YUCLAW 9.1.0 — minor release: adds the point-in-time research-export command and the prospective SourceLock R10 admission rule; replay verification bound to its inputs, honest evidence scope, bilingual onboarding (9.0.0 feature scope F1–F5 kept)

Research & education only. Not investment advice.

YUCLAW 9.1.0 — Evidence-First Financial AI · The Science Trust Layer for Financial AI

Financial AI normally gives you an answer. YUCLAW gives you the evidence, what that evidence can support, what it cannot support, and whether that conclusion survived time.

Changed in 9.1.0 — minor release: adds the point-in-time research-export command and the prospective SourceLock R10 evidence-admission rule, and corrects research verification and public evidence text (9.0.0 feature scope F1–F5 kept; no statistic, threshold or historical result changed)

• Replay verification: yuclaw replay-lab now fails when a required daily ledger root is missing, when a Newey–West p-value or any core expected statistic has been altered, or when forward analysis scores no longer match the ledger leaves; duplicate JSON keys, non-finite numbers, unsupported schemas and inconsistent dates or counts are refused. Internal numerical reproduction, bundle integrity and externally pinned provenance are three separate outcomes; exit 4 means the required external verification is incomplete (--require-external-ledger, which needs an independently obtained ledger and its SHA-256). The standalone verifier script uses the same implementation. The legacy bundle's 79 uncovered date/ticker inputs are disclosed; a passing internal reproduction does not authenticate them.
• Point-in-time research export: the new yuclaw research-export command builds a spreadsheet-safe CSV and an exact-JSON export from the complete, manifest-matched public history at a UTC cutoff, with a dictionary and file hashes; a capped preview is refused as a complete source, recorded availability is distinguished from filing dates, and self-consistency from independent authenticity. This command exists from 9.1.0; it is not present in the published 9.0.0 — added functionality, one reason this release is a minor release rather than a patch.
• Replay bundle (lab-replay/2): snapshot identities, the trading calendar and the observation units (signal snapshots and distinct issuer-dates; no count of independent observations is claimed) travel with the bundle; earlier bundle bytes are unchanged and the verifier still reads them.
• Event hashes: the public pages and the verify endpoint state what source_hash covers (the stored event record), that public event-inclusion verification is NOT AVAILABLE, and that the daily ledger covers signal snapshots; hash consistency, provenance, economic interpretation and source truth stay distinct.
• Earnings comparisons: the Suncor trace page flags that its EARNINGS_BEAT excerpt states profitability, not a comparison with expectations (UNSUPPORTED_COMPARISON); the original event is retained. Prospective SourceLock rule R10 refuses a new beat/miss admission whose excerpt does not state the comparison, refuses non-finite or out-of-range confidence, and adds the neutral event type EARNINGS_RESULT. The rule is registered as a dated addendum under the Science Trust umbrella protocol and applies from its deployment date: no historical event is reclassified and no reported estimate changes. A prospective change to evidence admission is the second reason this release is a minor release rather than a patch. The registered addendum names the carrying release by its preparation number 9.0.1; the owner renumbered that same release 9.1.0 on 2026-10-15, and the registered bytes are left unchanged (an immutable record is never rewritten) — the deployment date is the 9.1.0 push-to-main timestamp. The check is lexical and structural; it is not factual entailment and not an independent check of analyst expectations.
• Why pages: event counts, distinct accessions, objects without an accession identifier and the largest event-family share are reported separately, and no count of independent observations is invented. The composite anatomy lists each component's stored confidence, nominal weight, effective weight and contribution where recorded, names the weight provenance (recorded by snapshots written from 9.1.0 on; current-configuration reconciliation only for earlier snapshots), keeps C4 disclosed as frozen, and reports INCOMPLETE instead of reconstructing an absent input.
• Public text: cohort wording follows the displayed numbers, including rounded ties (the 4.15% versus 4.06% direction error is corrected); homepage claims are limited to supported filing-event coverage; market-only inputs and research limitations are stated; the Validation Lab carries its verification scope and statistical limits. The historical Lab packet keeps its bytes and receives a hash-bound limitations correction; later packets carry the current limitations.
• Ledger day files: the nightly rebuild leaves an existing per-day ledger file byte-for-byte unchanged when its root and entries are unchanged (only new days are written), so an externally recorded file hash keeps matching.
• Bilingual onboarding: a research-start page in English and French with the published v9 workflow, a fictional first exercise and the limits of what the checks establish. Mobile tables and long links fit or scroll inside focusable regions; the viewport check inspects the served document, fails on HTTP errors or wrong bytes, opens disclosure panels and uses real keyboard input; the Explorer keyboard test uses a portable native selection.
• Release tooling: the 9.x notes composer accepts 9.1.0 as a minor release of the 9.0 line with this tracked change list (the 9.0.0 feature scope F1–F5 kept; the patch definition — defect repairs only — is not widened) and binds Gate #15 to a decision record for exactly the released version (the 9.0.0-only exception is not carried); the nightly stages the research corrections after the weekly note, the last page it writes, so every current page carries its responsive regions.
• Status, stated once and unchanged: the workbench and its modules are experimental, local and owner-operated; no independent security review has been performed; no user study and no human comprehension study was run, scheduled or recruited (human benefit PENDING); backups remain cancelled. This release creates no missing historical anchor, no independent sample, no predictive result and no agent-reliability experiment. Built in Canada.

In the 9.0 line since 9.0.0 — research briefs with traceable AI assistance (local, loopback only; feature scope F1–F5 kept)

• Sentence and claim inspection (F1): every recorded statement is bound to an exact UTF-8 byte span of its text view and to the specific v8 claim version it uses; six roles (direct quotation, computed statement, attributed source statement, analyst interpretation, model-written commentary, unresolved claim) describe function, not correctness; a four-area inspector (sources and calculations · how the text was produced · changes · checks) and five independent answers per statement (byte integrity · recorded origin · issuer trust · substantive support · time scope) that are never merged into a score; exact quotations and registered arithmetic are decided deterministically, an operator's or model's link is ATTRIBUTED, contradictions and unresolved statements are preserved; coverage is stated over identified statements with extraction completeness NOT established.
• Financial brief composer (F2): briefs rendered from a stable, digest-identified evidence snapshot by deterministic bilingual templates (guidance change · numerical comparison · explicitly unresolved interpretation) that protect typed amounts, signs, currencies, units, metrics, bases, periods, accessions and quotations as slots over the v8 Decimal calculator; imported drafts are a separate route whose sentences start unassessed; raw and assembled bytes are distinct identities; a changed protected fact cannot keep its support; no API key, model download or hosted service is required.
• Edits, translations and source corrections (F3): every saved revision is a new immutable version with a parent reference and a transform record (kind, implementation, languages, exact-bytes span mapping, review findings); the original is preserved beside the revision; a parent's checks never carry to the child's bytes; a deterministic template re-rendering binds the other language fully, an entered translation records its provenance, is never sent to a provider and carries no certification claim; a corrected v8 source or an amended claim produces review items on exactly the statements whose recorded dependencies reference it, while earlier exports stay valid historical snapshots.
• Provenance records and detector-report import (F4): generation receipts, detector reports and calibration records are imported through a bounded parser (duplicate keys, floats, depth and size refused), bound to the exact text view and tested span, and kept apart by origin (operator assertion · connector observation · issuer-signed statement); signature validity, issuer trust under this workspace's own roots, revocation and payload binding are four separate answers; execution, signal (only when COMPLETED) and calibration (APPLICABLE only in the enrolled record's own scope, from an authenticated record) are separate dimensions; an unavailable account, unsupported input, too-short input or failure never becomes NOT_DETECTED; no live detector or provider connector exists in 9.0.
• Portable verification and measured operations (F5): a readable static HTML brief (no script, CSP), JSON records and a versioned packet (yuclaw.brief-packet/1) verify in a fresh workspace without the author's database; every check is declared separately (VERIFIED · FAILED · NOT_RECOMPUTABLE · REPORT_ONLY · UNSUPPORTED · NOT_APPLICABLE); rights filtering withholds excerpts, prompts and raw responses whose disclosure is not established; keys, secrets and local paths never travel; operation identifiers, attempts, retries, outcomes and durations are measured from the first run with their definitions and missingness, and a methods/limitations appendix is derived from the records.
• Candidate evidence (candidate b16cc33): current: 77/77 PASS; floor: 77/77 PASS; installed self-check PASS; browser inspection PASS; clean install FAIL on that pre-freeze candidate (readme transcript exact); the final pair is verified from fresh installs before any publication.

Evidence totals (public scoreboard at composition)

• Replication attempts: primary unavailable, successful unavailable; registration PENDING; exact-release coverage BOUND
• Witness reviews unavailable · audit-break attempts unavailable · refusals unavailable · challenges unavailable · document-use receipts 0 · pilots 0
• Board timestamp 2026-10-15T02:45:21.791661Z; these are counts of records, not a quality or independence verdict

Activation status (every proposed activation)

• real receipt program: INACTIVE — no reviewer appointed, no registration record adopted, the three-reproduction floor unadopted (D3)
• nightly status delivery: INACTIVE — adapter and preview only; delivery not activated (D4-NIGHTLY-STATUS)
• note-snapshot coordinator: INACTIVE — contract v3 stays live; coordinator not wired (D4-SNAPSHOT)
• sentinel policy: INACTIVE — unchanged; proposal only (D4-SENTINEL)
• Phase-C prospective protocol: INACTIVE — draft; unregistered; nothing runs (D4-PHASE-C)
• Phase 6 / A2 designation of S: INACTIVE — candidate record only; not registered; N_eff not computed (D4-A2-S)
• ETF class addendum: INACTIVE — proposed classification path; registered set untouched (D4-ETF-ADDENDUM)
• U-ladder promotion / window: INACTIVE — fixture validation only; no admission, promotion or registered window (D4-U-LADDER-WINDOW)
• Gate #15 human study: INACTIVE — kit ships; no study run; explicit owner exception for 9.1.0 only, MANUAL_REVIEW — not a pass; human benefit PENDING
• Phase-5 contribution anatomy: a READER of registered protocol lines and registered results; it is not a registered result and registers nothing

Release policy (recorded; the publisher refuses notes that do not match the record)

• Allocation: decision D1-ACCEPT on allocation document V9-ALLOC-9.1.0-2026-10-15-P1 (sha256 42b9cb1d035d7cb5…)

• Gate #15 (user comprehension test passes): MANUAL_REVIEW — requirement NOT satisfied; the owner accepted an explicit release-policy exception for 9.1.0 only on 2026-10-15 (decision record sha256 4c712d26a18ff7d8…); not carried to any later release; no human comprehension study was run, scheduled or recruited and none is claimed — not a pass; the automated consumer-posture scaffold check is retained; human benefit PENDING. The owner's decision, verbatim:
  3. GATE 15 — OWNER EXCEPTION FOR 9.1.0 ONLY

  For 9.1.0 only, I accept an explicit release-policy exception for Gate #15.
  The human comprehension requirement is not satisfied; Gate #15 remains
  MANUAL_REVIEW and must not be represented as passed.

  The human study is not authorized and must not be run, scheduled or
  recruited. Disclose this exception verbatim in the release-policy record
  and public release notes.

  This exception applies to 9.1.0 only. It is not carried to any later release.

Continuing objects (unchanged from 8.0.1) — name · receipt · status

• Layered Evidence Dependency v1, first read · chain lines 81–82, sha256 0b2ac8a5967b13aa… · STRUCTURE_PRINTED — structural_completeness = PARTIAL; N_eff PENDING; READ_SCOPE = STRUCTURAL_ONLY (chain 81–82)
• Science Trust surfaces — per-name research-state cards + machine JSON, 132 names · anchor ac51ddfe97eb… · gate GREEN (machine JSON equals the human card, byte-reproducible); staged preview, not linked from the live navigation
• Research states · sha256 116d6e5f86ae5771… · 132 names: INSUFFICIENT_EVIDENCE 132 — derived, never hand-maintained
• Discovery Ledger · sha256 ca26fb0e99bf43f4… · 37 hypotheses in bijection with 37 registered protocol lines; status counts ACCRUING 18, INCONCLUSIVE 1, OPEN 7, REGISTERED 7, SUPERSEDED 4 — negative and inconclusive findings preserved
• Anytime Evidence Record · sha256 935fa3494856b83e… · 3 prospective enrollments — ACCRUING, not adjudicated
• Evidence Completeness Profiles · sha256 3db6dc6c8cfaf234… · 132 names; ETF class membership BLOCKED_BY_REGISTRATION
• Protocol registry · 83 chained lines, tip 9a515414…, chain-verified
• Public daily evidence ledger · 106 daily blocks, latest 2026-10-14 root e36e5b8d53eb… · append-only, replayable
• C6 risk channel: rare-by-construction confirmed OOS (22% fire rate, n=9 held-out); sign positive at n=2 elevated — accruing · fourth read chain line 77 (98dcf74a827a…): DESCRIPTIVE
• Cross-lens reversal coherence · chain line 79 (be05bf7a9dc8…) · INSUFFICIENT — accruing, no coherence claim (chain 79)
• Consumer-posture gate · five deterministic stranger personas · GREEN (scaffold); user-comprehension study: not run — explicit owner exception for 9.0.0 only, MANUAL_REVIEW (not a pass)
• Replication · public log 1 entry, bundle sha256 f431f9c629ac38b5… · REPRODUCED — External-machine reproduction completed by an affiliated operator; unaffiliated replications: 0
• yuclaw 9.1.0 package · wheel + sdist sha256 attached to this release · CLI · REST · MCP · SDK

Not in this release

• 9.1.0 keeps the frozen 9.0.0 feature scope (F1–F5): every statement below is that scope's own wording; what 9.1.0 adds or changes is listed under Changed above.
• Backup functionality is cancelled: no backup scheduling, restoration feature or drill exists; a packet is a verification artifact, not a backup.
• Excluded by the owner: human comprehension study, backup or restore functionality.
• Human benefit: PENDING — no pilot, no user study; the journeys and tests demonstrate behaviour, not benefit.
• Review: every edit, link, resolution and import in the release evidence was recorded by the automated tests, the installed self-check and the browser inspection as a simulated action; there was no human review and no user study.
• Fixtures are clearly fictional demonstration data (issuer, filings, amounts); the fictional guidance example is never market data or a historical research record.
• Detector and watermark evidence: NOT_DETECTED never proves human authorship; a detector score is not the percentage of text written by AI; a watermark establishes neither identity, ownership, responsibility nor factual accuracy; a local receipt is neither legal compliance nor an independently anchored timestamp; an imported provider-positive result with unknown calibration is a provider-reported result whose calibration is not established here.
• Deferred, not in 9.0 (explicitly later work, not incomplete core): live provider generation or detection connectors, local entropy controlled sampler or open weight model integration, detector calibration campaigns registration or sequential inference methods. No public text-detection endpoint, watermark parameter, key or provider capability is invented; no OpenAI-compatible watermark claim is made; primary evidence is never altered to carry a watermark.
• The v8 workbench, its four modules, its export format and its published guides are unchanged; v9 writes a separate sidecar beside the v8 journal and an 8.0.1 client reads its own data with the sidecar present and untouched.
• N_eff PENDING
• Phase-5 contribution anatomy NOT YET
• user-comprehension study: not run; explicit owner exception for 9.0.0 only — MANUAL_REVIEW (not a pass; human benefit PENDING; not carried to any later release)
• unaffiliated replications 0

Built in Canada

Built in Canada — from Lake Ontario to Lake Louise and Kananaskis Lake — with gratitude to the country whose land and light frame this work.