# Fixed revision comparison and one integration This comparison replaces the misleading interpretation of historical percentages as one learning curve. The preserved V17 result was 143/160 useful answers (89.4%) on corpus V9 and passed that corpus's original 80% gate. V31 scored 279/320 (87.2%) on an expanded corpus with a later 90% target and two accepted language violations. V60's 68/80 (85%) came from exposed probes, not a full repeated trial. V77's 220/320 (68.8%) used different sources, grading and an isolated 2,400-token answer ceiling. Those historical results remain valid records of their own experiments; they cannot establish a like-for-like regression or improvement. ## Comparable design V17, V31 and V77 run the same 22 invented sources: the eleven original V9 cases and eleven V22 held-out cases, now called legacy and recent blocks because both are exposed. Each block has ten writable sources and one deliberate read-only hold. Every source runs twice, with English/Russian queries, English/Russian requested answers and explicit/inferred views. That gives 352 answer coordinates per revision, including 320 writable coordinates and 32 read-only controls. Repeated language coordinates are not independent source scenarios. Each revision owns its built runtime, protocol package and SQLite ABI. Submissions rotate sequentially across arms. The common public-operation procedure, corpus, source-first obligations and model policy were frozen before execution. The actual configuration uses a 2,400-token derive ceiling and a 1,024-token answer ceiling; no model or output-budget upgrade is credited as a code improvement. The separate reviewer grades original sources and anonymous public records, retrievals and answers. Private extraction or verifier rationales are unavailable to the grader. Complete retention requires all material source propositions; focused answers may omit independent neighboring facts but must retain every qualification governing the requested proposition. Generated knowledge is English. Requested answer language is independent of query language, and actual proper names keep their source spelling. One early packet omitted existing managed relation markers. The corrected formatter exposes complete public read lines uniformly for every revision. This changed no runtime output or trial input. The original packet, initial grades and exact source-based correction ledger remain preserved. Final grades must equal initial grades plus those explicit corrections. ## Decision rules and implementation The predeclared original target is 80% complete retention, qualified retrieval and useful answers in every block/run, with the later 90% answer target reported separately. Accepted source, qualification, language and promotion errors must be zero. Bytes, replay, read-only holds and ordinary Markdown admission must pass, and case availability failures must stay within 5% in every block/run. Pooled scores cannot rescue a failing block. There are too few independent sources to claim general or longitudinal reliability. The one proposed integration addresses observed boundaries: - An otherwise valid generated report with unfamiliar personal-limit wording reaches the existing source verifier directly, retaining the original text. A candidate-owned source/current-text witness is required alongside the existing semantic decisions. The former specialized report rewrite is removed. - Extraction and verification share a complete retrieval-unit rule. Source-coupled question, hypothesis and report limits remain in the same readable record; independent neighboring facts remain separate. Neutral provenance cannot invent a performed activity, and embedded hypotheses retain their commitment. This instruction is not a deterministic completeness proof. - Complete-record rendering preserves the exact spelling of a named source already selected in its text. - Bounded English/Russian report and preliminary-hypothesis category headings qualify ordinary Markdown; projection migration corrects derived classifications without editing source bytes. This is one fixed comparison and one integration verification, not a new sequence of tuning trials. Including the two previously charged iterations, the three comparison arms and this integration used six of the ten-iteration cap when frozen. Charging the later deterministic clock consistency fix separately brings closure to seven. No additional live tuning cycle follows. ## Completed frozen-revision comparison | Revision | Legacy run 1 | Legacy run 2 | Recent run 1 | Recent run 2 | Useful total | Accepted answer errors | | --- | ---: | ---: | ---: | ---: | ---: | ---: | | V17 | 68/80 | 71/80 | 24/80 | 30/80 | 193/320 | 56 | | V31 | 64/80 | 53/80 | 48/80 | 51/80 | 216/320 | 13 | | V77 | 63/80 | 59/80 | 36/80 | 53/80 | 211/320 | 11 | The comparison confirms a real original-scenario regression: V17 scores 139/160 (86.9%) against V77's 122/160 (76.3%). It also shows why a rollback is insufficient: V17 has 56 accepted answer errors across the full suite, while V31's five-answer net gain over V77 comes with lower legacy coverage and 13 accepted answer errors. No revision meets the fixed quality gate. Complete retention is 32/40, 31/40 and 30/40 for V17, V31 and V77 respectively; V17 also has one retained-set qualification error. Incomplete source coverage and accepted-content errors are separate judgments. All three revisions complete all 352 coordinates, preserve source bytes and replay, and justify all 32 read-only nulls. All share six ordinary Markdown admission mismatches. Case-level availability failures are zero, while finer answer-level diagnostics count one failed answer call for V17, four for V31 and zero for V77. These denominators remain distinct in the machine-readable results. The bound selection chooses one focused integration on V77, informed by the repeated alternatives, report, nested-report and assistant retention losses, the question-selection loss, and the source-name publication errors. V31's five-answer net gain is below the predeclared eight-answer threshold for a large port and does not have a consistent block advantage. The source-specific hypotheses distinguish which paths the correction can reach from improvements that still require live evidence. ## Final integration result The measured integration at `963bf68` scores **249/320 useful answers (77.8%)**, compared with V77's **211/320 (65.9%)** under the same sources, model policy, budgets and rubric. Original-scenario coverage recovers to **142/160 (88.8%)**, against V77's 122/160 (76.3%) and V17's 139/160 (86.9%). Newer-scenario coverage is **107/160 (66.9%)**. The original block's two runs remain visibly different; this is a bounded recovery, not proof of repeatable general reliability. | Block / run | Useful answers | Complete retained sets | Useful retrievals | | --- | ---: | ---: | ---: | | legacy / 1 | 76/80 | 10/10 | 40/40 | | legacy / 2 | 66/80 | 9/10 | 36/40 | | recent / 1 | 53/80 | 7/10 | 34/40 | | recent / 2 | 54/80 | 8/10 | 33/40 | By requested answer language, the integration scores **131/160 English answers** and **118/160 Russian answers**, with 4 and 5 accepted errors respectively. These are descriptive slices; they do not replace the fixed block/run gate. There are **9 accepted answer errors**, down from V77's 11, and **34/40 complete retained sets**. The integration has 1 unsupported retained set (V77 had none), 0 retained-set qualification errors, 0 retained-language violations, 0 answer-language violations and 0 unsafe factual promotions. Correct output language does not imply correct translated meaning. The newly unsupported retained set changes a rattle into vibration; this regression must be weighed alongside the coverage gain. All 32 read-only nulls are justified; source bytes and replay pass in all 44 case-runs. Ordinary Markdown admission mismatches fall from six to **0**. Case-level availability failures are 4; the separate answer-level diagnostic counts 5 failures. The **80% per-block/run quality gate remains unmet**, as does the later 90% answer target. The gate requires complete retention, qualified retrieval and useful answers in every block/run, plus zero accepted source, qualification, language and promotion errors. The newer scenarios still have missing independent facts, protective holds of defective candidates, missed view selection, and accepted qualification or component-translation errors. Pooled scores do not rescue a failing block. The integration removes the forced report rewrite before source verification, joins source-coupled qualifications into one retrieval unit, preserves selected source-name spelling, and corrects report/hypothesis heading projections. Independent forensics confirm the report path and complete open-question record were exercised. Their improved outcomes do not prove that every stochastic change was caused by the new instructions. ## Evidence and limits This closes the bounded implementation and evaluation work in PR #70 with reliability acceptance unmet. The first recent block alone is below the declared 80% coverage target and contains accepted answer errors; neither a pooled average nor a later successful run can rescue that gate. The opt-in language policy and bounded discourse protections are implemented, but this result does not close issues #61/#62, approve a merge, or establish reliable automatic retention. The remaining failures concern complete source coverage, view selection and semantic verification as well as false holds. More phrase-specific tuning on these exposed cases is not part of this closure. The live integration remains frozen at `963bf68`. A later deterministic consistency correction recognizes the exact Russian source-clock sentence already prescribed by generation for supported day, week, month and year labels. The original parser falsely withheld faithful “tomorrow” drafts before semantic verification. The correction keeps clause, quotation, unknown-date, language and semantic checks; it does not make a new semantic acceptance decision. Its separate tests and review are included. No live coordinate is rerun or rescored for that correction, and no hypothetical recovered answer is added to the integration score. The public bundle contains frozen sources and obligations, anonymous public review packets, preserved initial/final grades and correction ledgers, public operation receipts, completed-run records, exact selection/preflight declarations, validation output, case deltas, and independent code/documentation reviews. The source-name and report audit are bounded mechanisms; the retrieval-unit contract remains a model instruction and does not prove exhaustive source coverage. The manifest records original and published hashes. The exact repository/runtime paths and recognized operating-system temporary or mounted-volume paths are replaced with explicit placeholders. Remaining recognized home and system paths fail publication; this is a bounded check, not a proof against every possible private string. Source propositions, public answers and judgments retain their bytes. Raw model traces remain local with their original hashes. Public-operation projections are included without the private model reasoning. Archived harness scripts record the exact execution procedure and expect the original repository layout; this is an evidence archive rather than a self-contained replay package. The suite contains twenty independent writable source scenarios, each repeated across languages, views and runs. Those repeated coordinates are not independent scenario samples. These exposed cases can show regressions and recoveries under one policy; they cannot establish unseen-language, implicit- discourse, general or longitudinal reliability. Independent model grading remains fallible.