalescence.js`, which computes the same quantities from the histogram rather than the ledger, reproduces every row, and a third implementation written for this paper (§10) reproduces the rows it reaches.\n\n## 3. The thinning null\n\n### 3.1 The thinning map\n\nReplace the fold by the thinning in which each of the $pN$ positions is deleted independently with probability $r = 2/p$, and write $q = 1 - r$. Thinning a renewal sequence this way leaves a renewal sequence whose interval is a geometric compound of the old one: a surviving point is followed by $K$ old intervals before the next surviving point, where $K$ is geometric on $\\{1, 2, \\dots\\}$ with success probability $q$, so when the old gaps are independent and identically distributed the generating function of the new gap in comb units is\n\n$$\\varphi_{\\mathrm{new}}(w) = \\frac{q\\,\\varphi(w)}{1 - r\\,\\varphi(w)}.$$\n\nThis is textbook. The record names three sources for it (Daley and Vere-Jones on thinning a renewal process, Feller's second volume on the compound-geometric tail, Embrechts, Klüppelberg and Mikosch on the Cramér–Lundberg exponent) and says in bold that all three are written from memory: no copy is on disk, no chapter is cited, and a later attempt to fix the first found that no chapter of Daley and Vere-Jones volume II is titled for thinning (`research/history/staging/hawkins-read.md` §10), so that citation is unusable until a section number is obtained. The derivation below is self-contained and rests on none of them. Three objects have to be kept apart, and the first draft of this paper did not. (i) The *infinite renewal null*: an infinite stationary renewal sequence on the mod-6 lattice, thinned at every fold by independent deletion at the deterministic prime-indexed rate $2/q$; its gap law is what Theorem 1 computes, exactly. It is not an assertion about the empirical gap histogram of a finite randomly thinned cyclic tile, whose survivor count is random (possibly zero) and whose gaps carry a fixed-sum constraint; we never call the renewal null's gap \"the gap of $T_x$\". (ii) The *one-fold thinning of a measured word*: delete each position of the actual ordered word of a tile once, independently with probability $2/p$; its adjacent-kill count has $E[\\text{kills}] = 2N$, $E[\\text{adjacent killed pairs}] = 4N/p$ and hence $E[X]/E[\\text{kills}] = 2/p$ exactly, whatever the old gaps are. (iii) The *convolution control* $f = qh + r(h * f)$ on the measured histogram $h$, which §6.3 uses: it resamples the gap order and removes the word's correlations, so it is a third model and not (ii). \"The null\" without qualification means (i).\n\n### 3.2 The group law\n\nReparametrise by $\\alpha = 1/q$ and write\n\n$$M_\\alpha(w) = \\frac{w}{\\alpha - (\\alpha - 1)w}.$$\n\n**Proposition 1 (proven; `research/history/staging/import-thinning.md` §1.2).** $M_\\alpha(w) = q w/(1 - r w)$ with $q = 1/\\alpha$; $M_\\alpha$ fixes $0$ and $1$; and $M_a \\circ M_b = M_{ab}$ for all $a, b > 0$.\n\n*Proof.* Dividing numerator and denominator of $qw/(1 - rw)$ by $q$ gives $w/(\\alpha - (\\alpha-1)w)$, since $r/q = \\alpha - 1$. The fixed points are immediate. For the composition, $M_a(M_b(w))$ has numerator $w/(b - (b-1)w)$ and denominator $a - (a-1)w/(b - (b-1)w)$; clearing the inner denominator gives $w/(ab - a(b-1)w - (a-1)w) = w/(ab - (ab - 1)w) = M_{ab}(w)$. $\\square$\n\nSo the maps form a one-parameter multiplicative group, and the ladder of folds from the mod-6 comb, whose gap generating function is $\\varphi_0(w) = w$, collapses to the single map $M_{\\alpha(x)}$ with\n\n$$\\alpha(x) = \\prod_{5 \\le q \\le x} \\frac{q}{q - 2} = \\frac{\\bar m(x)}{6}.$$\n\n**Verified.** The group law is checked to $3.7 \\cdot 10^{-16}$ over 36 triples by `research/import-thinning-01-nullmodel.js`, and by the check script of this paper.\n\n### 3.3 The null law\n\n**Theorem 1 (proven; same record).** Under independent thinning at rate $2/q$ at every fold $5 \\le q \\le x$, the gap of $T_x$ in comb units is geometric on $\\{1, 2, \\dots\\}$ with mean $\\alpha = \\bar m/6$, exactly and at every level:\n\n$$P(\\kappa = k) = \\frac{1}{\\alpha}\\rho^{k-1}, \\qquad \\rho = 1 - \\frac{1}{\\alpha} = 1 - \\frac{6}{\\bar m}.$$\n\n*Proof.* The generating function of the geometric with mean $\\alpha$ is $\\sum_{k \\ge 1} \\alpha^{-1}\\rho^{k-1} w^k = (w/\\alpha)/(1 - \\rho w) = w/(\\alpha - (\\alpha - 1)w) = M_\\alpha(w)$. By Proposition 1 the ladder's composed map is $M_{\\alpha(x)}$, applied to $\\varphi_0(w) = w$, which is $M_{\\alpha(x)}(w)$ itself. $\\square$\n\nThe null law's tail is $P(G \\ge \\theta) = \\rho^{\\theta/6 - 1}$, exponential in $\\theta$ with exponent, in units of $\\theta/\\bar m$,\n\n$$c_{\\mathrm{null}} = \\frac{\\bar m}{6}\\ln\\frac{1}{\\rho} = 1 + \\frac{3}{\\bar m} + O(\\bar m^{-2}),$$\n\nwhich decreases toward 1, the thinning fixed point, from above. In the record's convention, where the fold by $p$ is scored against the mean gap $\\bar m$ of the tile before it, $c_{\\mathrm{null}}$ reads 1.527151 at fold 7 and 1.023916 at fold 1499 (`research/history/staging/import-thinning.md` §0; the convention note of `research/history/staging/null-limsup.md` §5 records that the scaling record indexes $\\bar m$ one fold later, a ratio of 1.00477 at $p = 421$, and that nothing scored turns on it).\n\n**Verified.** The closed form is checked against brute-force iteration of the exact recursion $f_{\\mathrm{new}}(k) = q f(k) + r\\sum_{j < k} f(j) f_{\\mathrm{new}}(k - j)$ from $\\delta_1$ through folds 5 to 29, maximum relative deviation $1.02 \\cdot 10^{-14}$, with the mean equal to $\\alpha$ to eight places at every level (`research/import-thinning-01-nullmodel.js`). Our own check script (§10) repeats the recursion and finds the mean equal to $\\alpha$ to $2.7 \\cdot 10^{-15}$ at every fold and the relative deviation below $4.6 \\cdot 10^{-14}$ wherever the mass exceeds $10^{-30}$; over the full support it reaches $4.8 \\cdot 10^{-13}$ at $k \\approx 2865$, where the mass is near $10^{-300}$ and the convolution's rounding accumulates. The record's figure is therefore reproduced on the bulk and not over the whole support, and the difference is floating-point.\n\n### 3.4 The amplitude the null forces, and the moment it shares with the fold\n\n**The geometric/CRT hybrid.** The quantity the record calls $r_{\\mathrm{null}}$, and the first draft called the null's adjacent-kill rate, is a geometric gap law scored with the arithmetic qualifying weights: under the geometric law of Theorem 1 the three qualifying classes of §2 have total weight, per gap,\n\n$$r_{\\mathrm{hyb}}(p) = \\frac{2\\rho^{p-1} + \\rho^{k_c - 1} + \\rho^{k_e - 1}}{2\\alpha(1 - \\rho^p)}, \\qquad k_c = \\frac{2p - 2\\eta}{6},\\quad k_e = \\frac{4p + 2\\eta}{6},$$\n\nwhich is $X/\\mathrm{kills} = \\sum_i \\omega_i/(2N)$ evaluated on a geometric word, each class summed along its progression of modulus $6p$. It is not the adjacent-kill rate of independent deletion, which is exactly $2/p$ (§3.1, model (ii)); at fold 7, with $\\alpha = 5/3$, the hybrid reads $0.130351$ against $2/7 = 0.285714$, and the producer's own C2 table keeps $X_{\\mathrm{null}}$ apart from $X_{\\mathrm{indep}} = 4N/p$ (`research/import-thinning-01-nullmodel.js`). The hybrid's leading term is $(3/(\\bar m\\rho))\\exp(-c_{\\mathrm{null}}\\theta_p/\\bar m)$, so its amplitude in the rate law of §7 is $3/\\bar m$ at leading order, with zero free parameters; every \"null\" figure of §6 (the OLS refit, the amplitude ratio 1.97, the pair ratio 2.73) is a finite comparison of this hybrid and is labelled so.\n\n**Lemma 1 (proven).** The fold and the null multiply the mean gap by the same factor $p/(p - 2)$, exactly.\n\n*Proof.* The fold keeps the modulus $pW$ and leaves exactly $(p - 2)N$ slots, since each slot dies in exactly two copies, so its mean gap is $pW/((p-2)N) = \\bar m\\,p/(p-2)$. Under the null the gap law after the fold is the geometric with parameter $\\alpha\\,p/(p-2)$ by Proposition 1 and Theorem 1, whose mean is $\\alpha\\,p/(p-2) = (\\bar m/6)\\,p/(p-2)$ in comb units. $\\square$\n\nLemma 1 is why the domination of §5 cannot buy an exponent: at first order in the moment parameter the two processes agree identically, so any margin between them is a second-moment effect.\n\n### 3.5 The owner of the null, and the null's maximal gap\n\nThe one-class instance of Theorem 1 is the Hawkins random sieve, in which the $n$-th surviving integer deletes each later survivor independently with probability $1/n$. Heyde (1976, p. 278), following Williams, writes the gap law as $P(X_{n+1} - X_n = j \\mid X_n, Y_n) = Y_n^{-1}(1 - Y_n^{-1})^{j-1}$ with $Y_n = \\prod_{k \\le n}(1 - X_k^{-1})^{-1}$, and Rivoal (2008, p. 800) restates it with the Mertens product $m_n$ as the geometric parameter; Neudecker and Williams (1974, p. 199) make the pair process Markovian and remark that elementary properties of geometric distributions make things neat. That the geometric parameter is the mean gap and is the Mertens product is explicit in print in 1974 (the Markov transition on printed page 198, the Mertens comparison on 199; the first draft's locator was one page late), and our $\\alpha = \\prod q/(q-2)$ is its two-class instance. One difference is carried into every ownership sentence: Hawkins' sieve has a *random* product indexed by the sieve's own surviving numbers and a conditional geometric law, whereas the null here thins at deterministic prime-indexed rates; the two stochastic processes share the fractional linear mechanism and are not identical (the Lorch paragraph below already says so, and the referee asked that it be said at the headline). The composition of thinnings on the generating function is the composition semigroup of probability generating functions (Bunge 1996, with the geometric case as the worked example). The rate $2/p$ is admissible in Lorch's family of Hawkins' $p$-primes, whose Theorem 2.1 needs only $\\sum p(k)^2 < \\infty$ and $\\sum p(k) = \\infty$; what the rate alone produces there is density $1/(2\\log n)$ and not the twin density, because Lorch's sieving numbers are the survivors of the same sieve while our moduli are the primes and our comb is separate (`research/history/staging/hawkins-read.md` §4). Of the Hawkins tradition, the record marks eight items read in full at source on 2026-08-19 (Neudecker–Williams, Wunderlich 1974 and 1969, Heyde 1976 and 1978, Bui–Keating, Lorch 2007, Rivoal); Hawkins 1957 and 1974, Neudecker 1975, Wunderlich 1976, Deheuvels 1985, Lorch–Ökten 2007 and Williams 1974 were not reached, and Bunge 1996 is cited from the registry's convention row without a source read. §8 lists what that costs.\n\nThe null's maximal gap is owned too. Neudecker (1975) proves $\\limsup (p_{n+1} - p_n)/\\log^2 p_n = 1$ almost surely for Hawkins' primes, carried here second-hand from Rivoal p. 808. The proposal's trigger asked that the extinction half be retired if the null's maximal gap disagreed with that law, and scoring it required a translation the record writes out (`research/history/staging/null-limsup.md` §1): in Hawkins' sieve $\\log p_n$ is at once the mean gap and the logarithm of the point count, whereas here depth and window length are independent, so the transferable statement is \"maximal gap ≈ mean gap × ln(number of points), constant 1\". For a fixed integer number $n$ of independent geometric gaps the maximum $K$ has the exact law $P(K \\le k) = (1 - \\rho^k)^{n}$, with Gumbel form $E[G_{\\max}] \\approx (\\bar m/c_{\\mathrm{null}})(\\ln n + \\gamma)$. A fixed spatial window does not contain such a sample: its point count is random, the retained internal gaps are those that fit inside the window, and conditional on the count they are constrained by the window length (a four-site Bernoulli window at $0, 6, 12, 18$ has no internal gap of $24$ or more in any outcome, while a geometric gap with $\\alpha = 2$ is $\\ge 24$ with probability $1/8$); the producer passes the non-integer $Y/\\bar m - 1$ as $n$. So the law and its Gumbel form are large-window approximations here, not exact laws (the first draft called them exact), and forced back onto $\\ln^2 p \\cdot \\ln W$ the constant is $e^{2\\gamma}/(2C_2) = 2.402607$. Neudecker's almost-sure limsup for Hawkins' self-indexed process does not become a theorem about this fixed-depth, fixed-window process by matching mean gap times log point count; the calculation above supplies its own derivation and carries its own approximation. Scored (pre-registered, `null-limsup-prereg.md`; producer `research/null-limsup-01-score.js`): the iid law and its Gumbel form agree to 0.2%; fed Hawkins' rate and diagonal the same closed form returns $E[\\max]/\\ln^2 x \\to 1$ (0.985231 at $\\ln x = 400$); a coupled simulation reproduces the exact mean at $z = +0.56$ and $-0.36$; and the real fold's record gaps sit at 0.9178, 0.8298, 1.0223, 0.9013 of the null's expectation at the four windows $Y = 2 \\cdot 10^7$ to $2 \\cdot 10^{10}$, below it at three of four (measured). The trigger does not fire. Read literally, \"constant 1 on $\\ln^2 p$\" would have predicted 53.5 at fold 1499 against record gaps near 2000; that factor of 42 is translation and not disagreement, and a literal reading would have retired the extinction half on it.\n\n## 4. The Fold Moment Identity\n\n### 4.1 Statement and proof\n\nFix $\\lambda > 0$ and define, on the cyclic gap word of $T_x$,\n\n$$\\Phi = \\frac{1}{N}\\sum_i e^{\\lambda g_i}, \\qquad \\Psi = \\frac{1}{N}\\sum_i e^{\\lambda(g_i + g_{i+1})}, \\qquad \\Omega = \\frac{1}{N}\\sum_i \\omega_i e^{\\lambda g_i},$$\n\nwith $\\Phi_{\\mathrm{new}}$ the same $\\Phi$ on the folded word, and\n\n$$\\Delta = \\frac{1}{N}\\Big(\\sum_{\\text{runs } R} e^{\\lambda\\,\\mathrm{span}(R)} - 2N\\Psi\\Big),$$\n\nwhere the sum is over the kill runs of the fold and $\\mathrm{span}(R)$ is the new gap the run creates, the sum of the $\\ell + 1$ old gaps it merges.\n\n**Theorem 2 (Fold Moment Identity; proven, `research/history/staging/import-thinning.md` §2.3, from the two identities of §2).**\n\n$$(p - 2)\\,\\Phi_{\\mathrm{new}} = (p - 4)\\,\\Phi + \\Omega + 2\\Psi + \\Delta.$$\n\n*Proof.* Every gap of the folded word is of exactly one of two kinds: a surviving copy of an old gap, or the span of a kill run. By the Merge Rate Identity, gap $i$ has exactly $p - 4 + \\omega_i$ surviving copies, so the surviving copies contribute $\\sum_i (p - 4 + \\omega_i)e^{\\lambda g_i} = N(p - 4)\\Phi + N\\Omega$ to $N_{\\mathrm{new}}\\Phi_{\\mathrm{new}} = (p - 2)N\\Phi_{\\mathrm{new}}$. The runs contribute $\\sum_R e^{\\lambda\\,\\mathrm{span}(R)}$, which is $2N\\Psi + N\\Delta$ by the definition of $\\Delta$. Dividing by $N$ gives the identity. $\\square$\n\nThe definition of $\\Delta$ is what makes the identity exact. If every run had length 1, each killed position would merge its two neighbouring gaps, each slot dies in exactly two copies, and the run term would be exactly $\\sum_i 2e^{\\lambda(g_i + g_{i+1})} = 2N\\Psi$; $\\Delta$ is the correction from the runs of length $\\ell \\ge 2$, each of which replaces $\\ell$ single-merge terms by one term with a larger exponent. $\\Delta$ is supported entirely on the $X$ adjacent kill pairs and vanishes when $X = 0$, as at fold 11.\n\nThe two terms $\\Omega$ and $\\Delta$ live on the same runs, and together they are one non-negative quantity.\n\n**Proposition 3 (proven here).** For a kill run $R$ of length $\\ell$ merging the gaps $h_0, h_1, \\dots, h_\\ell$, with $h_0$ before the first killed position, $h_\\ell$ after the last and $h_1, \\dots, h_{\\ell-1}$ the interior gaps, put $x_j = e^{\\lambda h_j}$ and\n\n$$F_R = \\prod_{j=0}^{\\ell} x_j - \\sum_{j=1}^{\\ell} x_{j-1}x_j + \\sum_{j=1}^{\\ell-1} x_j.$$\n\nThen $F_R \\ge 0$, $F_R = 0$ for $\\ell = 1$, $F_R = x_1(x_0 - 1)(x_2 - 1)$ for $\\ell = 2$, and\n\n$$\\Omega + \\Delta = \\Gamma := \\frac{1}{N}\\sum_R F_R \\ \\ge\\ 0, \\qquad\\text{so}\\qquad (p - 2)\\,\\Phi_{\\mathrm{new}} = (p - 4)\\,\\Phi + 2\\Psi + \\Gamma.$$\n\n*Proof.* The interior gaps of the runs are exactly the adjacent killed pairs, so $N\\Omega = \\sum_i \\omega_i e^{\\lambda g_i} = \\sum_R \\sum_{j=1}^{\\ell-1} x_j$; the single-merge accounting charges each killed position its two neighbouring gaps, so $2N\\Psi = \\sum_R \\sum_{j=1}^{\\ell} x_{j-1}x_j$; and the run term is $\\sum_R \\prod_j x_j$. Hence $N(\\Omega + \\Delta) = \\sum_R F_R$. For the sign, $F_1 = 0$ and\n\n$$F_\\ell(x_0, \\dots, x_\\ell) = F_{\\ell-1}(x_0, \\dots, x_{\\ell-1}) + (x_\\ell - 1)\\Big(\\prod_{j=0}^{\\ell-1} x_j - x_{\\ell-1}\\Big),$$\n\nwhere the second term is non-negative because every $x_j \\ge 1$; induction on $\\ell$ gives $F_\\ell \\ge 0$. The $\\ell = 2$ form is the identity $x_0x_1x_2 - x_0x_1 - x_1x_2 + x_1 = x_1(x_0 - 1)(x_2 - 1)$. $\\square$\n\nAt the exact folds 7 to 29 other than 23, and at the sampled window folds $100 \\le p \\le 1499$ of the four scaling windows other than $p = 103$ in the fifth window (§7.1, clause 4), all runs have length at most 2, so $\\Gamma$ is the sum over the $X$ length-2 runs of $x_1(x_0 - 1)(x_2 - 1)$: the merged pair's own weight times the excess weights of its two outer gaps. Fold 23's 62 runs of length 3 add their $F_3 \\ge 0$; later exact folds also carry length-three runs (fold 41 of return #23), so \"at most 2\" is a statement about the sampled range and nothing more.\n\n**Verified.** At folds 7 through 29 (the exact tiles $T_5$ through $T_{23}$ folded once) and $\\lambda = u/\\bar m$ with $\\bar m$ the mean gap of the tile *before* the fold and $u \\in \\{0.25, 0.5, 0.75, 0.9\\}$ (this convention is the record's throughout, and it matters: taking $\\bar m$ after the fold moves the $c'_{\\min}$ cells of §5 by up to 0.16) by `research/import-thinning-02-coalescence.js`, with the run-length spectra, $X$ and $M = 2N - X$ reproducing the spend ledger of §2 from an engine that computes them differently; and by our check script to machine precision at all 28 cells, including $\\Gamma = \\Omega + \\Delta$ and $F_R \\ge 0$ run by run (§10). The share of $\\Delta$ in $(p - 2)\\Phi_{\\mathrm{new}}$ at $u = 0.5$ reads $-3.78 \\cdot 10^{-3}$, $-1.43 \\cdot 10^{-3}$, $-9.53 \\cdot 10^{-4}$, $-1.11 \\cdot 10^{-3}$, $-1.10 \\cdot 10^{-3}$ at folds 13, 17, 19, 23, 29 (measured). Since $\\Delta$ is the entire contribution of every sum of $m \\ge 3$ adjacent gaps, the recursion closes at the pair level up to $\\Gamma$.\n\n### 4.2 The deviation from the null, exactly\n\nWrite $\\mathcal N_p$ for the null map on exponential moments at rate $2/p$, $\\mathcal N_p\\Phi = (1 - 2/p)\\Phi/(1 - 2\\Phi/p)$, which is Proposition 1 read at $w = e^{6\\lambda}$ in comb units. As the exponential moment of a compound-geometric law it is defined only for $\\Phi < p/2$; at the pole it is undefined and beyond it the rational continuation is not the moment of any null, whose exponential moment diverges. Corollary 1 is an algebraic identity for every $\\Phi$; its probabilistic reading, and the reading of $K_{\\mathrm{crit}}$ below, carry the domain $\\Phi < p/2$.\n\n**Corollary 1 (proven here; the leading term is the record's).**\n\n$$(p - 2)\\big(\\Phi_{\\mathrm{new}} - \\mathcal N_p\\Phi\\big) = 2(\\Psi - \\Phi^2) + \\Gamma - \\frac{4\\Phi(\\Phi - 1)^2}{p - 2\\Phi}.$$\n\n*Proof.* $(p-2)\\mathcal N_p\\Phi = (p-2)^2\\Phi/(p - 2\\Phi)$. Subtracting it from $(p - 4)\\Phi + 2\\Phi^2$ over the common denominator $p - 2\\Phi$ gives numerator $((p-4)\\Phi + 2\\Phi^2)(p - 2\\Phi) - (p-2)^2\\Phi = -4\\Phi + 8\\Phi^2 - 4\\Phi^3 = -4\\Phi(\\Phi - 1)^2$. Substitute into Theorem 2 with $\\Omega + \\Delta = \\Gamma$. $\\square$\n\nTerm by term against $\\mathcal N_p$: the survival coefficient $(p-4)/(p-2)$ is smaller than the null's $1 - 2/p$, the merge coefficient $2/(p-2)$ is larger than $2/p$, both by $4/(p(p-2))$, and the merge term carries $\\Psi$ where the null carries $\\Phi^2$. The record reads the last difference as the whole story: \"the entire deviation of CRT thinning from independent thinning, at first order in $1/p$, is $\\Psi - \\Phi^2$\" (`research/history/staging/import-thinning.md` §0). Corollary 1 says exactly when that reading holds. All three terms on the right are $O(\\lambda^2)$, so the comparison is in $p$, not in $\\lambda$: the last term is $O(1/p)$ relative to the first, and $\\Gamma$ is supported on the qualifying gaps, whose share of the word is $(Q_0 + Q_+ + Q_-)/N$ (0.0307 at fold 29 on the exact ladder) and falls like the pair density $\\exp(-c\\,\\theta_p/\\bar m)$ deeper. On the exact ladder to $T_{29}$ neither is small: at fold 29 and $u = 0.25$ the four quantities $2(\\Psi - \\Phi^2)$, $\\Gamma$, the last term and their total read $-0.0062$, $+0.0033$, $-0.0185$, $-0.0213$, and at $u = 0.5$ they read $-0.0665$, $+0.0338$, $-0.1676$, $-0.2003$ (our check script, §10), so there the $1/p$ term is three times the pair term. In the localized frame the ordering reverses: $1 - K$ sits flat between $2.0 \\cdot 10^{-3}$ and $3.7 \\cdot 10^{-3}$ at $u = 0.25$ from $p = 31$ to $1009$ (measured, §5.2) while the last term is below $5 \\cdot 10^{-4}/\\Phi^2$ by $p = 1009$, and $\\Gamma$ carries the vanishing qualifying share. So the record's sentence is an asymptotic reading in $p$: wrong on the tiles a computer reaches, and at the localized depths only approximately right, since with the printed $\\Phi$ and $K$ at $u = 0.25$ the rational term is about 40.8%, 23.3%, 14.2% and 7.6% of the pair term at $p = 101, 211, 421, 1009$, sparse support does not by itself bound an exponentially weighted $\\Gamma$, and an $O(1/p)$ absolute correction is not $O(1/p)$ relative to a pair covariance that may tend to zero; no full $\\Gamma$ calculation or uniform bound on those windows is supplied, and the cyclic identity needs boundary adjustments before it is applied to non-cyclic window averages. Where the pair term dominates, $\\Psi - \\Phi^2 < 0$ is a moment condition *related to* $H''$ at $m = 2$ and not that hypothesis: $H''$ of `research/a3-05-bound-L.md` §8 is a conditional count inequality for consecutive gaps all exceeding a threshold, for every $m \\ge 2$, and the sign of one exponential covariance neither gives that family nor is given by its $m = 2$ member. The referee's abstract cyclic word $[18, 6, 12, 6, 18]$ at $\\lambda = \\ln 2/6$ has exponential covariance $-16/25$ and threshold-18 covariance $+1/25$ (both reproduced for this revision); it is not a tile and refutes only the claimed equivalence of the two statistical conditions. We write \"related moment condition\" throughout and state no implication between them.\n\nCorollary 1 also gives the exact criterion for the fold to be dominated by its own null map. $\\Phi_{\\mathrm{new}} \\le \\mathcal N_p\\Phi$ if and only if $K = \\Psi/\\Phi^2 \\le K_{\\mathrm{crit}}$ with\n\n$$K_{\\mathrm{crit}} = 1 + \\frac{1}{\\Phi^2}\\Big[\\frac{2\\Phi(\\Phi-1)^2}{p - 2\\Phi} - \\frac{\\Gamma}{2}\\Big],$$\n\nwhich is the record's $[\\Phi(2 + (p-4)\\Phi)/(p - 2\\Phi) - (\\Omega + \\Delta)/2]/\\Phi^2$ rearranged. The bracket is a race between the slack the fold has because it destroys $4 - \\omega \\ge 2$ copies of every gap where the null destroys 2 in expectation, and the run term $\\Gamma \\ge 0$, which pushes the other way. Measured, the slack wins: $K_{\\mathrm{crit}}$ is 1.00443540 at fold 29 and $u = 0.25$, rising to 1.11721750 at $u = 0.9$, and above 1 at every cell checked. Negative association, $K < 1$, is therefore sufficient for domination at those cells and far from necessary; it is not sufficient in general, since Proposition 3 gives the sign of $\\Gamma$ and not its size. The pre-registered proxy $K < 1 - 4/p$ was a large-$p$ approximation of $K_{\\mathrm{crit}}$ and is withdrawn.\n\n### 4.3 What is measured about $K$\n\n$K = \\Psi/\\Phi^2 < 1$ at all 28 cells of the exact ladder (folds 7 to 29, four values of $\\lambda\\bar m$), range 0.89362755 to 0.99894799, and at all 24 cells of the localized ladder out to $p = 1009$ (measured, `research/import-thinning-02-coalescence.js` and `research/import-thinning-03-deepfolds.js`). Adjacent gaps are negatively associated at exponential order at every fold in the two sampled ranges. $K$ is not the mechanism of the per-fold field of §7.4: it spans a factor 1.009 across folds 101, 211 and 421 while the field spans 2.28, and the record refutes it as a field predictor (`research/OUTCOMES.md`, row \"K = Ψ/Φ² as the M_p field mechanism\", 2026-08-20).\n\n### 4.4 The technique is the tradition's\n\nExact stage-to-stage moment recursions with no error term are how the Hawkins literature works. Wunderlich (1974, eq. (5)) has $E[a_{n+1}^k] - E[a_n^k] = ((1 - 1/n)^k - 1)E[a_n^{k+1}]$ for the survival product, Lorch (2007, Lemma 3.3) the same at an arbitrary rate, and Bui and Keating (2006, eq. (4)) a joint recursion mixing the survival density with a pair count; Bui and Keating also expand a joint two-point probability about the product of marginals with the first correction named (their p. 2 and Lemma 2). Theorem 2 must be presented as an application of that standard technique to a different object, and the novelty of $\\Psi - \\Phi^2$ is in which two objects are differenced, not in the move of differencing them. Three differences carry the identity's own content: the tradition's moments are power moments of the survival density and ours is an exponential moment of the gap word; theirs close in the moment index at one stage and ours mixes four objects across two stages; and every one of theirs describes the random sieve, whereas Theorem 2 describes the arithmetic fold, which is the whole point of subtracting the null (`research/history/staging/hawkins-read.md` §5). The deviation of the random sieve from the true sieve is a question that tradition asks out loud and answers only as a constant: the missing $e^\\gamma$ in Mertens (Neudecker–Williams p. 199), the missing twin-prime constant (Bui–Keating p. 2), $n/a$ against $n/\\varphi(a)$ (Rivoal, Theorem 2). No identity for it was found in print.\n\n## 5. Domination on exponential moments, and why it is not a route\n\n### 5.1 The measured domination\n\nAsk for the fold to be dominated by an independent thinning at some rate $c'/p$ in place of $2/p$: $\\Phi_{\\mathrm{new}} \\le (1 - c'/p)\\Phi/(1 - (c'/p)\\Phi)$. Solving for the smallest such $c'$ gives the closed form\n\n$$c'_{\\min}(\\lambda) = \\frac{p\\,(\\Phi_{\\mathrm{new}} - \\Phi)}{\\Phi\\,(\\Phi_{\\mathrm{new}} - 1)},$$\n\nand $c'_{\\min} < 2$ means the fold is dominated by a thinning at a rate better than its own. This is a bound on one exponential moment and, by Markov, a Chernoff tail bound; it is not a stochastic order and is not presented as one.\n\n**Measured: $c'_{\\min} < 2$ at 52 of 52 cells.** On the exact ladder, folds 7 to 29:\n\n| fold | $\\lambda\\bar m = 0.25$ | 0.50 | 0.75 | 0.90 |\n|---|---|---|---|---|\n| 7 | 1.9181 | 1.8105 | 1.6722 | 1.5743 |\n| 11 | 1.8553 | 1.6797 | 1.4763 | 1.3441 |\n| 13 | 1.9010 | 1.7796 | 1.6321 | 1.5297 |\n| 17 | 1.9434 | 1.8772 | 1.7922 | 1.7252 |\n| 19 | 1.9508 | 1.8866 | 1.7909 | 1.7066 |\n| 23 | 1.9490 | 1.8769 | 1.7636 | 1.6621 |\n| 29 | 1.9524 | 1.8785 | 1.7537 | 1.6389 |\n\nIn the localized frame at $Y = 2 \\cdot 10^9$, folds 31, 53, 101, 211, 421 and 1009, the same four columns lie in $[1.8579, 1.9811]$, $[1.8311, 1.9298]$, $[1.7318, 1.8332]$ and $[1.4812, 1.7187]$, the maximum of every column at fold 211 and the minimum at 1009 (`research/history/staging/import-thinning.md` §3.2). The 1009 row at $u = 0.25$ is the one cell where the instrument is thin, and the record reports it without leaning on it.\n\n### 5.2 The margin is a second-moment effect\n\n**Lemma 2 (proven; `research/history/staging/import-thinning.md` §4.3).** At every fold, $c'_{\\min}(\\lambda) \\to 2$ as $\\lambda \\to 0$.\n\n*Proof.* By Lemma 1, $\\Phi = 1 + \\lambda\\bar m + O(\\lambda^2)$ and $\\Phi_{\\mathrm{new}} = 1 + \\lambda\\bar m\\,p/(p - 2) + O(\\lambda^2)$. Substituting, the numerator of $c'_{\\min}$ is $p\\lambda\\bar m\\cdot 2/(p - 2) + O(\\lambda^2)$ and the denominator is $\\lambda\\bar m\\,p/(p - 2) + O(\\lambda^2)$; the ratio tends to 2. $\\square$\n\nThe margin is linear in $\\lambda$, not quadratic; the first draft said $O(\\lambda^2)$ and was wrong. Write $\\Phi = 1 + m\\lambda + a\\lambda^2 + O(\\lambda^3)$ and $\\Phi_{\\mathrm{new}} = 1 + n\\lambda + b\\lambda^2 + O(\\lambda^3)$ with $n = mp/(p-2)$, $a = E[g^2]/2$, $b = E[g_{\\mathrm{new}}^2]/2$; dividing the two first-order-vanishing quantities gives\n\n$$c'_{\\min}(\\lambda) = 2 + \\Big(\\frac{(p-2)b - pa}{n} - 2m\\Big)\\lambda + O(\\lambda^2),$$\n\nwhich at $T_5 \\to T_7$ with exact integer gap moments is $c'_{\\min} = 2 - (20/7)\\lambda + O(\\lambda^2)$, checked in 40-digit arithmetic: $(2 - c'_{\\min})/\\lambda = 2.858621, 2.857291, 2.857158$ at $\\lambda = 10^{-4}, 10^{-5}, 10^{-6}$ (referee report 72 §2; reproduced for this revision). The difference of two moment generating functions can begin at second order while this normalised rate margin begins at first order; the data of §5.1 at $u = 0.01$ and $0.05$ already show it. A tail bound at level $x$ needs one absolute $\\lambda$ carried through every fold of the ladder, and along that trajectory $u_q = u_x\\bar m(q)/\\bar m(x)$ is small at every fold below the top, so the ladder spends nearly all of its $\\sum 1/q$ where the margin is small. A pointwise limit at each fold does not by itself show that the accumulated margin along an expanding ladder changes only a constant: that needs a bound uniform in the fold index before the errors are summed over primes, and none is proved here. What follows is a model calculation, pooling and interpolating measurements, and not such a bound. Composed on the fixed-$\\lambda$ trajectory, with $c'$ clamped to 2 at $u = 0$ and interpolated linearly between the four measured values, the implied exponent $c = \\bar m/(6\\prod_q q/(q - c'_q))$ reads\n\n| $u$ at the top fold | implied exponent, first draft (faulty interpolator) | implied exponent, corrected interpolation |\n|---|---|---|\n| 0.25 | 1.0302 | 1.0302 |\n| 0.50 | 1.0614 | 1.0589 |\n| 0.75 | 1.0885 | 1.0900 |\n| 0.90 | 1.1096 | 1.1167 |\n\nThe first-draft column came from `research/import-thinning-03-deepfolds.js`, whose `cOfU` selects its index with `US[i+1] < u` and then interpolates on the preceding interval, so that it does not return the supplied grid value at $u = 0.50$ and $0.75$ (referee report 72 §4, with a static checker on the captured rounded cells); the corrected column uses the interval containing $u$ on the same rounded inputs and is approximate at the fourth decimal, pending a full-precision regeneration from the producer. Against 1.1104, 1.2322, 1.2988, 1.8158 if the top fold's margin were held at every fold, which is not licensed. The pre-registered prediction U2, that the margin decays in $p$, is refuted: at $u = 0.25$ it reads $4.650 \\cdot 10^{-2}$, $4.775 \\cdot 10^{-2}$, $4.259 \\cdot 10^{-2}$, $1.891 \\cdot 10^{-2}$, $4.628 \\cdot 10^{-2}$, $1.421 \\cdot 10^{-1}$ at $p = 31, 53, 101, 211, 421, 1009$, with no trend, and $p(2 - c'_{\\min})$ grows from 1.44 to 143.4; U2's conclusion, that the accumulated margin is a constant, is not established by this: the margin decays in $\\lambda$ at each fixed fold, and the fixed-$\\lambda$ trajectory forces $\\lambda\\bar m \\to 0$ down the ladder, but the sum of first-order margins over the ladder is not bounded here. One further consequence of Lemma 2: for any fixed $0 < c' < 2$, the hypothesis of Proposition 2 for every $x$ is impossible already at a fixed early fold, since $\\lambda_x \\to 0$ makes that fold's $c'_{\\min}$ exceed $c'$; among fixed positive rates $c' \\le 2$ the nonvacuous candidate is $c' = 2$.\n\n### 5.3 The conditional theorem, and what its condition costs\n\n**Theorem 3 (Fold Tail Propagation; proven as a conditional statement, `research/history/staging/import-thinning.md` §4.1).** Fix $\\lambda > 0$ and $c' \\le 2$, put $\\alpha' = \\prod_{5 \\le q \\le x} q/(q - c')$ and $\\rho' = 1 - 1/\\alpha'$, and suppose $e^{6\\lambda} < 1/\\rho'$. Suppose that at every fold $q$ of the ladder $\\Phi(\\lambda) < q/c'$ and $\\Phi_{\\mathrm{new}}(\\lambda) \\le D_q(\\Phi(\\lambda))$, where $D_q(\\Phi) = (1 - c'/q)\\Phi/(1 - (c'/q)\\Phi)$. Then the gap word of $T_x$ satisfies $\\#\\{i : g_i \\ge \\theta\\} \\le N\\,M_{\\alpha'}(e^{6\\lambda})\\,e^{-\\lambda\\theta}$ for every $\\theta$, that is, $P(G \\ge \\theta) \\le C\\exp(-c\\,\\theta/\\bar m)$ with $c = \\lambda\\bar m$ and $C = M_{\\alpha'}(e^{6\\lambda})$ at the fixed $\\lambda$. If moreover the domination hypothesis holds at $\\lambda = 1/(6\\alpha')$, which lies in the admissible range, then the bound holds with $c = \\bar m/(6\\alpha') \\ge 1$ and $C = M_{\\alpha'}(e^{1/\\alpha'})$. Here $0 < c' \\le 2$; the case $c' = 0$ and negative rates are excluded from the probabilistic reading, since $q/c'$ appears as a pole.\n\n*Proof.* $D_q$ is the thinning map $M_{\\alpha_q}$ with $\\alpha_q = q/(q - c')$, read at $w = \\Phi$; it is increasing on $[0, q/c')$, its pole. By Proposition 1 the maps compose to $M_{\\alpha'}$, and the base comb has $\\Phi = e^{6\\lambda}$ exactly, so the induction $\\Phi_q \\le D_q(\\Phi_{q^-}) \\le D_q(B_{q^-}) = B_q$, with $B$ the running composed bound, is valid as long as each $B_{q^-}$ lies below the pole of $D_q$. That is guaranteed by the hypothesis $e^{6\\lambda} < 1/\\rho'$: the composed value $M_{\\alpha'}(e^{6\\lambda})$ is finite, and since $M_{ab} = M_a \\circ M_b$ with $M_b$ a bijection of $[0, \\text{pole of } M_b)$ onto $[0, \\infty)$, the pole of $M_{ab}$ is the $M_b$-preimage of the pole of $M_a$, so finiteness of the whole forces every intermediate value below its own pole. Hence $\\Phi_x(\\lambda) \\le M_{\\alpha'}(e^{6\\lambda})$, the exponential moment of a geometric with mean $\\alpha'$. Markov's inequality gives $\\#\\{i : g_i \\ge \\theta\\} \\le N\\Phi_x(\\lambda)e^{-\\lambda\\theta}$. In units of $\\theta/\\bar m$ the exponent is $\\lambda\\bar m = 6\\lambda\\alpha$. The admissible range $6\\lambda < \\ln(1/\\rho') = c_{\\mathrm{null}}(\\alpha')/\\alpha'$ contains $6\\lambda = 1/\\alpha'$ because $c_{\\mathrm{null}}(\\alpha') = \\alpha'\\ln(1/(1 - 1/\\alpha')) > 1$, and if the domination hypothesis is assumed at that $\\lambda$ the same argument gives the exponent $c = \\alpha/\\alpha' = \\bar m/(6\\alpha') \\ge 1$ with $C = M_{\\alpha'}(e^{1/\\alpha'})$ finite. The first draft asserted the second form while retaining the $C$ of the originally fixed $\\lambda$; the hypotheses hold at one $\\lambda$ and the proof cannot choose another without assuming it there. The referee's first-fold counterexample to the first draft's statement: $x = 5$, $c' = 2$, $\\lambda = 0.001$, $\\theta = 6$; the old comb has moment $e^{0.006}$, the pole conditions hold, $T_5$'s gaps are $6, 12, 12$ and its moment lies below the thinned-comb moment $C = 1.010070$, yet $P(G \\ge 6) = 1$ against the printed $C e^{-6/10} = 0.554338$ (reproduced for this revision; the domination at $\\lambda = 1/(6\\alpha') = 1/10$ fails there, since $3w/(5 - 2w) - (w + 2w^2)/3 = 4w(w-1)^2/(3(5-2w)) > 0$ at $w = e^{6\\lambda}$). $\\square$\n\nBy Theorem 2, the hypothesis $c' \\le 2$ rearranges into an upper bound on $\\Psi$ in terms of $\\Phi$, $\\Omega$ and $\\Delta$: a moment condition related to $H''$ at $m = 2$, with $m \\ge 3$ entering only through $\\Delta$, and not equivalent to $H''$ (§4.2). The base case $m = 1$ is not assumed anywhere, since the induction starts at the mod-6 comb whose gap law is $\\delta_6$. On the naming question the condition is therefore clean: it is not the $L = 1$ residue count that `research/attack-l1-residue.md` showed to be the Zone Postulate in residue notation, and it is not Assumption A. It is a moment condition on adjacent pairs reached from a third direction.\n\n**Proposition 2 (proven here; the record states it as an argued chain, `research/history/staging/import-thinning.md` §4.2 and `research/OUTCOMES.md`).** Suppose that for every $x$ the hypothesis of Theorem 3 holds along the ladder to $x$ at $\\lambda = \\lambda_x = 1/(6\\alpha'(x))$, with a fixed $0 < c' \\le 2$ (by §5.2 only $c' = 2$ is nonvacuous for all $x$; the domination is assumed at $\\lambda_x$ itself, which is the assumption Theorem 3's second form needs). Then $G_2(x\\#) \\ll x\\ln^2 x$, which is the naive tail form rejected in `research/a3-05-bound-L.md` §8, and the Zone Postulate $G_2(x\\#) < x'^2 - 2$ holds for all large $x$.\n\n*Proof.* At $\\lambda_x$ the proof of Theorem 3 gives $\\#\\{i : g_i \\ge \\theta\\} \\le N\\,C_x\\,e^{-\\lambda_x\\theta}$ with $C_x = M_{\\alpha'}(e^{1/\\alpha'})$. Taking $\\theta = G_2(T_x)$, where the count is at least 1, gives $G_2 \\le (\\ln N + \\ln C_x)/\\lambda_x = 6\\alpha'(\\ln N + \\ln C_x) \\le \\bar m\\,(\\ln N + \\ln C_x)$, since $6\\alpha' \\le 6\\alpha = \\bar m$. Here $\\ln N = \\sum_{5 \\le q \\le x}\\ln(q - 2) = \\theta(x) - \\ln 6 + \\sum_{5 \\le q \\le x}\\ln(1 - 2/q) = x(1 + o(1))$ by the prime number theorem; $\\bar m = 6\\prod q/(q-2) \\sim (e^{2\\gamma}/(2C_2))\\ln^2 x$, which is $2.4\\ln^2 x$ to three figures at $x = 10^6$ (`research/history/staging/null-limsup.md` §1); and $C_x \\le 5\\alpha'$ for $\\alpha' \\ge 2$, because $\\alpha' - (\\alpha' - 1)e^{1/\\alpha'} \\ge 1/(3\\alpha')$ there and $e^{1/\\alpha'} \\le e^{1/2}$, so $\\ln C_x = O(\\ln\\ln x)$. Hence $G_2(x\\#) \\le 2.5\\,x\\ln^2 x$ for all large $x$, and $2.5\\,x\\ln^2 x < (x')^2 - 2$ for all large $x$ because $x' \\ge x$ and $x\\ln^2 x = o(x^2)$. $\\square$\n\nThat is the statement `research/a3-05-bound-L.md` §8 calls far too strong, stronger than the Zone Postulate and stronger than anything known. It is not known to be false, and the hypothesis of Theorem 3 at $\\lambda_x$ implies it. So the moment hypothesis is at least as strong as the Zone Postulate: strong and unproved. That is what the implication shows; it is not a proof that such a sufficient hypothesis can never be established, nor that an implication toward the target is itself a wrong direction, and the first draft's sentence to that effect is withdrawn. The record's closure is of the induction route that would have proved the hypothesis level by level; the measurements of §5.1 are evidence for the Zone Postulate and not a route to it.\n\n### 5.4 The coupling obstruction\n\nA monotone coupling would need the fold's deleted set to be a randomisation of an independently deleted set. The first draft argued that it is not because \"a gap merges only if it qualifies\", so that the merge event is a function of the gap value and the two laws are mutually singular; both halves of that argument are wrong and are withdrawn (referee report 72 §6). The qualifying condition concerns *both* endpoints of a gap dying in the same copy, that is, an interior gap of a kill run of length at least two; a singleton kill merges its two outer gaps whatever their sizes, and in the actual $T_5 \\to T_7$ fold the singleton kill at slot 47 consumes the gaps 6 and 12, neither of which qualifies. And on a finite set, Bernoulli deletion with probability strictly between 0 and 1 gives every configuration positive probability, each CRT configuration included, so no mutual singularity holds. What can be proved is narrower. At the same rate the fold's deletion count is fixed at $2N$ while the independent count is non-degenerate with the same expectation; an almost-sure set-inclusion coupling would force equal cardinalities and hence equal sets, which is impossible for those two laws. Likewise a scalar stochastic order between the two gap laws together with equality of means (Lemma 1) forces equality in law. Neither argument rules out altered-rate or moment comparisons, which is what §5.1 measures and what a tail theorem needs. The recorded total-variation and tail differences of §6.3 stand as finite observations, without the withdrawn mechanism.\n\nThe Merge Rate Identity fixes the number of adjacent killed pairs ($X = \\sum_i \\omega_i$) and the survival counts exactly; that is what the null is compared against in §6.3, and the comparison is an observation and not a mechanism.\n\n## 6. The null against the data\n\n### 6.1 The derived exponent and the pre-registered predictions\n\nFive predictions were written into the banner of `research/import-thinning-01-nullmodel.js` and printed by a stage (`STAGE=predict`) that stops before touching any measurement; all five are in the embedded tail. The record's fitted model is the two-parameter Poisson maximum-likelihood fit of `research/history/staging/attack-foldL-06-scaling.md` §3.1 over folds $p \\ge 100$ at $Y = 2 \\cdot 10^9$, $A = 2.4312 \\cdot 10^{-2}$, $c = 1.0818 \\pm 0.0317$. The geometric/CRT hybrid $r_{\\mathrm{hyb}}(p)$ of §3.4 fitted by the same shape, by ordinary least squares in the logarithm over the same folds, gives $A = 4.7843 \\cdot 10^{-2}$, $c = 1.0577$.\n\n| prediction | outcome |\n|---|---|\n| R1, shape: fitted $c_{\\mathrm{null}}$ within 10% of 1.0818 | held, 2.2% ($c_{\\mathrm{null}} = 1.0577$) |\n| R2, amplitude: $A_{\\mathrm{null}}/A_{\\mathrm{record}} \\in [1.5, 5]$ | held, 1.97 |\n| R3, residue and sign: null overshoots measured pairs by 2 to 4 at $Y = 2 \\cdot 10^9$ | held, 2.73 (5486.4 against 2006) |\n| R4, extinction: brackets 421 at $2 \\cdot 10^9$, misses high at $2 \\cdot 10^{10}$ | held as written |\n| R5, decomposition: the overshoot is mostly tail, hazard within 30% of 1 | refuted (§6.3) |\n\nThe 2.2% is a zero-parameter agreement on the exponent, and the two misses in R2 and R3 are part of the same claim. A referee's reading of it, which the record shares, is that the derived value is the closed form refitted through the record's own estimator over the record's own window, from a model that is wrong by a factor of two on the amplitude and by 2.73 on the pair count at the calibration window; a model that wrong on two quantities and right to 2.2% on a third invites a coincidence charge. The three routes to $c \\approx 1.06$ to $1.09$, the fit to measured pairs, the closed-form hybrid, and the domination composed along the ladder, are three models of one measurement and not three measurements; R1 to R3 are predictions of the hybrid, not of independent deletion.\n\n### 6.2 Per decade, and the direction of the miss\n\nMeasured adjacent pairs against the null by decade of $p$ at $Y = 2 \\cdot 10^9$:\n\n| decade of $p$ | folds | mean $\\theta/\\bar m$ | $X$ measured | $X$ null | ratio |\n|---|---|---|---|---|---|\n| [5, 10) | 2 | 1.600 | 19,047,619 | $7.449 \\cdot 10^6$ | 2.5572 |\n| [10, 30) | 6 | 1.747 | 1,087,762 | $3.161 \\cdot 10^6$ | 0.3441 |\n| [30, 100) | 15 | 2.904 | 180,556 | $1.963 \\cdot 10^5$ | 0.9200 |\n| [100, 300) | 37 | 5.674 | 2,003 | $5.467 \\cdot 10^3$ | 0.3664 |\n| [300, 1000) | 106 | 12.588 | 3 | $1.962 \\cdot 10^1$ | 0.1529 |\n| [1000, 1500) | 71 | 20.606 | 0 | $1.512 \\cdot 10^{-4}$ | 0 |\n\nFrom the [30, 100) decade on the ratio falls with depth: the fold is progressively further below independent thinning the deeper the fold, which is the safe direction for every bound that leans on it. The two shallow decades carry no prediction and are printed for completeness. On the exact tiles to $T_{29}$ the geometric law models the word within a factor 2.5 at every fold of that ladder, while independent deletion models the fold badly, off by 3 to 7 at the same folds and by orders of magnitude deeper (`research/history/staging/import-thinning.md` §2.1). The two statements are different: the null is a fair model of the gap histogram and a poor model of the merge process.\n\n### 6.3 Where the fold differs from its null: bulk, not tail\n\nR5 expected the overshoot to be almost all tail. In the window at $Y = 2 \\cdot 10^9$:\n\n| fold | $\\theta$ | tail factor $\\#\\{g \\ge \\theta\\}$, true/null | hazard $P(g = \\theta \\mid g \\ge \\theta)$, measured | null $1/\\alpha$ | hazard factor |\n|---|---|---|---|---|---|\n| 101 | 204 | 0.4470 | 0.08643 | 0.11490 | 0.7442 |\n| 211 | 420 | 0.3256 | 0.24040 | 0.08553 | 2.8057 |\n| 421 | 840 | 0.1644 | 0.19091 | 0.06762 | 2.8232 |\n\nThe tail at $\\theta \\approx 2p$ is two to six times lighter than geometric and getting lighter with depth, while the discrete hazard is about 2.8 times the null's constant at the two deeper folds. A tail that falls faster than geometric puts more of its conditional mass on the threshold itself; the gap word above $2p$ is steeper than geometric, and the steepness is what the fitted amplitude absorbs.\n\nThe total-variation comparison isolates the thinning structure from the gap law by thinning the measured old histogram itself, $f = qh + r(h * f)$, and comparing with the fold. The pre-registered prediction S5, that the excess of the fold over independent thinning sits near $\\theta$ and $2\\theta$, is refuted in the opposite direction: the excess sits in the bulk, at gaps 18, 30 and 42 at every fold, and in the far tail the fold is at or below independent thinning, by a factor 2.4 at $3\\theta$ for folds 23 and 29. Total variation falls monotonically, 0.18069 at fold 7 to 0.01397 at fold 29, and is under 0.05 from fold 13 on (measured, `research/import-thinning-02-coalescence.js`). The first draft explained this by \"the fold merges only at qualifying gaps and leaves the bulk alone\", which is false (§5.4: singleton kills merge outer gaps of any size). What the identities fix is the count of adjacent killed pairs, $X = 2Q_0 + Q_+ + Q_-$, against independent thinning's $4N/p$ in expectation, and the exact survival multiplicities $p - 4 + \\omega$; the convolution control of model (iii) also resamples the gap order. The measured total-variation and tail differences are recorded as observations, and no mechanism is claimed for them here.\n\n## 7. The extinction law\n\n### 7.1 The statement, with its known failures inside it\n\nConventions are those of `research/history/staging/attack-foldL-06-scaling.md` §0: a window $[A, A + Y)$ with $6 \\mid A$, positions $n \\equiv 5 \\pmod 6$, $\\mathrm{key}(n) = \\min\\{q \\ge 5 : q \\mid n(n + 2)\\}$ with key 0 for survivors of every fold $\\le 1499$; fold $p$ deletes the positions with key $p$; a kill run is a maximal block of positions adjacent at level $p$ and all killed at $p$; $X_p = \\mathrm{kills} - \\mathrm{runs}$ is the number of adjacent kill pairs; $\\bar m_p$ is the Mertens-product mean spacing before fold $p$, $6\\prod_{5 \\le q < p} q/(q-2)$, a model quantity; the window's empirical mean gap $Y/N_p$ depends on $Y$ and on the anchor, since periodicity fixes the full-period density and not every partial-period count (in $T_7$, windows of length 60 at anchors 0 and 18 contain 5 and 4 points, with empirical means 12 and 15 against the periodic 14), and the first draft's \"proven, no dependence\" is withdrawn. The producer distinguishes the two and reports their discrepancy (1.00000 at folds 5, 13, 29 and 1.00050 at fold 421, larger at 1009); $\\mathrm{kills} = (Y/\\bar m_p)(2/p)$ below is the model exposure, not an exact count in every window. The two end gaps are excluded. A window is not cyclic and has no copies, so the straddle question of the tile does not arise.\n\n**The law (measured).** For folds $p \\ge 100$,\n\n$$X_p \\sim \\mathrm{Poisson}\\big(\\lambda_{\\mathrm{model}}(p, Y)\\,M_p\\big), \\qquad \\lambda_{\\mathrm{model}} = \\mathrm{kills}(Y, p)\\,A\\,\\exp(-c\\,\\theta_p/\\bar m_p), \\qquad \\mathrm{kills}(Y, p) = \\frac{Y}{\\bar m_p}\\cdot\\frac{2}{p},$$\n\nwith $(A, c) = (2.4312 \\cdot 10^{-2}, 1.0818 \\pm 0.0317)$ fitted once at $Y = 2 \\cdot 10^9$, and with four clauses that are part of the statement:\n\n1. *The one-parameter form is refuted as a fitting model.* With $A = 1$, which is the law as the amortized record first wrote it, the fit gives $c = 1.9496$ and puts the last pair at fold 277 against the truth 421; the crossing calibration gives $c = 1.22$; two calibrations of the same shape disagree by 60%, and the amplitude is what reconciles them. The mechanism produces an amplitude, since a gap qualifies in three residues out of $p$ among multiples of 6.\n2. *The law over-predicts the counts by about 20%.* Measured over predicted $N_2(p \\ge 100)$, the number of folds with a run of length at least 2, reads 0.75, 0.82, 0.87, 0.82, 0.86 at five windows; the pre-registration of the fifth window fixed in advance that a fifth confirmation of this systematic puts it inside the law as a known overprediction rather than inside the acceptance band a sixth time.\n3. *The per-fold dispersion is Poisson around $M_p$, not around the smooth law.* The registered clause $|X_p - \\lambda_{\\mathrm{model}}| \\le 3\\sqrt{\\lambda_{\\mathrm{model}}}$ at 90% of folds fails: 43.2% at $Y = 2 \\cdot 10^{11}$ (32 of 74 folds with $\\lambda \\ge 1$), 78.4% at the blind anchor of §7.4, and it is retired for future larger-$Y$ pre-registrations with a scope clause, since it was never separately refuted at $Y \\le 2 \\cdot 10^8$. The failure mechanism is the first-moment shift $(M_p - 1)\\sqrt\\lambda$, which grows with window length. Around $\\lambda_{\\mathrm{model}}M_p$ the dispersion is Poisson: the anchor-replicate $\\chi^2/\\mathrm{df}$ is 1.17 where independent overdispersion would give about 12, and the cross-window fold-factor deviance is 0.98 (measured, `research/history/staging/perfold-error-model.md` §2, confirmed on an independent sieve in `research/history/staging/redteam-0820-empirical.md` §T1.a).\n4. *Runs longer than 2 do not follow the rate.* Chaining the same rate to runs of length $\\ell$ predicts 8.3 folds at $p \\ge 100$ with $L \\ge 3$ at $Y = 2 \\cdot 10^9$; measured, zero at the four scaling windows, and one at the fifth window, where the C1 row has $L = 3$ at $p = 103$ (its totals $S_1 = 70$ against $N_2 = 69$ already force it; the first draft's \"all window runs at $p \\ge 100$ have length at most 2\" was false). A run of three needs two consecutive qualifying gaps in opposite classes, and the independence model does not know the alternation cap (`research/attack-foldL-01-census.md` §1). $S_1 = \\sum_{p \\ge 100}(L - 1)$ is therefore predicted equal to $N_2$, and no prediction is made below $p = 100$, where the model predicts $L = 11$ at fold 7 against the truth 2.\n\nThe mechanism under the law is an exponential density for the chance that a killed slot has a killed neighbour, at a rate set by $\\theta_p$ against $\\bar m_p$, a statement in the neighbourhood of $H''$ and not $H''$ itself (§4.2). $H''$ is unproven and the law does not prove it.\n\n### 7.2 The pre-registered test, five windows\n\nThe predictive machine, the bands (10th to 90th percentile of the last-pair distribution under $P(\\mathrm{last} \\ge p) = 1 - \\exp(-\\sum_{q \\ge p}E[X_q])$, unioned over $c \\in [c - 3\\,\\mathrm{se}, c + 3\\,\\mathrm{se}]$ with $A$ profiled), the count bands (prediction × [0.7, 1.3], then ±2 Poisson standard deviations), and the kill criterion were fixed before any new window was sieved (`STAGE=predict node research/attack-foldL-06-scaling.js`, deterministic, its text reproduced line by line in the full run). The fifth window's pre-registration was committed alone before its producer existed (§10 lists the commits).\n\n| $Y$ | last fold with $L \\ge 2$ | survival median | band | in band | $N_2$ measured / predicted | $S_1$ measured |\n|---|---|---|---|---|---|---|\n| $2 \\cdot 10^7$ | **181** | 211 | [163, 311] | yes | 8 / 10.6 | 8 |\n| $2 \\cdot 10^8$ | **331** | 311 | [241, 439] | yes | 21 / 25.7 | 21 |\n| $2 \\cdot 10^9$ | *421* | *431 (calibration)* | *[347, 577]* | | *37 / 42.6* | *37* |\n| $2 \\cdot 10^{10}$ | **457** | 557 | [457, 719] | yes | 50 / 60.8 | 50 |\n| $2 \\cdot 10^{11}$ | **631** | 683 | [571, 877] | yes | 69 / 80.1 | 70 |\n\nVerdict under the pre-registered rule: CONFIRMED at the first three test windows (`research/history/staging/attack-foldL-06-scaling.md` §5) and HIT on all four criteria at the fifth (`research/history/staging/foldL-window5.md` §1), the sequence 181, 331, 421, 457, 631 strictly increasing across four decades of $Y$. The record's own literal crossing arithmetic, the last fold with $2c\\,p/\\bar m_p \\le \\ln\\mathrm{kills}(Y, p)$ at $c = 1.22$, predicts 233, 317, 523, 653 against 181, 331, 457, 631, within 22%, 4%, 13% and 3.4%. Four things keep this from being a comfortable pass and all four were visible in the pre-registration: the model is hot in one direction at every window; the $2 \\cdot 10^{10}$ measurement sits exactly on the floor of its band, which is the 10th percentile of the predicted distribution; the deep decade [300, 1000) is over-predicted by about 2 at the two largest windows (3 against 6.7, 13 against 23.8; 32 against 43 at the fifth); and the bands are wide. The fifth window's full sweep to $p = 2999$ puts the last fold whose $\\theta_p$ fits under the window's record gap at 1039, so the ceiling of 1499 never bound and the extinction is arithmetic rather than a sweep artifact.\n\nThe full-window totals over all 237 folds to $p = 1499$, no prediction attached below $p = 100$: folds with $L \\ge 2$ are 29, 42, 58, 71 at the four scaling windows, $\\sum(L - 1) = 31, 46, 64, 81$, maximum $L = 3, 3, 3, 4$, and $G_2$ at fold 1499 is 1458, 1560, 2220, 2220.\n\n### 7.3 The derived constants, and why there is no Stein derivation\n\nThe amplitude the scaling record calls fitted and not derived is now a computed first moment (`research/history/staging/import-stein.md` §2.4; measured). Randomising the fold's offset $u \\in \\mathbb{Z}/p$ uniformly, which the Merge Rate Identity licenses since $u = 0$ is the arithmetic truth and $W(0) = X_p$ at 237 of 237 folds (verified), the expected number of adjacent kill pairs at fold $p$ is $\\lambda_p = \\sum_\\alpha \\omega_\\alpha/p$ over the qualifying gaps of the window's level-$p$ word, with no fit and no model of the gap word. It reproduces the measured $X_p$ fold by fold: $\\sum X/\\sum\\lambda = 0.914, 1.009, 1.046, 1.009$ at the four windows, and $|X_p - \\lambda_p| \\le 3\\sqrt{\\lambda_p}$ at 100%, 100%, 97.1%, 100% of folds with $\\lambda_p \\ge 1$. Fitted the record's own way, the derived curve gives\n\n| $Y$ | $A$ from $\\lambda_p$ | $c$ from $\\lambda_p$ | $A$ fitted on $X$ | $c$ fitted on $X$ |\n|---|---|---|---|---|\n| $2 \\cdot 10^9$ | $2.2091 \\cdot 10^{-2}$ | 1.0701 | $2.4312 \\cdot 10^{-2}$ | 1.0818 ± 0.0317 |\n| $2 \\cdot 10^{10}$ | $2.2010 \\cdot 10^{-2}$ | 1.0698 | $2.1867 \\cdot 10^{-2}$ | 1.0662 ± 0.0101 |\n\nso $c$ lands 0.37σ low and $A$ at a factor 0.909 at the calibration window, and 0.36σ and a factor 1.007 at the window with the tighter error bar. Against it the geometric null gives $A = 4.7843 \\cdot 10^{-2}$, a factor 1.97 high, and $c = 1.0577$, 0.76σ low: the first moment that counts the window's actual qualifying gaps beats the null that models the gap word as geometric on both coordinates. The proposal's upgrade trigger named two amplitudes, the law's and the null's, and only the first is derived; the null's factor of two stands, and the grade stays WEAKENED on the trigger's narrow reading (`paper/proposals/prop-thinning-null.md` §5, scored 2026-08-20). $\\lambda_p$ is derived from the window's own word, so it is a derivation of the rate law's shape and constants from data the record also had, not a forecast of a new window.\n\nWhat the first moment does not come with is an error term. Arratia, Goldstein and Gordon's Poisson approximation needs $b_3$, the dependence between an event and the occurrences outside its neighbourhood, to be small, and here every indicator is a function of the single shared offset $u$, so the indicators outside any proper neighbourhood reconstruct it and $b_3$ sits at its ceiling $\\sum_\\alpha 2p_\\alpha(1 - p_\\alpha)$: measured at 0.8503, 0.9834, 0.9984, 0.9998 of that ceiling at the four windows, 200 to 800 times $b_1 + b_2$ (`research/history/staging/import-stein.md` §2.3; the AGG formulas verified against the authors' 1990 restatement in Statistical Science 5, read as page images). The one neighbourhood choice that makes $b_3 = 0$ costs $b_1 = \\lambda^2$ and is informative only for $\\lambda \\lesssim 1/4$, below the extinction fold; so the extinction band's Poisson assumption is licensed deep and unlicensed exactly where it is used (P4, refuted at the two large windows). Three qualifications the first draft dropped are restored. The Stein producer approximates some conditioning cells as singletons and states an aggregate error ceiling $2b_1(N_1)$ of $0.3069, 3.172, 31.61, 315.0$ at the four windows; the quoted $b_3$ ratios are therefore not exact evaluations, and the qualitative large-$b_3$ obstruction is what the numbers support. A single shared offset does not mean that every proper complement reconstructs it, nor that $b_3$ equals its ceiling: the equality condition is measurability of the particular indicator with respect to its outside $\\sigma$-field, for which reconstructing the whole offset is sufficient and not necessary (on uniform $\\mathbb Z/5$ with two singleton events at $0$ and $1$ and singleton neighbourhoods the exact $b_3$ is $4/25$ against a ceiling of $16/25$); the large observed ratios support an approximate reconstruction statement in the tested regime. For the all-event neighbourhood $b_3 = 0$ and $b_1 = \\lambda^2$, but $b_2 = E[W(W-1)]$ still matters, so a small mean alone does not license Poisson approximation; and multiplying zero-event probabilities across primes for the last-fold distribution needs a joint model across folds, which separate one-fold bounds do not supply. The extinction bands are pre-registered model predictions with measured scores, including the failed clause; the pre-registered kill criterion fired, and the constants are recorded as a first-moment computation and not as a Stein-derived result. One further measurement from the same producer bears on $H''$: $b_2$ under the adjacent-gap neighbourhood is exactly zero at all four windows, meaning that of the 0, 1, 24 and 221 adjacent pairs of qualifying gaps in the four windows, not one fires at a common offset at any of the $p$ copies. Two adjacent gaps can both qualify and can never both merge, which is the alternation cap measured over the whole offset ensemble rather than at the one arithmetic offset.\n\n### 7.4 The per-fold field $M_p$\n\nThe clause-3 failure has a model, and the model was tested blind twice (`research/history/staging/perfold-error-model.md`, `research/history/staging/mp-derivation.md`; both red-teamed, `research/history/staging/redteam-0820-empirical.md` §T1 and `research/history/staging/redteam-0820-night-empirical.md` §T1).\n\n**Measured.** $M_p$ is *defined* as the fold factor of the pooled fitted field, and the identity $M_p = \\lambda_{\\mathrm{derived}}/\\lambda_{\\mathrm{model}}$ is an empirical comparison with the first moment of §7.3, which holds within uncertainty at the 14 folds where both are printed (at $p = 211$, $2.189$ against $2.225$ with $2\\,\\mathrm{se} = 0.089$; six of the fourteen rows are low-information) and not as numerical equality; over the 43 best-measured folds $\\mathrm{sd}(\\ln M) = 0.560$, range $[0.19, 2.23]$. The field is the same number at every window length from $2 \\cdot 10^7$ to $2 \\cdot 10^{11}$ to within about 2 to 3% relative standard deviation, and constant across the three anchors tested to within the 6 to 10% that one replicate pair can resolve; \"deterministic\" is the surviving model class, not a measured identity in the anchor direction (the red team's corrected sentence). The exposure-weighted band means of $M$, 0.959, 0.858, 0.520 over [100, 200), [200, 300), [300, 500), concern pair-count exposure; the response of the nonzero-fold count to the field is nonlinear, $E[N_2] = \\sum_p (1 - \\exp(-\\lambda_p M_p))$ under the Poisson model, so these means do not by themselves identify the 20% overprediction of clause 2, and that identification stays an interpretation.\n\n**Blind, first anchor.** At $[6.6 \\cdot 10^{10}, 6.6 \\cdot 10^{10} + 2 \\cdot 10^9)$, pre-registered alone before the producer existed, the negative-binomial predictive $X_p \\sim \\mathrm{NB}(r = Sx_p + \\tfrac12,\\ q = \\lambda_h/(\\lambda_h + S\\lambda_p))$ from the pooled exposure of the embedded windows scored 33 of 37 folds inside the 90% bands (needed 28) and 0 of 37 outside the 99.73% bands, re-scored digit-identically from the red team's independent sieve.\n\n**Partially derived.** The field's comb is $W_1(v) = \\prod_{q \\mid v}(q - 2)/(q - 4)\\cdot\\prod_{q \\mid v \\pm 2}(q - 3)/(q - 4)$ over the folded primes $5 \\le q < p$, the exact pair-survival weight of a distance $v$ relative to a generic distance (proven, elementary residue counting: the forbidden set $\\{0, -2, -v, -v - 2\\}$ has size 2, 3 or 4), and\n\n$$M_p \\approx k\\,W_1(\\theta_p)\\,\\exp(-\\delta\\,\\theta_p/\\bar m_p), \\qquad (k, \\delta) = (1.9468, 0.2771)$$\n\nfitted on 37 folds $p \\le 293$ only, drops the residual roughness from 0.560 to 0.177 and reads a free exponent 1.031 on $\\ln W_1$ where the derivation says 1. It hit a second sealed blind test at $[1.32 \\cdot 10^{11}, 1.32 \\cdot 10^{11} + 2 \\cdot 10^9)$: 34 of 37 inside the 90% bands, 0 of 37 outside 99.73%, likelihood margin +25.4 nats over the constant-bias rival (predicted +23.6). Using $W_1$ as the gap-count weight is a modelling step, endpoints exact and interior smooth. $(k, \\delta)$ are refits of constants the law already had. A residual sub-field remains, a far-tail flattening (band [500, 710) measured 0.357 against 0.233 predicted) with four folds pulled above 3, of which only 409 and 631 stay above 3 under the formula's own Poisson-lognormal variance. Two riders from the red team travel with this. The claim that the field is a function of $\\theta_p$ rather than of $p$ rests on the matched control (non-twin fold pairs matched on predicted $M$ agree like twin pairs, rms 0.181 against 0.155), not on the raw twin contrast, which is confounded because sharing $\\theta$ and $\\Delta p = 2$ coincide. And the second producer first exists in the repository history a few minutes after its seal, a custody residual mitigated but not removed by the outcome sitting at the incumbent expectation rather than above it (§10).\n\n### 7.5 The shape of the extinction sequence is not established\n\nMoving the same window length $2 \\cdot 10^9$ to the anchor $10^{10} + 2$ moves $\\bar m$ and the kill count by two parts in a thousand, the counts by 3, and the last multi-kill fold from 421 to 349: a swing of 72 in $p$ at identical length, the only empirical measurement of the sampling spread of the extreme statistic on the real fold. The null supplies an independent estimate of the same order from many windows: its ceiling statistic has simulated spread 93.3 and 104.6 in $p$ at the two simulated windows. The per-decade increments 150, 90, 36, 174 of the measured sequence are inside that noise, and the apparent deceleration of the first four is not established by anything in the record (`research/history/staging/attack-foldL-06-scaling.md` §5; `research/history/staging/null-limsup.md` §3). The fold 631 was seen twice, at the fifth window and at the first blind anchor, a $2.8 \\cdot 10^{-3}$ Poisson event at that fold under the comb formula; $W_1(1260) = 5.55$ explains a preference for 631 and does not explain two sightings.\n\nTwo corrections to the amortized record travel with the ledger: the crossing at fold 421 is $\\ln\\mathrm{kills} = 11.569$ against $2c\\,p/\\bar m = 11.522$ at $c = 1.22$, not the 11.6 the record rounded up to; and the per-decade effective exponents 1.439, 2.435, 1.738, 1.550, 1.171 are biased low by a Jensen effect, maximum likelihood on the same folds giving 1.545, 2.594, 2.052, 1.950, 1.656, so the reading that $c$ drifts below the bracket at the closure point is that bias and not a drift. Neither changes a conclusion. One near miss belongs in any methods note: the first offset run was anchored at $10^{10}$, which is 4 modulo 6, so the walk visited multiples of 3 rather than twin slots and nothing it printed looked wrong; the engine now refuses an anchor that is not a multiple of 6.\n\n## 8. Prior art\n\nThe position is the registry's (`research/PRIOR-ART.md`; `research/SEARCH-CONVENTIONS.md` §1, rows written 2026-08-19; `research/history/staging/proposals-prior-art.md` §3; `research/history/staging/hawkins-read.md`), not a hopeful one.\n\n| result | ours | theirs | relation |\n|---|---|---|---|\n| gaps exactly geometric at every stage, parameter the Mertens product | Theorem 1 | Hawkins 1957; Neudecker–Williams 1974; Heyde 1976 p. 278 | owned by them; ours is the two-class instance |\n| composing thinnings on the generating function | Proposition 1 | Bunge 1996 (composition semigroup of pgfs; geometric case worked) | owned by them |\n| a random sieve at a general deletion rate | the rate $2/p$ | Lorch 2007 Thm 2.1 (Hawkins' $p$-primes) | adjacent; $p(n) = 2/n$ satisfies his hypotheses |\n| exact stage-to-stage moment recursion, no error term | the technique of Theorem 2 | Wunderlich 1974 eq. (5); Lorch 2007 Lem. 3.3; Bui–Keating 2006 eq. (4) | technique owned by them, applied to the null; ours applies it to the fold |\n| joint minus product of marginals, first correction named | $\\Psi - \\Phi^2$ | Bui–Keating 2006 p. 2 and Lem. 2 | adjacent; same move, different objects, theirs null-internal |\n| the deviation of the random sieve from the true sieve | $\\Psi - \\Phi^2$ as first-order deviation | Bui–Keating p. 2; Neudecker–Williams p. 199; Rivoal Thm 2 | question owned by them, answer not found in print |\n| maximal gap under the null | §3.5 | Neudecker 1975 (via Rivoal p. 808) | owned by them, constant 1 |\n| twins and $k$-tuples under the null | not used here | Wunderlich 1974 Thm 4; Bui–Keating; Rivoal Thm 1 | a theorem since 1974, no singular series |\n| the tile, the fold, the closure | §2 | Holt and Rudd 2014 (one-class recursion, Lemma 2.1) | owned by them; the two-class operator here adapts it and is not the printed statement |\n| Merge Rate and Consumption Identities; Fold Moment Identity; Corollary 1 | §2, §4 | searched under \"Hawkins' random sieve\" and \"Hawkins' p-primes\", six primaries read | not found |\n\n**What the primary read established.** Within the dated search and bibliography of `hawkins-read.md` the forward-citation walk converges on eleven items; the \"small and closed\" description is scoped to that search and is not a theorem about the literature. Nothing in it deletes residue classes; the calibrated negative for the strings \"residue\", \"congruen\", \"modul\" and \"two classes\" is zero hits in five of the six full texts read, and Rivoal p. 802 states the structural reason. The random-sieve twin conjecture is a theorem (Wunderlich 1974, Theorem 4), extended to all $l$-tuples (Bui–Keating) with an error term (Rivoal, Theorem 1), with no singular series anywhere, and the tradition records that absence as its own limitation.\n\n**What is unverified.** Daley and Vere-Jones, Feller, and Embrechts–Klüppelberg–Mikosch for the classical thinning identity are written from memory with no section cited; the first was checked and found to have no chapter titled for thinning, so it is unusable as a locator. Pyke 1965 carries volume and pages and no verification is recorded, and it is load-bearing only for the claim in §5.4 about which solvable model is nearest; the proposal's downgrade trigger fires if any of the four misdescribes the classical result. Hawkins 1957 and 1974, Neudecker 1975, Deheuvels 1985, Lorch–Ökten 2007, Williams 1974 and Wunderlich 1976 were not reached; Neudecker 1975 is the one that matters, since the limsup law of §3.5 is carried second-hand. One date is corrected in passing: the record's earlier \"Neudecker–Williams 1979\" is 1974, Compositio Math. 29 fasc. 2, submitted 30 May 1974, and no 1979 item exists in the tradition. Locators corrected in this revision: Neudecker–Williams' Markov geometric transition is on printed page 198 and the Mertens comparison on 199; Bui–Keating's detailed pair expansion is on printed pages 6 to 7 (Lemma 2), not page 2, which carries the tuple results and the missing twin constant; Arratia–Goldstein–Gordon's article spans pages 403 to 424, the ensuing discussion running to 434. Lorch and Bunge publisher downloads and Heyde's AMS download were not obtained by the referee, so those page-level checks rest on the record's read and not on the review's.\n\n**What was not searched.** $W_1$ of §7.4 has no owning-convention row; it is singular-series-shaped and nothing about it is claimed as new. The standing assumption of the registry applies: prior art exists for more of this than has been found, and the burden is on us to look again. No claim of literature novelty is made for Corollary 1; it is one line of algebra on Theorem 2, new to this corpus and not claimed beyond that.\n\n## 9. Refuted and corrected claims on this line of work\n\nKept visible, as the house rules require.\n\n1. **\"An exactly solvable null for an object that had none.\"** The proposal's headline, dead hours after filing: the object had one since 1957, under a name this corpus had mentioned zero times (§1, §8). The grade WEAKENED is that event.\n2. **R5, the decomposition of the overshoot.** Predicted mostly tail with the hazard within 30% of 1; the hazard factor is 2.8 and the tail factor 0.16 to 0.45 (§6.3).\n3. **S5, where the fold exceeds independent thinning.** Predicted near $\\theta$ and $2\\theta$; measured in the bulk at 18, 30, 42, with the far tail at or below the null (§6.3).\n4. **U2, the margin decays in $p$.** Refuted; it decays in $\\lambda$, and the conclusion survives for that reason (§5.2).\n5. **The proxy $K < 1 - 4/p$** for domination by the fold's own null map was a bad large-$p$ approximation of $K_{\\mathrm{crit}}$ and is withdrawn (§4.2).\n6. **The one-parameter rate law** $r = \\exp(-c\\,\\theta/\\bar m)$ with a single $c$ is refuted as a fitting model on its own calibration window; the run-length extension to $L \\ge 3$ is refuted, 8.3 predicted against 0 measured (§7.1).\n7. **The ±3√λ per-fold clause** is refuted as an error model at $Y \\ge 2 \\cdot 10^9$ and retired with a scope clause (§7.1, §7.4).\n8. **P3 of the amortized record**, that most large gaps are young, is refuted in the form stated and replaced by the age filter: at $T_{29}$ against fold 31 the cohorts born at folds 3 to 11 contribute nothing to the qualifying gaps while carrying 52.9% of the population, because a gap's size is frozen at birth and the largest gap fold 11 can make is 42 (`research/history/staging/attack-foldL-04-amortized.md` §4).\n9. **\"Gaps of scale $2p'$ are extreme-tail objects\"** is false at every computable fold: $\\theta_p/\\bar m$ never exceeds 2.14 on the exact ladder (same record, §9).\n10. **The amplitude trigger** of the proposal names two amplitudes; the law's is now derived (0.909 of the fit) and the null's is not (1.97), so the trigger is scored narrowly and the grade holds (§7.3).\n11. **$K$ as the mechanism of the $M_p$ field** is refuted as a field predictor (§4.3).\n12. **The Neudecker trigger** read literally would have retired a sound half on a factor of 42 that is translation (§3.5).\n13. **\"The entire first-order deviation is $\\Psi - \\Phi^2$\"** is an asymptotic reading in $p$. Corollary 1 gives the exact form $2(\\Psi - \\Phi^2) + \\Gamma - 4\\Phi(\\Phi-1)^2/(p - 2\\Phi)$, and on the exact ladder to $T_{29}$ the last term is about three times the pair term; the reading holds in the localized frame at $p \\ge 100$ (§4.2). Proposition 3 replaces the record's two separate smallness claims for $\\Omega$ and $\\Delta$ by the identity $\\Omega + \\Delta = \\Gamma \\ge 0$.\n14. **Three routes agreeing on the exponent** are three models of one measurement (§6.1).\n15. **Numbers corrected in the record:** the crossing 11.6 is 11.522; the per-decade effective exponents are biased low; the $10^{10}$ anchor walked the wrong residue class; Neudecker–Williams is 1974, not 1979 (§7.5, §8).\n16. **The M_p anchor-constancy sentence** was weakened by the red team from \"identical\" to \"within 6 to 10% at one replicate pair\", and the $\\theta$-not-$p$ claim rests on the matched control, not the raw contrast (§7.4).\n17. **First draft, three models conflated.** Theorem 1's object is the infinite renewal null; the first draft let its gap stand for the gap of a finite thinned tile, called the geometric/CRT hybrid rate the null's adjacent-kill rate (independent deletion's is exactly $2/p$; at fold 7 the hybrid is 0.130 against 0.286), and treated the convolution control as the one-fold thinning of the word (§3.1, §3.4, §6).\n18. **\"$2 - c'_{\\min} = O(\\lambda^2)$\"** is false; the margin is linear, $2 - (20/7)\\lambda$ at $T_5 \\to T_7$ (§5.2, referee report 72 §2).\n19. **The composed exponent table** used a faulty interpolator in `import-thinning-03-deepfolds.js` (`cOfU`); corrected values 1.0302, 1.0589, 1.0900, 1.1167 from the rounded cells, and the table is labelled a model calculation, not a uniform bound (§5.2, referee report 72 §4).\n20. **Theorem 3's second conclusion** substituted the exponent $\\bar m/(6\\alpha')$ while keeping the $C$ of the fixed $\\lambda$; counterexample at $x = 5$; repaired by assuming domination at $\\lambda = 1/(6\\alpha')$, which Proposition 2 does (§5.3, referee report 72 §3).\n21. **\"$H''$ at $m = 2$ written as a moment\"** is withdrawn; the word $[18, 6, 12, 6, 18]$ has exponential covariance $-16/25$ and threshold covariance $+1/25$; the paper now says \"related moment condition\" (§4.2, referee report 72 §5).\n22. **\"The record's sentence is right at the depths the rate law lives at\"**: the rational term is still 40.8% to 7.6% of the pair term at $p = 101$ to $1009$, and no uniform bound on $\\Gamma$ or on the window boundary terms is supplied (§4.2).\n23. **\"The fold merges only at qualifying gaps\"** and the mutual-singularity coupling argument are withdrawn; slot 47 of $T_5 \\to T_7$ merges the gaps 6 and 12; the surviving obstruction is the fixed deletion count $2N$ against a non-degenerate count (§5.4, §6.3, referee report 72 §6).\n24. **\"So the hypothesis is at least as strong as the target and the route is closed\"**: the implication shows the hypothesis strong and unproved; it does not show it can never be established (§5.3, referee report 72 §3).\n25. **The finite-window maximum law** $P(K \\le k) = (1 - \\rho^k)^{n}$ and its Gumbel form are approximations in a fixed window, not exact laws; Neudecker's limsup does not transfer as a theorem (§3.5, referee report 72 §7).\n26. **\"$\\bar m_p$ carries no dependence on $Y$ or the anchor, proven\"** is false for the empirical mean; the model mean and the empirical mean are now distinguished (§7.1, referee report 72 §8).\n27. **\"All window runs at $p \\ge 100$ have length at most 2\"** is false at the fifth window ($p = 103$, $L = 3$); \"at every exact fold except 23\" is scoped to folds 7 to 29 (§4.1, §7.1).\n28. **$M_p = \\lambda_{\\mathrm{derived}}/\\lambda_{\\mathrm{model}}$ \"at 14 of 14 folds\"** was a definition stated as an equality; it is an agreement within uncertainty; the 20% overprediction is not identified by exposure-weighted band means; the Stein error ceilings and the $b_2$, joint-model and measurability qualifications are restored (§7.3, §7.4, referee report 72 §8).\n29. **Locators and tolerances**: Neudecker–Williams p. 198, Bui–Keating pp. 6 to 7, AGG 403 to 424; the Corollary 1 residual $2.7 \\cdot 10^{-14}$ is absolute (relative maximum $7.8 \\cdot 10^{-13}$); the fifth-window pre-registration commit is cited as `4391c2c` here and as `7038e0b` in the served producer's retrospective note, a relationship not yet reconciled (§8, §10, referee report 72 §9).\n\n## 10. Reproduction\n\nEvery number in this paper is read from an embedded OUTPUT block or a staging record that names its producer, or from the check run made for this paper whose log is listed here.\n\n- `node --max-old-space-size=8000 research/import-thinning-01-nullmodel.js` (20.5 s): the group law, the null law against the recursion, $c_{\\mathrm{null}}$, the R1 to R5 predictions (`STAGE=predict` stops before any measurement, 0.1 s), the fits of §6.1, the decade table of §6.2.\n- `node --max-old-space-size=8000 research/import-thinning-02-coalescence.js` (19.6 s; `DEEP=1` adds fold 31): the exact-tile columns of §2, the Fold Moment Identity at seven folds, $\\Delta$, $K$, $c'_{\\min}$ on the tile, the total-variation table of §6.3.\n- `node --max-old-space-size=8000 research/import-thinning-03-deepfolds.js` (21.0 s): the localized $c'_{\\min}$ and $K$ to $p = 1009$, the hazard table, the composition of §5.2.\n- `node research/attack-foldL-04-genealogy.js` (41 s) and `node research/attack-foldL-04-localized.js` (25 s): the spend ledger and the window closure.\n- `node research/attack-foldL-06-scaling.js` (95 s, all windows; `STAGE=predict` 8 s): the pre-registration and the four-window test; `research/foldL-window5-01-extinction.js` (1228.7 s): the fifth window.\n- `node research/import-stein-01-multikill.js` (386.5 s): $\\lambda_p$, the derived $(A, c)$, the $b$-table.\n- `node research/qc/embed.js --check` on `research/attack-perfold-01-error-model.js`, `research/attack-perfold-02-blindwindow.js`, `research/attack-mp-derive-01.js`, `research/attack-mp-derive-02-anchor.js`: the field, the two blind tests.\n- `node research/null-limsup-01-score.js`: the null's maximal gap against Neudecker's law.\n- `python3 job68-check.py --max-fold 29 > job68-check.log` (Python 3 and numpy, 22 s, about 2 GB at fold 29; the version published with return #183, file `6e1324b2…`, which moves the six timing lines to stderr and replaces the wall-clock gate on fold 29 by the explicit `--max-fold` parameter, so that stdout is the results computed and nothing else; return #27's original `cbba1128…` and its log `c689daf5…` remain on record), written for this paper: the group law as an exact polynomial identity and at 36 random points; the null law against the recursion to fold 29 (§3.3 gives the figures); $c_{\\mathrm{null}}$ at folds 7 and 1499 and the refit of §6.1, reproduced to the printed digits with $\\bar m$ before the fold and $\\theta_p = 2p - 2\\eta$ (six other conventions printed, none matching); the exact tiles $T_5$ to $T_{29}$ built by the fold, with $T_7$ to $T_{19}$ also sieved directly and found equal; the Merge Rate Identity on the whole word at every fold; the Consumption Identity and the full spend ledger of §2; the Fold Moment Identity at 28 cells with residual $3.8 \\cdot 10^{-16}$; $\\Gamma = \\Omega + \\Delta$ to $7.1 \\cdot 10^{-15}$ with every $F_R \\ge 0$; Corollary 1 to an absolute $2.7 \\cdot 10^{-14}$ (relative maximum $7.8 \\cdot 10^{-13}$); the $\\Delta$ shares, the 28-cell $c'_{\\min}$ table to the record's four decimals, and the range of $K$. Every exact and integer result of that script reproduces identically across machines (folds 7 to 23 checked on two platforms in return #183); tasks 2 and 3 print raw float64 roundoff magnitudes, which differ between numpy builds, so a byte comparison of stdout is not portable and the check compares named invariant values with declared tolerances. The referee's `small-checks.py` and `interpolation-check.py` (artifacts of report 72) check the linear margin, the Theorem 3 counterexample, the covariance word and the interpolator on the captured cells.\n\nCustody of the pre-registrations cited in §7, as the records carry it: the scaling test's predictions are the deterministic output of `STAGE=predict`; the fifth window's pre-registration is commit `4391c2c`, alone, while the served producer's retrospective source note names `7038e0b`, a relationship this revision has not reconciled and does not treat as independently checked; the first fold-factor blind test's is `199dd33` (10:38:46), with its producer first present at `f0eb201` (10:50:44); the second's is `303711b` (17:16:41), with its producer first present at `254b689` (17:22:02), the 5 minute 21 second gap being the residual §7.4 names.\n\nThe tile ladder in every producer is generated from the mod-6 comb by the fold, never sieved, and reproduces $N = \\prod(q - 2)$ and the $G_2$ ladder inside the script; the window engines reproduce the scaling record's embedded tail digit for digit and abort on disagreement.\n\n## References\n\nLocators are those the research record carries; a locator identifies the evidence and does not mean the cited page was re-read for this draft unless the record says it was read at source.\n\n- D. Hawkins, The random sieve, Math. Mag. 31 (1957/58) 1–3; Random sieves, II, J. Number Theory 6 (1974) 192–200. Not reached; construction quoted in four items read at source.\n- W. Neudecker and D. Williams, The \"Riemann hypothesis\" for the Hawkins random sieve, Compositio Math. 29 (1974) 197–200; the Markov geometric transition on printed p. 198, the Mertens comparison on p. 199. Read in full (numdam; page images re-checked in referee report 72).\n- W. Neudecker, On twin \"primes\" and gaps between successive \"primes\" for the Hawkins random sieve, Math. Proc. Cambridge Philos. Soc. 77 (1975) 365–367. Abstract only; the limsup law carried via Rivoal p. 808.\n- M. C. Wunderlich, A probabilistic setting for prime number theory, Acta Arith. 26 (1974) 59–81, eq. (5), Theorem 4. Read in full.\n- C. C. Heyde, On asymptotic behavior for the Hawkins random sieve, Proc. Amer. Math. Soc. 56 (1976) 277–280, p. 278; A log log improvement to the Riemann hypothesis for the Hawkins random sieve, Ann. Probab. 6 (1978) 870–875. Read in full.\n- H. M. Bui and J. P. Keating, On twin primes associated with the Hawkins random sieve, J. Number Theory 119 (2006) 284–296; arXiv:math/0607196v3: p. 2 the tuple results and the missing twin constant, p. 4 eq. (4), pp. 6–7 Lemma 2 and the pair expansion. Read in full.\n- J. Lorch, A generalized Hawkins sieve and prime k-tuplets, Rocky Mountain J. Math. 37 (2007) 533–550, Theorem 2.1, Lemma 3.3, §6. Read in full.\n- T. Rivoal, On the distribution of Hawkins' random \"primes\", J. Théor. Nombres Bordeaux 20 (2008) 799–809, pp. 800–802, 808. Read in full.\n- J. Bunge, Composition semigroups and random stability, Ann. Probab. 24 (1996) 1476–1489. Cited from the registry's convention row; not read at source.\n- R. Pyke, Spacings, J. Roy. Statist. Soc. Ser. B 27 (1965) 395–449. Unverified.\n- D. J. Daley and D. Vere-Jones, An Introduction to the Theory of Point Processes; W. Feller, An Introduction to Probability Theory and Its Applications, vol. II; P. Embrechts, C. Klüppelberg and T. Mikosch, Modelling Extremal Events. All three cited from memory in the record with no section; unverified.\n- R. Arratia, L. Goldstein and L. Gordon, Poisson approximation and the Chen–Stein method, Statist. Sci. 5 (1990) 403–424 (article; discussion to 434), Theorem 1 on p. 406 and the definitions of $b_1, b_2, b_3$ on p. 405 (read as page images); the Ann. Probab. 17 (1989) 9–25 original not reached.\n- Referee report 72 on return #27 and its artifacts (`small-checks.py`, `interpolation-check.py`); return #183 (the reproducibility fix of the check script). Platform records this revision builds on.\n- F. B. Holt and H. Rudd, Eratosthenes sieve and the gaps between primes, arXiv:1408.6002 (2014), §2 Lemma 2.1 (the recursion); verified verbatim in `research/history/staging/lit-pdf-holt-rudd.md`.\n- H. Diamond and H. Halberstam, the dimension-2 sifting limit $\\beta_2 = 4.26645$, as recorded in `research/dhr-verification.md`.\n\n## Authorship and AI disclosure\n\nSole author: Chris Benjaminsen. The framework, vocabulary, and driving questions are the author's, developed over six years of independent work. Formal derivations, literature audits, computations, and manuscript drafting were carried out using AI assistants under the author's direction. Computations have reproducible code and recorded outputs; asymptotic arguments require their stated mathematical inputs and are not proved by finite checks. Refuted intermediate claims are retained in the record.\n\n"}