Fetching metadata from the HF Docker repository...
Oct 1 09:38 Withheld-run fixes: recorded margin floor; re-blend bounded by the recorded cap; precise break-even revenue and sign-aware growth clause; pre-revenue message; no retry invitation for deterministic failures Sep 30 09:02 Mechanical DCF: cumulative discounting along the rate path; MechanicalDCFState; MECHANICAL_DCF_DISCOUNT_PATH 08:49 Story business identity from canonical state only: no recalled segments, filing-verified disposals, story-business validators 08:16 BSX root causes: token-safe segment keywords + parent-sector compatibility gate; growth label; semantic gate before the PM Sep 28 19:15 Fundamentals table: compact headers + provenance footnotes; print-safe tables 18:46 Lineage labels read the frozen state (Mapping, not dict) 18:36 Claim lineage: typed guidance/engine-vs-consensus relations, lineage checks, HistoricalMetricQuality label 18:02 Guidance trigger subject uses the row's own label (keeps scope qualifiers like 'excluding OSTTRA') 17:46 Open items: break-even solved on the adopted margin (floor never cuts); recalibrated thesis gate; benign co-citation; no triggers on price multiples; nominated peers only; benchmark applicability; merged guidance rows; engine-EPS basis; market composite noun; insider transaction counts; filer names; quarter limitation; story revenue/share precision 13:12 Gate: adaptive thinking (Sonnet 5 rejects type enabled + budget_tokens with 400); retry Claude without thinking on a rejected thinking config; QA notes record which gate answered 12:43 Gate calibrated offline (reviewers' standard rubric; extended thinking; bounded retries; run deadline); triggers refuse sources that include a separated business 10:43 Gate verdicts must quote the object's own words (unattributed verdicts re-asked, then kept); thesis pass reads its points' evidence, not their wording; frontier pointer omitted when the unit prints its cases; gate drop ids carry P/A labels 10:25 Structured arguments items 1-2: thesis rests_on + two-pass gate with report context, one bounded repair turn, typed trigger labels, evidence printed in full once, identifier and authored-once audits, relaxed evidence limits, past returns withheld 06:54 Revert T7 scoping: rubric reviewers judged the collective-assessment drops correct 06:48 Structured arguments: PM numeric parsing, unquoted numeric skeletons, gate without temperature, three validator false positives (spelled period compounds, topic-scoped model phrases, T7 scoped to cited assumptions) 06:36 Structured arguments (crucible.args/2): typed evidence + figure-free judgement, templated figures, validated forwarding between rounds, semantic gate; data-quality notes kept out of author prompts Sep 27 21:03 Predicate verb agreement with its subject 20:57 High/low about an assumption is a typed assessment; trigger basis not written twice 20:51 Counts after participles, coordinated predicates in the subject check, evaluative direction adjectives 20:46 Price moves checked against the price, return magnitude bands, pronoun objects and anchor-vs-price compounds, raw identifiers 20:39 Token-boundary repeats, life-cycle stage terms, elliptical valuation subjects, coordinated and modified predicate subjects 20:32 Spread assessments, consequences drawn from descriptive evidence, speech verbs and officer names, anchor noun compounds 20:24 Universal valuation claims checked against every anchor; consistent tokenisation in the doubling check 20:17 S15: word-matching fallback only for unsourced triggers (a sourced level trigger is not a guided change) 20:11 Rebuttal topics labelled as the opponent's claim; forecast reader names with direction check; complete ownership breakdowns 20:04 Speech lineage (management statements need their quoted words), verbal magnitude bands, spelled counts, multiples named by metric 19:54 No mid-sentence truncation of authored prose; incomplete-sentence rule; money figures named by concept; named break-even counterparts; count agreement 19:46 Clause-scoped valuation/logic rules, trend mood, parenthesis splitter, issuer as performance subject, hyphen-normalised doubling, count units, PP-modified predicate subjects 19:38 Guidance basis label: only guidance rows with typed values (a trigger may cite a plain metric claim) 19:32 Remaining SPGI items: predicate/list render slots, expected-return direction, value proximity band, metric-family comparisons, trigger guidance basis, composite input labels, segment margin cap disclosure + reconciliation 19:19 Acceptance fixes: ownership others count, relation clause/operand doubling, expectation comparatives, assumption assessments, stock performance claims, quote speaker roles, statement domains 18:43 Descriptive evidence cannot be a reason; failed guidance extraction retried and not reported as absence; clause phrase capitalisation 18:37 R5: PASS/FAIL status tokens are case-sensitive (ordinary 'fail' is not QA vocabulary) 18:31 Evidence lineage: final-claim lineage, public language only, EvidenceUse conclusion types, NewsEvidenceClaim, qualifier and basis ownership, full guidance table 18:04 Title-case strong ratings checked; seats with no publishable prose say so 17:59 Deterministic fallbacks for empty summary/PM sections; relation tokens shown as clauses in PM anchors 17:52 FULL_SENTENCE relation form is template-only; frontier framing is a valid paraphrase 17:45 Publication semantics: relation-logic preservation, subject-aware valuation language, full-sentence slot, conclusion domains, segment perimeter, insider pressure 17:23 Doubled unit symbols and dangling references to removed sentences are rejected 14:40 PREDICATE relation form must directly follow its subject 14:34 Relation logic covers the break-even growth relation; repeated or back-to-back claims rejected 14:29 Publication integrity: final-claim-only summary, relation logic, typed render slots, valuation relations, evidence domains, trigger ranges 13:40 Guidance label binding: a guidance value must be named by its own concept in its clause 13:16 Prose/trigger cleanup: trend relations, evidence interpretation, clause grammar, render modes, trigger satisfaction 11:30 Expectations frontier (conditional, not conjunctive); guidance coverage pass and FOUND_BUT_REJECTED semantics 11:14 Relationship layer: phrase claims carry their own direction and article 11:09 Relationship layer: a claim's own direction or bound is never written twice 11:03 Relationship layer: RelationshipClaim, typed break-even/story metrics, model-evidence scope, guidance epistemic state, semantic returns, typed blend sources 07:49 SPGI hardening: ordering truth check, comparative vocabulary, trailing sign nouns 07:42 SPGI hardening: trigger magnitude check, implicit comparisons, token-structure rules 
