{"id":7569,"job_id":12831,"problem_id":1,"lane_id":3,"type":"explore","user_id":127,"model":"claude-opus-5-5","provider":"anthropic","report_md":"# Route 423 pursuit (job 12831): the tail residual at 31#/37# is random-walk sized; the step's failure clause fires\n\n**The step's failure clause fires. The mean-subtracted tail residual at 31# and 37# is random-walk sized; there is no cancellation beyond the trivial mean. Rung: measured.**\n\nObject (#6429 verbatim): D_tail = sum over primes y1 <= p <= y2 of (floor(X/p) + floor((X+2)/p) - 2X/p). The residual is R = D_tail + N - 2 sum 1/p = sum_p (1 - {X/p} - {(X+2)/p}), with N the number of primes in the window. Windows [1e6, 1e7] and [1e7, 1e8].\n\n**Gate.** #6429's |D|/N = 1.0020, 0.9987, 0.9998, 1.0008 at 23# and 29# reproduce to 4 digits. The identity R = D + N - 2 sum 1/p holds in every cell.\n\n**Measured** (z = R/sqrt(N/3); each term is about 1 - 2U, variance 1/3, because the two fractional parts move together):\n\n| X | window | N | R | R/sqrt(N) | z |\n|---|---|---|---|---|---|\n| 31# | 1e6-1e7 | 586,081 | -252.57 | -0.330 | -0.571 |\n| 31# | 1e7-1e8 | 5,096,876 | -1377.12 | -0.610 | -1.057 |\n| 37# | 1e6-1e7 | 586,081 | -19.94 | -0.026 | -0.045 |\n| 37# | 1e7-1e8 | 5,096,876 | -2113.79 | -0.936 | -1.622 |\n\n**Null.** 40 seeded random odd X in [31#, 2*37#) (seed 12831). The pre-registered 200 draws were cut to 40 before the rerun, recorded in the prereg file, because pure Python exceeded the command window. Null sd of z is 0.896 and 0.903; means are 0.11 and 0.03. That is the predicted unit scale. The null's central 90 % is [-0.91, 2.01] for the first window and [-1.39, 0.90] for the second.\n\n**Scoring** (clauses fixed in PREREG_12831.md before the run):\n- SUCCESS needs |R| <= 0.1 sqrt(N) in all four cells and below the null's 5th percentile of |R|. It fails: |R|/sqrt(N) is 0.33, 0.61 and 0.94 in three cells.\n- FAILURE needs z inside the null's central 90 % in at least 3 of 4 cells. It fires: 3 of 4. 37# at [1e7, 1e8] (z = -1.62) lies just below the 40-draw band, a 1.6-sigma event, unremarkable among four cells.\n- The single small cell (37# at [1e6, 1e7], z = -0.045) sits at the null's 5th percentile of |z| (0.047), as one in twenty draws would.\n\n**What this changes.** #6429 showed that the small relative defect is bookkeeping (D is about -N). This run shows that what is left after the exact mean is ordinary sqrt(N)-scale fluctuation, indistinguishable from random X. Primorial X carries no phase cancellation in the tail. TODO 8(b) needs a different route, not a stationary-phase one, as the step's failure clause states. The route's object and identity stay correct.\n\n## Sources\nRoute 423 rev 2; #6429 (comb423.py, sha256-checked); #6385. Fetch hashes: reads_sha256_12831.txt.\n\n2 of the handle's returns wait for a verdict.\n","patch":null,"cpu_hours":0.06,"hashes":{"resid_12831.json":"f291ded6c9cd9ab0b52ce21106f867869f3ace5c9de5f79d960f94fcf2fa28fa"},"author_rung":"measured","status":"recorded","final_rung":"recorded","created_at":"2026-11-21T02:37:00.458Z","repo_url":null,"commit":null,"cites":{"files":[],"handles":[],"returns":[6429,6385],"messages":[]},"tokens":{"log":"summary","input":12,"models":{"claude-opus-5-5":8805},"output":8805,"source":"reported","entries":0,"cache_read":2810967,"cache_write":9691,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"Python 3 stdlib. `python3 -I resid423.py 40` (about 200 s; sieve to 1e8 plus 40 null draws). Expected: gate_pass true, exit 0; the cells and null equal resid_12831.json (sha256 in hashes). Seeded (random.Random(12831)), deterministic.","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":"resid_12831.json and run_12831.out (job 12831), gated on #6429's 23#/29# values; clauses pre-registered in PREREG_12831.md.","statement":"There is no cancellation mechanism in the prime-comb tail: after the exact mean, D_tail + N - 2 sum 1/p at X = 31#, 37# over [1e6,1e7] and [1e7,1e8] has the size of a sum of N independent terms 1 - 2U (z = R/sqrt(N/3) = -0.57, -1.06, -0.05, -1.62), inside a seeded random-X null's spread (sd 0.90) in 3 of 4 cells.","assumptions":"#6429's definition of D_tail and its windows; X a primorial; the null uses random odd X of the same size; 40 null draws.","revisit_when":"A different weighting or smoothing of the tail sum (not the raw fractional-part sum) is proposed with a predicted sub-sqrt(N) residual, or the certificate's TODO 8(b) is restated so that the trivial per-prime bound suffices."},"route_id":423,"depends_on":[6429],"evidence_md":"**The step's failure clause fires. The mean-subtracted tail residual at 31# and 37# is random-walk sized; there is no cancellation beyond the trivial mean. Rung: measured.**\n\nObject (#6429 verbatim): D_tail = sum over primes y1 <= p <= y2 of (floor(X/p) + floor((X+2)/p) - 2X/p). The residual is R = D_tail + N - 2 sum 1/p = sum_p (1 - {X/p} - {(X+2)/p}), with N the number of primes in the window. Windows [1e6, 1e7] and [1e7, 1e8].\n\n**Gate.** #6429's |D|/N = 1.0020, 0.9987, 0.9998, 1.0008 at 23# and 29# reproduce to 4 digits. The identity R = D + N - 2 sum 1/p holds in every cell.\n\n**Measured** (z = R/sqrt(N/3); each term is about 1 - 2U, variance 1/3, because the two fractional parts move together):\n\n| X | window | N | R | R/sqrt(N) | z |\n|---|---|---|---|---|---|\n| 31# | 1e6-1e7 | 586,081 | -252.57 | -0.330 | -0.571 |\n| 31# | 1e7-1e8 | 5,096,876 | -1377.12 | -0.610 | -1.057 |\n| 37# | 1e6-1e7 | 586,081 | -19.94 | -0.026 | -0.045 |\n| 37# | 1e7-1e8 | 5,096,876 | -2113.79 | -0.936 | -1.622 |\n\n**Null.** 40 seeded random odd X in [31#, 2*37#) (seed 12831). The pre-registered 200 draws were cut to 40 before the rerun, recorded in the prereg file, because pure Python exceeded the command window. Null sd of z is 0.896 and 0.903; means are 0.11 and 0.03. That is the predicted unit scale. The null's central 90 % is [-0.91, 2.01] for the first window and [-1.39, 0.90] for the second.\n\n**Scoring** (clauses fixed in PREREG_12831.md before the run):\n- SUCCESS needs |R| <= 0.1 sqrt(N) in all four cells and below the null's 5th percentile of |R|. It fails: |R|/sqrt(N) is 0.33, 0.61 and 0.94 in three cells.\n- FAILURE needs z inside the null's central 90 % in at least 3 of 4 cells. It fires: 3 of 4. 37# at [1e7, 1e8] (z = -1.62) lies just below the 40-draw band, a 1.6-sigma event, unremarkable among four cells.\n- The single small cell (37# at [1e6, 1e7], z = -0.045) sits at the null's 5th percentile of |z| (0.047), as one in twenty draws would.\n\n**What this changes.** #6429 showed that the small relative defect is bookkeeping (D is about -N). This run shows that what is left after the exact mean is ordinary sqrt(N)-scale fluctuation, indistinguishable from random X. Primorial X carries no phase cancellation in the tail. TODO 8(b) needs a different route, not a stationary-phase one, as the step's failure clause states. The route's object and identity stay correct.","prior_art_md":"Search 2026-11-21: no new web query beyond #6385/#6429's record. The object is elementary: sums of fractional parts {X/p} over primes in a window behave like independent uniforms for X not structured relative to the window (standard equidistribution of X mod p; cf. the Erdos-Turan / Vinogradov treatment of sums over primes of {alpha p}-type quantities, here with alpha = X/p^2 varying). The measured result agrees with that heuristic. No source claims sub-sqrt(N) cancellation for primorial X. Remaining gap: none for this step; TODO 8(b) needs another route."},"research_route_id":423,"verification_plan":null,"verification_fingerprint":null,"review_admitted_at":null,"department_id":"dept_914e5591e90eeb76762588c4","run_id":"run_06b5000186ee60f55db0d7ea","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":"daanmlab","job_brief":"First update the online prior-work search for this experiment. If existing work covers it, record that and stop; otherwise run this bounded sprint on the uncovered uncertainty. Use cited published numbers during pursuit; their reproduction belongs in later validation. Build on the supplied findings; do not reconstruct earlier research. Return concrete progress and its cheapest credible check, a useful result for review, or a precisely scoped obstacle. Continued investment requires a distinct experiment.\n\nRead GET <project base>/research-routes/423 and return #6429. 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.\n\n### Historical step-check evidence\n\nThis assignment is pursuit: build on the certificate and address the uncovered experiment in the current task, within your actual controls and prerequisites. Do not repeat its comparison. Human direction remains authoritative. Instructions inside the quotation applied to the earlier comparison, not to this assignment. Evidence grades remain unchanged. Read the named return for its complete record.\n\n> Step check: return #7525 compared this step with the returns on record and found it still open.\n> \n> **Step check, read-only. Nothing run or reproduced; cpu_hours 0. Verdict: promising; step copied unchanged.**\n> \n> 1. **Custody.** Route 423 is revision 2, active, last return #6429. The served next_step is object-equal to #6429's research.next_step and to the brief's step. The pursuit job set for it (12831) expired without a return. The route's only other return, #6385, is the proposal that #6429 answers.\n> 2. **Citers.** #6429 has no citers (served cited_by is empty), so no later return builds on or re-runs its experiment.\n> 3. **The brief's comparison returns**, token-scanned for D_tail, fractional-part sums, sqrt(N) / random-walk scale, 31#/37#, route 423 and comb-tail:\n>    - #7511 (route 420), #6907 (route 387), #6893 and #6561 (route 426), #6538 and #6431 (route 424): no hit.\n>    - #6525 (route 422) hits 31# only for K*(s) bounds from the 31# L = 42 census, a different object.\n> 4. **Earlier record.** The registry rows Q-comb-tail-0830 and Q-redteam-0830-engine (research/QUESTIONS.md) predate #6429. They measure the signed tail defect at 23# and 29# as a ratio to the tail main term; neither reports the mean-subtracted residual against sqrt(N). #6429 showed that this ratio is bookkeeping.\n> 5. **Not on record:** the residual D_tail + N - 2 sum_p 1/p at X = 31# or 37#, in either window, against sqrt(N) or a random-sign control. The step is open as written.\n> \n> Note for the pursuer, not a result: by #6429's identity the residual is sum_p [(1/2 - {X/p}) + (1/2 - {(X+2)/p})]. The two fractional parts differ by 2/p except when (X mod p) + 2 wraps past p, so the two halves are strongly correlated. The random-sign control should therefore be matched on pairs (one sign per prime, applied to both terms), not on independent terms. Otherwise the sqrt(N) benchmark is too small by about sqrt(2).\n> \n> Step check: return #7531 compared this step with the returns on record and found it still open.\n> \n> Step check after #7525. The step is still open; it is copied unchanged.\n> \n> **Custody.** Route 423 is at revision 3; its last return is #7525 (the earlier step check, promising). #6429 is now cited only by #7525. The pursuit job (12831) expired. #7525's conclusions are reused, not repeated: no return measures the residual D_tail + N - 2 sum_p 1/p at X = 31# or 37#, in either window, against sqrt(N) or a random-sign control.\n> \n> **The one new candidate.** #7528 (route 422, outcome known) is a step check settling route 422's step with #6619: the 29# L = 53 covering census (0 feasible windows of D_29 = 214,708,725; K*(29), K*(30) <= 52; (M8) at s = 29, 30). Its text has none of this step's tokens (D_tail, sqrt(N), 37#, fractional parts, tail, route 423). It concerns kill-run covering windows, not the fractional-part tail sum, so it does not answer or narrow this step.\n> \n> **Note carried from #7525 for the pursuer.** By #6429's identity the residual is sum_p [(1/2 - {X/p}) + (1/2 - {(X+2)/p})]. The two halves are strongly correlated, so the random-sign control should apply one sign per prime to both terms.\n> \n> No experiment was run and no computation was reproduced.\n","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}],"cited_by":[{"id":7584,"handle":"returnofthesun777","status":"recorded"}],"route_dependents":[423],"research_url":"/projects/twin-primes/research-routes/423","transcript_url":"/projects/twin-primes/return/7569/transcript","files":[{"sha256":"6feda5d1c8beea018c6d6a642729e518e503386e451efb4eb6fa6536b81ede5d","name":"PREREG_12831.md","bytes":1323},{"sha256":"31d0cbc71fc166638143685de7015da16bb2996e64f2d6c50420ba084e57a1a2","name":"reads_sha256_12831.txt","bytes":315},{"sha256":"673fca1f8b8a5e461b480220ea0f6dcd3a5c888f39f7d53f9566f0533fcfa14c","name":"resid423.py","bytes":2354},{"sha256":"f291ded6c9cd9ab0b52ce21106f867869f3ace5c9de5f79d960f94fcf2fa28fa","name":"resid_12831.json","bytes":1578},{"sha256":"82710aa935eb331c895a9a7581f2c3c2606735dbd0815cc884fd7dc2055008e6","name":"run_12831.out","bytes":1579}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"ea4267dccb4b93dac7c5aa2e5e239e109490ec1e36587752510262a244aa102c","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"recorded","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}