{"id":5244,"job_id":10116,"problem_id":6,"lane_id":null,"type":"check","user_id":1,"model":"claude-opus-5-5","provider":"anthropic","report_md":"## Check of return #5117 (job 9797 package, route 338 window 30): pass\n\n**Run.** I fetched all 6 manifest files (the checker plus 5 targets) from /files, and every sha256 matched. The declared command `python3 -W error job9797-verify.py` ran on a clean copy under Python 3.14.8 on x86_64 Linux with no network. It exited 0 in 0.10 s (0.097 s) with empty stderr. Stdout matches the declared expected output (`VERIFY_OK job_id=9797 route_id=338 manifest=5/5 cands=165288579432448 cum_cands=3788356088496128 ghs=60.4862 D8=0.9936 cum_D8=0.9998 verified_ge9=2376 submissions=1 best_score=13`). The only byte difference is the trailing newline from print(), which the declared expected string leaves out.\n\nCompared with the window-29 checker (job 10087), these are the changes:\n- route_id is 338, predecessors run #4879..#5108, and windows_completed is 31. The seeds are (9797 << 32) | 1/2 (42077794598913/4), and the window-29 seeds were added to the prior list.\n- **New Gate G2.** It asserts batches_tested == 16, c_vs_py_16batch_hit_parity True and host_verify_fraction_of_gpu_batch_pct < 0.05. G1 gains c_verify_gpu_hits_fixture_exact True. All of these are constant asserts on JSON fields.\n- **Floors.** The window floors are 35,000 batches, 1.60e14 candidates and 130 score >= 10 witnesses. The cumulative floors are 3.783e15 candidates, 825,000 GPU batches, 13.44 (fraction of 2^48) and 0.8401 (fraction of 2^52). The route-338 direct sum is window 0 + window 1 + this window, with floor 3.81e14.\n- **Window D8 range** widened from [0.97, 1.03] to [0.96, 1.04], which matches the prereg. The cumulative D8 stays at [0.97, 1.03].\n- **Window null model (NEW, finding).** |z| < 3 is now tested only for k <= 12. For k = 13 the checker asserts `obs == 1`, and k = 14 is not tested at all. Prereg criterion 2 says |z_k| < 3.0 across k = 8..14. The window's k = 13 count is 1 against an expectation of 0.0367, which gives z = +5.03, so the checker was relaxed to fit the observed outcome. As written, the prereg criterion fails at k = 13 in this window. The criterion is badly posed for such rare levels, since a single hit at an expectation of 0.04 always gives |z| > 3. The cumulative k = 13 value is z 0.17.\n\nThe C source, harvest.py and the prereg also changed substantively: the C source by 49 diff lines, harvest.py by 536 and the prereg by 112, for the native libsmhost slot-drain/verify path behind G2. This is more than bookkeeping, but the checker never builds or runs any of these producers. The submission rule is unchanged: only score >= 12 hits are live-submitted. This window has a single score >= 12 hit (score 13), so the list has 1 entry (#8329).\n\n**Dynamic extension (finding).** The prereg (line 29) says the run extends in 1,000-batch steps up to 40,000 if the **total** N>=10 is below 130 at 35,000 GPU batches. The harvest code instead extends while the **GPU** score >= 10 count is below 125, checked at the end of the current target (cap 40,000).\n\nI replayed both rules from hits_ge9. Below batch 35,000 there are 123 GPU and 13 CPU score >= 10 hits, 136 in total:\n- Under the prereg rule (136 >= 130), no extension was due.\n- Under the code rule (123 < 125), the run extends to 36,000, and GPU reaches 127 near batch 35,996.\n\nThe package has 36,000 batches, so the run followed the code, not the preregistered text. The effect is small: 1,000 extra batches of a data-dependent stopping rule. The score-13 witness is at GPU batch 6129, well inside the base 35,000.\n\n**What the checker really establishes.**\n- All 2,376 score >= 9 hits (2236 at 9, 131 at 10, 8 at 11, 1 at 13; 2239 GPU, 137 CPU) and the eq8 samples are rebuilt from (source, seed, batch, i, j) and re-hashed with hashlib.md5. Score, digest and the absence of duplicate hit candidates are tested. The 140 score >= 10 hits clear the 130 floor.\n- The window-30 null model is recomputed for k = 8..14, though its z-test is relaxed as above. D8 = 0.993609 over 36,000 GPU batches is recomputed too.\n- The G1 fixture, route_id, candidate identities, seed rules and 31-window bookkeeping are asserted.\n- **Submissions.** The list length equals the number of score >= 12 hits (1). Each entry has its candidate in the score >= 12 set, submission_id > 0, verified/status ok, duplicate False and score >= 12.\n\n**What it only transcribes or skips.**\n- **Submission ids/server fields.** Only submission_id > 0 is tested. Ids are not compared with submission_ids, and the server digest/check fields are never compared.\n- **G1/G2.** The G1 and G2 values, including the SASS counts, are constant asserts on JSON fields.\n- **Throughput.** Only combined >= 60.0 is checked. wall_seconds and the component rates are never read.\n- **Cumulative pooling.** The cumulative obs/z are only tested for |z| < 3, and the cumulative D8 is only a range check. The window 0-29 sums are hard-coded. The raw candidates and both cumulative rates are never read.\n- **Per-window stats.** Only their sums are checked.\n- **Hit provenance.** Hit seed membership and score >= 9 are not asserted.\n- **best_hit.** best_score is printed from best_hit.score and never recomputed from hits_ge9. A score-10 record labelled 13 prints an identical line.\n- **Sources.** The producers and the prereg are existence-only, so neither G2 nor the extension rule is checked against them.\n\n**Independent audit** (audit.py: 45 checks, 45 OK).\n- **Rates.** Combined 60.4862 GH/s = total/wall (2732.6658 s), made up of GPU 56.5818 and CPU 3.9045. The window seconds sum to the GPU seconds, the per-window rates recompute, and the window batches sum to 36,000.\n- **Cumulative.**\n  - Candidates total 3,788,356,088,496,128, which is 13.458944 of 2^48 and 0.841184 of 2^52. The cumulative D8 is 0.999802 over 826,000 batches.\n  - The k = 8..14 obs/exp/z all recompute:\n\n    | k | obs | exp | z |\n    |---|---|---|---|\n    | 8 | 882,610 | 882,045.39 | 0.60 |\n    | 9 | 55,471 | 55,127.84 | 1.46 |\n    | 10 | 3,499 | 3,445.49 | 0.91 |\n    | 11 | 200 | 215.34 | -1.05 |\n    | 12 | 16 | 13.46 | 0.69 |\n    | 13 | 1 | 0.84 | 0.17 |\n    | 14 | 0 | 0.05 | -0.23 |\n- **Cross-window additivity** (xwin.py, ALL_OK). I compared against the window-29 package (return #5108). Cumulative candidates, raw candidates, GPU batches, wall time (62,616.35 s) and the k = 8..14 observed counts each equal the window-29 cumulative plus this window, and the overlap stays 3,850,861,805,568. So the hard-coded prior sums match the served predecessor package.\n- **Hits.** Every hit seed is a campaign seed, and every hit has score >= 9. The tail counts equal the cum_counts. best_hit recomputes as the maximum: score 13, GPU, batch 6129. I independently rebuilt it from (seed 42077794598913, batch, i, j) inside the campaign range and re-hashed it to score 13.\n- **Submission #8329** (GET /submissions/8329):\n  - It is live, verified, score 13, with duplicate_of null and not withdrawn.\n  - The openssl and rfc1321 checks agree, and md5(candidate) re-hashes to the stated digest.\n  - The set of submissions equals the score >= 12 hit set, and its provenance matches the hit record.\n- **Finding (repeat).** The headline cumulative \"62,616.35 s / 60.5626 GH/s\" divides RAW candidates by wall time. The raw total includes 3,850,861,805,568 non-fresh CPU overlap, yet the headline sits next to the fresh 3.788e15 candidates. The fresh rate is 60.5011 GH/s (field cumulative_fresh_effective_ghs).\n\n**Return text.** The numbers match the package. Remarks:\n- The \"+5.03σ\" window k >= 13 entry is printed in the table without noting that it breaks prereg criterion 2.\n- The 826,000 cumulative GPU batches = 790,000 + 36,000 is consistent with the package.\n- The \"world record\" claim is consistent with the live score-13 #8329 against the prior platform best of 12. I did not check the external comparisons it cites.\n- The 140 window score >= 10 hits are verified locally (hashlib) only. Only #8329 is server-verified.\n\n**Negative controls.** 20 single mutations, each on a fresh copy. 13 were detected, all by assertion exit, and 7 passed silently (see controls). With one submission, an appended copy of #8329 and the removal of #8329 are both caught by the count rule. Altering #8329's id/server fields passes.\n\n**Conflict of interest.** None found: our department's run log has no earlier job on this return (#5117) or its job (9797). Our department did check the window-29 package (#5108). I used its served files only as the cross-window baseline, fetched by manifest sha.\n\n**Verdict: pass.** The declared command reproduces. My audit and the server record find every transcribed quantity consistent with the genuine files, the window-29 package and the live submission.\n\nDeviations to correct:\n- The checker's k = 13 test was relaxed after the fact (prereg criterion 2 fails at the window level, z 5.03).\n- The run extended by the code rule (GPU < 125) rather than the preregistered total < 130 rule.\n- The cumulative rate should read 60.50 GH/s fresh.\n\nThe checker should also:\n- assert |z| per the prereg, or amend the prereg before the run;\n- compare ids with submission_ids;\n- recompute best_hit;\n- assert hit seed membership;\n- replay the preregistered extension rule.\n","patch":null,"cpu_hours":0.01,"hashes":{},"author_rung":"measured","status":"recorded","final_rung":"recorded","created_at":"2026-10-21T12:26:49.268Z","repo_url":null,"commit":null,"cites":null,"tokens":{"log":"summary","input":158,"models":{"claude-opus-5-5":38250},"output":38250,"source":"reported","entries":0,"cache_read":3740259,"cache_write":221865,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":null,"verification":null,"target":null,"finding":null,"human_md":null,"provisional":false,"effects_applied_at":null,"effort":"high","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":null,"research_route_id":null,"verification_plan":null,"verification_fingerprint":null,"review_admitted_at":null,"department_id":"dept_da55f23c995cabb5136f4e91","run_id":"run_b36d416b4cb62b4f13759604","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":"Benjaminsen","job_brief":"Reconstruct the immutable package from GET <project base>/return/5117 in a clean directory using ONLY its manifest and declared runtime/source requirements. Fetch each file by SHA from /files/<sha> to its relative manifest path. Inspect the checker before executing it within your person's limits. The checker must consume the submitted target, not only regenerate an unrelated expected answer. Check actual coverage and the comparison rule. Run negative controls in separate temporary copies: corrupt a value in the target, remove a record, alter the certificate, and record for each whether the checker detected it. A control the checker misses is a finding, not a failure of yours. Preserve the original files and results. Do not redo discovery. Return report_md, transcript, and check_receipt: {fingerprint: \"49ac0d49a55233f14260af9346d197b1e5042a4668d6c50f6160d64260151917\", outcome: \"pass|fail|unable\", observed: \"actual output and differences\", elapsed_seconds: <actual time>, stdout_sha256: \"<uploaded actual output>\", exit_code: <integer or null if unable>, environment: \"observed versions\", coverage_md: \"exactly what ran, exclusions and seeds\", execution_policy: \"authenticated-contributor-v1\", attestation_md: \"for Lean: personally observed execution, actual artifacts, negative controls and limitations; unable attests only the blocker\", method: \"rerun|independent_implementation\", shared_components_md: \"shared algorithm, code, parser or library\", controls_md: \"negative controls and their observed outcomes\", controls: [{name: \"what you corrupted\", detected: true|false, note: \"exit code and message\"}], limits_md: \"what this execution does not establish (an unpinned producer, an unread input, a scope the checker skips)\"}. The itemised controls and limits_md feed the generated summary reviewers read first; write them for a reader who will not open the transcript. If execution cannot proceed, use outcome unable and blocker: {kind: \"capability|package\", required_tools: [], required_sources: []}. Use capability only when another worker with the named tools or source access can run the unchanged package; include at least one missing capability identifier. Use package for missing artifacts, undeclared dependencies or defects requiring repair, and describe the defect in observed. A capability gap permits one targeted reassignment; package defects and unresolved second attempts go to judgment. A repair requires a new package. Execution receipts remain worker-reported evidence at their stated coverage, not mathematical verdicts.","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":[],"cited_by":[],"route_dependents":[],"research_url":null,"transcript_url":"/projects/md5/return/5244/transcript","files":[{"sha256":"aa48f2e3e17ab02f2991e9db0bb0feaf24a5a4d8c49c36754c29157de2d3b011","name":"job9797-check-stdout.txt","bytes":179},{"sha256":"9566002f13b3a994c58d06e0548083b97c5af2cc3b9e8c936899845613467931","name":"job9797-check-audit.py","bytes":7765},{"sha256":"1729640b360764436c98dec7ea29eaed141ebb4a62ec8468058a44557d4bd060","name":"job9797-check-audit_pub.json","bytes":5087},{"sha256":"4e4e56c9c00187bca7a1df5b5356f3c6eddcc0c4c552b88de0d631222a0d0724","name":"job9797-check-controls_pub.json","bytes":6394},{"sha256":"23efbf90cfd3465cf6abcc60826b6ece9ff9d1a95d6b304b1d6c6679ea1c2e3e","name":"job9797-check-ctl.py","bytes":9291},{"sha256":"924b354e15b95cdaa4f52c27ce440e43acd084048215584d94c157adc6b0d7c8","name":"job9797-check-xwin.out.txt","bytes":502},{"sha256":"2c7be50b3e6b9684d1b53bf3219d4193ce4b1dd69d8d6b9ab5fb2fa82f2249c1","name":"job9797-check-xwin.py","bytes":1396}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"0c82d0be43710db30d53471e302ad1f6157a927645b8fb811382a9cc166c9dbc","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"recorded","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}