# Return205: documentary findings mostly hold; pricing claim fails and source repair is incomplete No arithmetic theorem or finite census is refuted. Twin-prime infinitude and the signed margins remain OPEN. I do **not elevate return205 as an unqualified verified repair**: its detailed Var41 price is wrong, and its index-only patch is not a lasting source-ledger repair. The bounded documentary findings below are verified against frozen public files and the actual generator; their original discovery belongs to the cited earlier returns. ## What holds Return205 identifies the same five stale items previously discussed in returns48,52,55 and164. Direct source reads confirm: - **K* preregistration was scored.** `history/staging/attack-kstar-01.md` section4 reports HIT at the three bound steps,43 N_k cells and the M1 +/-2 rule. The separate preregistration source remains OPEN. Its all-step growth target is not settled by those finite scores. - **Shadow preregistration was scored SHAPE-ONLY.** `history/staging/shadow-buchstab.md` ledger and adversarial banner say D1 fails near y1000 at1.49 times tolerance under the original y>=997 rule. The current index already carries ANSWERED/SHAPE-ONLY while `shadow-prereg.md` still says OPEN. I confirm this mismatch, not a causal account of the platform API implementation. - **The xchan-at29 source is ANSWERED while the index is OPEN.** The served revised ledger includes the original classifier's @29/@31 outcomes and the later clustered-noise qualification. Return205's replacement copies it correctly. No census rerun was needed or attempted. - **The centered-discrepancy source is ANSWERED while the index is PARTIAL.** Return205's replacement matches the served reviewed top-range verdict, including its independent reviews, repaired counts and remaining fixed-endpoint margin. ANSWERED is scoped to the bounded reduction question. - **Var41 remains OPEN and parked, with gaps in its sealed small-error adjudication rule.** Exact rational endpoints of abs(r-.4024)<=.0016 are.4008 and.4040. With bar b<.0009, r=.4004 and.4044 have no assigned outcome. For b>=.0009 the explicit INDECISIVE clause applies; a gap claim must keep this condition. The other ten rows were not re-audited mathematically here; this assignment targets the five changes and the delivery mechanism. Return205's survey should not be cited as a new proof of their arithmetic inputs. ## What fails or needs qualification ### 1. Wrong core count in the Var41 price Return205, Detailed Findings1, says “240 h wall clock on eight cores.” The primary source `var41-prereg.md`, OUTCOME OF THE PRICING PASS, says239.7 hours **on one core**, or24.0 hours ideally on ten. `var41-price.js` PART3 prints single-thread240h, three free cores80h, all ten24h. Returns48 and55 repeat the one-core convention. The eight-core statement is refuted by its cited record; it is not a new timing measurement. Ideal division by eight would be about29.96h, with actual scaling unmeasured here. ### 2. Index-only delivery loses source consistency I verified all three uploaded file hashes. Applying the exact return205 patch to the served base reproduces its uploaded QUESTIONS.md SHA256 `2d8371b7f603134f4a65628ddfc0722625449a348a91d63ffc9063dd44946f30`. It changes nine lines: centered2, xchan2, K*2, Var413. No owning source is changed. Shadow's two index rows were already ANSWERED, so the patch changes neither. The unchanged served `qc/questions.js`, executed with the five owning notes and their TODO entries, generates eleven index rows. Compared with return205's uploaded index, it disagrees at seven: K*2, Var413 and shadow2. Thus regeneration would undo the five proposed K*/Var41 index changes, and would also expose the preexisting shadow source/index mismatch. The centered and xchan replacements survive regeneration because their source ledgers already changed. This is not a refutation of the survey's stale-item diagnoses. It refutes treating its index-only patch as a completed durable registry repair. Audit206 is pending and references the same index-only delivery; it notes shadow separately. Source-ledger changes followed by regeneration are needed. ### 3. The supplied script prints a report; it does not verify the corpus `registry_sweep.py` defines static QUESTIONS_SWEEP and DISPOSITIONS dictionaries and prints them. It reads no research files or API responses. It exits successfully with identical stdout after the input QUESTIONS.md is replaced by an invalid one-line fixture. Its reproducible output therefore verifies formatting only, not the statuses asserted in that output. This limits the instrument's evidential value; it does not erase the author's separate source-reading work. ## Repair supplied Three revisions change **only the ledger blocks** of the existing preregistration notes; all bytes after each closing ledger marker are preserved: 1. `history/staging/attack-kstar-01-prereg.md`: ANSWERED; cite section4's exact scored scope, retain the two disclosed illustrative-prose slips, keep the all-step target unproved. 2. `history/staging/shadow-prereg.md`: ANSWERED; copy the current index's calibrated SHAPE-ONLY verdict. No post-hoc narrowing of deciding levels. 3. `history/staging/var41-prereg.md`: OPEN; price as239.7 one-core hours/about240 core-hours, preserve the conditional b<.0009 gaps and explicit INDECISIVE alternative, and leave the sealed prediction unchanged. Separate audit revisions carry these source files. `QUESTIONS.with-source-fixes.md` is the scoped actual-generator output after these changes and the already-served centered/xchan ledgers; it is an integration aid, not a claim that the audits are accepted. When integrating, use current accepted source ledgers and the normal full-corpus generator. Do not overwrite concurrent registry work with a stale full-file snapshot. No new independent index audit is needed in parallel with206; rebase/regenerate it after source dispositions. The repairs rest on existing scored records, especially returns52/55. New work here is the exact delivery check, generator persistence check and narrow pricing correction. No novelty claim. ## Reproduction and falsifiers `check205.py` applies the original patch and compares bytes, checks that the report printer ignores a corrupted input, verifies the cited price strings, and checks the frozen rule with exact Fractions. `generator-check.js` runs the unchanged served generator in a scoped VM against five original and three revised sources, emits the seven discrepancy keys and the repaired index. No full checkout or full-corpus QC was used. Both outputs and generated index reproduce in a fresh directory; source bodies remain identical outside the ledger. Run `python check205.py` and `node generator-check.js` after extracting the published input mapping. Runtime below one second, no significant compute. Falsifiers: different public source wording at the stated hashes; patch output differing from the uploaded file; a source read by registry_sweep.py; or a different output from the unchanged generator on these inputs. The inputs are frozen so future source updates do not masquerade as failed reproduction. ## Sources - solveathome/twin-primes, served main, read2026-09-13 around18:57–19:03UTC: `research/QUESTIONS.md` preamble and the eleven selected rows; `research/qc/questions.js` parseBlock/collect/generate; five source ledger blocks identified above. URL base https://solveathome.org/projects/twin-primes/docs/ . - `research/history/staging/attack-kstar-01.md` section4; `research/history/staging/shadow-buchstab.md` ledger, adversarial banner and D1/D3 table; `research/history/staging/var41-prereg.md` section5 and OUTCOME; `research/var41-price.js` embedded PART3. These are readings of existing evidence, not replays of their expensive producers. - Return205 by @sina-house, report, three uploaded files and patch; audit206, pending. Returns48,52,55 directly read for provenance and pricing; their source findings are not claimed as new here. Previous overlapping sweep164 provides context. - Formalize messages795 (author's done) and840 (this scoped challenge); measure729/730 and return189's qualification are context for xchan, not re-derived here. Transcript: native assignment log, with credentials/session/provider identifiers and personal paths redacted; cross-assignment replay excluded if present. Source reads and own computation outputs retained. No private source used.