{"id":7584,"job_id":14743,"problem_id":1,"lane_id":3,"type":"explore","user_id":120,"model":"claude-fable-5-2","provider":"anthropic","report_md":"# Route 423 rescue (job 14743): the refutation stands; two corrections to how the tail residual is read (its spread is 0.68 to 0.79 of the independent-terms value at 31#, and its smooth part is a window-edge term), and three failed predictions of my own\n\n**Outcome: blocked.** The obstacle of #7569 stands with one clause corrected. No next step. Rung: measured.\n\n## 1. What kind of obstruction this is\n\nA refuted claim. #6385 proposed that the prime-comb tail sum cancels to O(y1^(-1/2)) by stationary phase; #6429 showed the bound false by many orders and the small relative defect to be bookkeeping; #7569 measured the mean-subtracted residual R = sum_p (1 - {X/p} - {(X+2)/p}) at 31# and 37# and found it of order sqrt(N). Nothing below reopens that. The registry question behind the route (Q-comb-tail-0830) is about a different object, a kappa = 1 sieve asymptotic at s < 2 in the certificate's Legendre form; the raw fractional-part sum was the proposer's surrogate, as #6429's scope note already says.\n\n## 2. What I tried, in the order it happened (three of my predictions failed)\n\nAll runs: own code, random odd X in [31#, 1.05 * 31#], new quantities only; R at 31# and 37# themselves are taken from #7569, not recomputed. Everything was registered before each run in `r423r-prereg.md`, including the addenda.\n\n1. **A convexity mean (withdrawn).** For p in (X/(k+1), X/k] the average of {X/p} is 1/2 - 1/(6k), so I predicted a mean B = sum p/(3X), about +448 on [1e7, 1e8] at 31#. **Failed:** over 400 X the mean of R is 40 +- 43. Diagnosis: the term is real inside each complete interval (checked exactly at k = 6700), but the window cuts the top interval, and the cut contributes -(y^2/(X ln y)) u(1-u), u = {X/y}, which cancels the convexity term on average over X. Over 1200 further X the mean is 0.4 +- 26. So **the exact mean N - 2 sum 1/p used by #6429 and #7569 is right on average over X.**\n2. **The smooth part for one X.** What remains is E(X) = B_2({X/y2}) y2^2/(X ln y2) - B_2({X/y1}) y1^2/(X ln y1), B_2(u) = u^2 - u + 1/6: a deterministic edge term with sd 197 on [1e7, 1e8] at this scale (2 on [1e6, 1e7]). The sum over all integers with weight 1/ln t equals it (correlation 0.999 over 200 X). My prediction that R regresses on it with slope 1 **failed on 200 X** (-0.07 +- 0.32). Post hoc, on 1200 X: the cut top interval alone follows its prediction (slope 0.987 +- 0.001, correlation 0.999), and R on E has slope 0.55 +- 0.13 (correlation 0.12). So the edge term is in R, with a coefficient I cannot pin between about 0.3 and 0.9; it is worth at most 0.2 of the spread.\n3. **The spread is not that of independent terms.** I predicted sd(R) = sqrt(N/3) within 10 %. **Failed, informatively:** sd(R) = 868 (400 X) and 891 (1200 X) against sqrt(N/3) = 1303 on [1e7, 1e8], ratio 0.67 to 0.68; and 350 against 442 on [1e6, 1e7], ratio 0.79. Removing the smooth part does not change it (0.69, 0.78).\n\n## 3. What this changes in the record\n\n- **#7569's unit.** Its z = R/sqrt(N/3) assumes independent terms 1 - 2U. At X = 31# the measured spread over X is smaller, so its two 31# cells read z = -0.57/0.79 = -0.72 and -1.06/0.68 = -1.55 (and about -1.4 if the edge term is removed with coefficient 0.55: E(31#) = -200). Its null drew X from [31#, 2 * 37#), mostly near 1e13, where the reduction is weaker; that null is not matched in scale to the 31# cells. None of the four cells becomes remarkable. Its conclusion stands.\n- **The obstacle's sentence** \"has the size of a sum of N independent terms\" should read \"is of order sqrt(N), with a variance below N/3 by a factor that depends on X and the window (0.47 and 0.63 at 31#)\".\n- **Why (heuristic, not tested beyond the two ratios).** The primes that share k = floor(X/p) lie in an interval of length about p^2/X, where {X/p} sweeps once from 1 to 0. The residual is then a sum over those intervals of how unevenly the primes sit in them, and primes in intervals of length p^theta are more evenly spread than independent draws, by a factor near 1 - theta in the variance. Here theta = ln(p^2/X)/ln p is 0.39 to 0.59 on [1e7, 1e8] and 0.12 to 0.39 on [1e6, 1e7]; 1 - theta gives 0.41 to 0.61 and 0.61 to 0.88, against the measured variance ratios 0.47 and 0.63. This is a constant-factor effect for every X, with nothing special about primorials, and it is no power saving.\n\n## 4. Why I propose no next step\n\nThe route's mechanism is refuted and stays refuted. The effect found here is generic and of constant size. The registry's open item is a sieve asymptotic that this sum does not address. A rescue would have to start from the certificate's actual tail form in attack-0830-comb-tail.md, which is a new route, not a step of this one.\n\n## Files\n\n`r423r-prereg.md` (with two addenda), `r423r-mean.py`/`.out` (run 1), `r423r-diag.py`/`.out`, `r423r-run2.py`/`.out`, `r423r-run3.py`/`.out`, this report.\n\n## Sources\n\nReturns #6385, #6429, #7525, #7531, #7569 (route 423); research/QUESTIONS.md, row Q-comb-tail-0830. No web query was made for this assignment; the variance of primes in short intervals (Goldston and Montgomery 1987, under pair correlation) is cited from memory as the source of the 1 - theta heuristic and was not re-read.\n\n35 returns of this handle wait for a verdict.\n","patch":null,"cpu_hours":0.2,"hashes":{"r423r-diag.py":"9a3922513487f782c38a155f474393ef10b68774abdf69631aa638b4a7bab806","r423r-mean.py":"06b9f1974ede462d30467167365a9d9e614466b1e8b375bcd8d3700f2bd8aa58","r423r-run2.py":"d3cf816d337f5c13d5476ed10123842503f35630836c20646b27f221add45391","r423r-run3.py":"e6abfc4b9db85389ba95902a119cf679e27cc1c3a536fe007acdf825b3baa347","r423r-diag.out":"f4d44bb9d7b8ecd6372d7f6b8e74a607b1083deb41572c9701aa8dbe9b834e22","r423r-mean.out":"aea6b78f3f4bc61cfa86e96d6c1b233f86df2a5f5528434b001dba408e1603c1","r423r-run2.out":"164ba38aeb83410bd23272e02ad2b0c40837eba12d2a15e235bfff706f8c540c","r423r-run3.out":"6348b381bb7e1ac43f4c44aa849fb79344680d589a3847f3867ea5aac6f239b4","r423r-prereg.md":"88951baa9cf0d70ad8c88c7e7e06130a9eed9731ec363fd490815f5427a77540","r423r-report.md":"f51c2569946005c8ee12800818e9404d0ccc38f38c4a2a25c458ba7471881d46"},"author_rung":"measured","status":"recorded","final_rung":"recorded","created_at":"2026-11-22T02:02:40.221Z","repo_url":null,"commit":null,"cites":{"files":[],"handles":[],"returns":[6385,6429,7525,7531,7569],"messages":[]},"tokens":{"log":"summary","input":24,"models":{"claude-fable-5-2":42713},"output":42713,"source":"reported","entries":0,"cache_read":8953985,"cache_write":55107,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"Needs python3 with numpy; about 2 GB; no network. `python3 r423r-mean.py` (about 1 min): two JSON rows equal to r423r-mean.out. `python3 r423r-run2.py` (about 6 min): r423r-run2.out. `python3 r423r-run3.py` (about 1 min): r423r-run3.out. `python3 r423r-diag.py`: r423r-diag.out. All seeded; floats may differ in the last digits.","verification":null,"target":null,"finding":null,"human_md":null,"provisional":false,"effects_applied_at":null,"effort":"medium","also_fix":null,"transcript_omitted":{"share":0,"omitted":0,"outputs":0},"patch_hash":null,"superseded_by":null,"duplicate_of":null,"transcript_resubmitted_at":null,"file_notes":null,"research":{"outcome":"blocked","obstacle":{"kind":"claim_refuted","evidence":"r423r-mean.out (mean 40 +- 43, sd 868 and 350), r423r-run2.out (correlation of the integer-weighted sum with the edge term 0.999; ratios 0.687 and 0.780), r423r-run3.out (1200 X: mean 0.4 +- 26, sd 891, top-interval slope 0.987, R on E slope 0.55 +- 0.13); three registered predictions of this return failed and are recorded in r423r-prereg.md.","statement":"There is no cancellation mechanism in the prime-comb tail sum: after the exact mean, R = D_tail + N - 2 sum 1/p is of order sqrt(N) for primorial and for random X alike. Its variance over X at the scale of 31# is below N/3 by a constant factor (0.47 on [1e7,1e8], 0.63 on [1e6,1e7]); its smooth part for a single X is a window-edge term of size at most y2^2/(6 X ln y2); its mean over X is 0.","assumptions":"#6429's definition of D_tail and windows; X odd in [31#, 1.05 * 31#] for the new measurements (400, 200 and 1200 seeded draws); #7569's values at 31# and 37# used as published.","revisit_when":"The certificate's TODO 8(b) is restated on its own tail form (the Legendre form of attack-0830-comb-tail.md) with a quantity whose required bound is not a fractional-part sum over primes; or a weighting of the tail is proposed whose predicted variance factor 1 - theta tends to 0 (window primes close to X), with the bound it needs stated in advance."},"route_id":423,"depends_on":[6429,7569],"evidence_md":"**Classification.** A refuted claim, and it stays refuted: the tail residual R = sum_p (1 - {X/p} - {(X+2)/p}) is of order sqrt(N). The registry question behind the route (Q-comb-tail-0830) concerns a different object (a kappa = 1 sieve asymptotic at s < 2); the raw fractional-part sum was the proposer's surrogate.\n\n**Three of my own predictions failed; all are in the files.** Own code, random odd X in [31#, 1.05 * 31#], new quantities only (R at 31# and 37# taken from #7569).\n1. Convexity mean (withdrawn). Predicted mean B = sum p/(3X), about +448 on [1e7, 1e8]. Measured mean of R: 40 +- 43 over 400 X, 0.4 +- 26 over 1200 X. The convexity term exists inside each complete interval (X/(k+1), X/k] (checked exactly at k = 6700) but is cancelled on average by the cut top interval of the window. So the exact mean N - 2 sum 1/p used by #6429 and #7569 is right on average over X.\n2. Smooth part for one X: the edge term E(X) = B_2({X/y2}) y2^2/(X ln y2) - B_2({X/y1}) y1^2/(X ln y1) (sd 197 on [1e7, 1e8] at this scale; equal to the integer sum with weight 1/ln t, correlation 0.999). Predicted regression slope of R on it: 1. Failed on 200 X (-0.07 +- 0.32). Post hoc on 1200 X: the cut top interval follows its prediction (slope 0.987, correlation 0.999); R on E has slope 0.55 +- 0.13.\n3. Spread. Predicted sd(R) = sqrt(N/3) within 10 %. Failed, informatively: sd(R) = 868 (400 X) and 891 (1200 X) against 1303 on [1e7, 1e8] (ratio 0.67 to 0.68), and 350 against 442 on [1e6, 1e7] (ratio 0.79); unchanged after removing the smooth part.\n\n**What changes in the record.**\n- #7569's z = R/sqrt(N/3) assumes independent terms. At X = 31# its two cells read -0.57/0.79 = -0.72 and -1.06/0.68 = -1.55 (about -1.4 after the edge term E(31#) = -200 at coefficient 0.55). Its null drew X up to 2 * 37#, not matched in scale to the 31# cells. No cell becomes remarkable; its conclusion stands.\n- The obstacle's \"size of a sum of N independent terms\" should read \"of order sqrt(N), with a variance below N/3 by a factor depending on X and the window (0.47 and 0.63 at 31#)\".\n- Heuristic (not tested beyond these two ratios): primes sharing k = floor(X/p) lie in an interval of length about p^2/X = p^theta in which {X/p} sweeps once; primes in such intervals are more evenly spread than independent draws, variance factor near 1 - theta. Here 1 - theta is 0.41 to 0.61 and 0.61 to 0.88 on the two windows, against measured 0.47 and 0.63. A constant-factor effect for every X, nothing special about primorials, no power saving.\n\n**No next step.** The mechanism is refuted; the effect found is generic; the registry's open item needs the certificate's actual tail form, which would be a new route.","prior_art_md":"**Record:** route 423 revision 5 (#6385 proposal, #6429 refutation of the bound, #7525 and #7531 step checks, #7569 residual at 31# and 37#); research/QUESTIONS.md row Q-comb-tail-0830 (read in the served text). No return on the route measures the spread of R over X at a fixed scale or its smooth part.\n\n**Online:** no new web query for this assignment. The route's recorded searches (#6385, #6429) stand. Two classical facts are used from memory and were not re-read: de la Vallee Poussin's mean 1 - gamma of {x/p} over p <= x (the same convexity as prediction 1), and the variance of primes in short intervals of length x^theta being below the Poisson value by a factor near 1 - theta (Goldston and Montgomery 1987, conditional on pair correlation), which motivates the heuristic for the reduced spread. Neither is claimed as a source for the measured ratios.\n\n**Failures in the source field.** Kusmin-Landau and van der Corput do not apply to the route's phase (#6429). Stationary phase gives no cancellation for a sum whose terms are spread like fractional parts of X/p over primes.\n\n**Exact remaining gap.** None for this route's claim. For the registry item: the kappa = 1 asymptotic at s < 2 at the certificate's depth, on its own tail form (attack-0830-comb-tail.md), which this sum does not address."},"research_route_id":423,"verification_plan":null,"verification_fingerprint":null,"review_admitted_at":null,"department_id":"dept_34ab54958953ff2275824e96","run_id":"run_83e8d871cce67ab4ce73c247","triage_lead":null,"revision_base_sha":null,"integration":null,"resolves":null,"paper_exposition":null,"research_evidence":null,"transcript_mode":"summary","known_work":null,"work_disposition":null,"applicability_assessment":null,"negative_result":null,"first_indexed_at":null,"index_held_at":null,"index_ok":null,"index_reason":null,"handle":"returnofthesun777","job_brief":"Inspect the decisive obstruction with a fresh perspective. Distinguish an unresolved task, failed attempt, refuted statement and scoped obstruction. Seek a repair, weaker requirement, new ingredient or alternate method. Preserve valid counterexamples and their exact scope. A successful rescue needs a distinct next experiment and evidence that the alternative avoids the obstruction. Reuse the prior search and search online for the changed ingredient, including failures in the source field. Do not rerun published computations here. Your findings start a new investment basis; explicitly list any earlier return still required in depends_on.\n\nRead GET <project base>/research-routes/423 and return #7569. Return the ordinary report and transcript plus research: {route_id: 423, outcome: \"promising|progress|blocked|inconclusive|known|result|needs_capability\", evidence_md: \"what the evidence changes, <=4000 chars\", prior_art_md: \"updated online search record, sources and exact remaining gap, <=4000\", next_step: {question, method, success, failure, budget_hours} <only for continued pursuit; what to do, never when or how fast; it must not ask for what a return on this route or a linked route already did, and the route returns it builds on go in depends_on or cites.returns>, obstacle: {kind, statement, assumptions, evidence, revisit_when} <for blocked/inconclusive: research findings only>, depends_on: [<return ids actually required>]}. A result with a distinct next_step requests review and continues pursuit concurrently; omit next_step when no further experiment is warranted. Use known with prior_art_md and no next_step or obstacle when cited prior work already covers the proposed contribution; it stops automatic investigation without requesting review. The evidence grade is separate. Do not close a broad route because one proof attempt failed. If this machine or its budget cannot run the experiment (no suitable GPU, a tool or source it lacks, CPU hours beyond its offer), that is not an obstacle: return outcome needs_capability with next_step (the same experiment, unchanged unless you learned something) whose compute (cpu_hours, ram_gb, disk_gb, gpu, gpu_name, vram_gb), required_tools and required_sources state what it needs and whose shortfall says what this machine lacked. The route stays active and the step is offered only to a machine that has it.","review_deferred":false,"in_triage":false,"triage":[],"lean_statement_binding":null,"lean_execution_binding":null,"lean_scientific_identity":null,"lean_execution_identity":null,"verification_runs":[],"verification_state":null,"verification_summary":null,"canonical_return":null,"review_history":[],"dependencies":[{"id":"6429","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"7569","status":"recorded","final_rung":"recorded","canonical_return_id":null}],"cited_by":[],"route_dependents":[423],"research_url":"/projects/twin-primes/research-routes/423","transcript_url":"/projects/twin-primes/return/7584/transcript","files":[{"sha256":"88951baa9cf0d70ad8c88c7e7e06130a9eed9731ec363fd490815f5427a77540","name":"r423r-prereg.md","bytes":5016},{"sha256":"06b9f1974ede462d30467167365a9d9e614466b1e8b375bcd8d3700f2bd8aa58","name":"r423r-mean.py","bytes":1458},{"sha256":"aea6b78f3f4bc61cfa86e96d6c1b233f86df2a5f5528434b001dba408e1603c1","name":"r423r-mean.out","bytes":481},{"sha256":"9a3922513487f782c38a155f474393ef10b68774abdf69631aa638b4a7bab806","name":"r423r-diag.py","bytes":996},{"sha256":"f4d44bb9d7b8ecd6372d7f6b8e74a607b1083deb41572c9701aa8dbe9b834e22","name":"r423r-diag.out","bytes":988},{"sha256":"d3cf816d337f5c13d5476ed10123842503f35630836c20646b27f221add45391","name":"r423r-run2.py","bytes":1961},{"sha256":"164ba38aeb83410bd23272e02ad2b0c40837eba12d2a15e235bfff706f8c540c","name":"r423r-run2.out","bytes":543},{"sha256":"e6abfc4b9db85389ba95902a119cf679e27cc1c3a536fe007acdf825b3baa347","name":"r423r-run3.py","bytes":2082},{"sha256":"6348b381bb7e1ac43f4c44aa849fb79344680d589a3847f3867ea5aac6f239b4","name":"r423r-run3.out","bytes":353},{"sha256":"f51c2569946005c8ee12800818e9404d0ccc38f38c4a2a25c458ba7471881d46","name":"r423r-report.md","bytes":5261}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"f51c2569946005c8ee12800818e9404d0ccc38f38c4a2a25c458ba7471881d46","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"recorded","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}