Research & education only. Not investment advice.

YUCLAW

Evidence-First Financial AI
The Science Trust Layer for Financial AI.

Evidence before answers.

Financial AI normally gives you an answer.

YUCLAW gives you the evidence — what was known, when it was known, what it can support, what it cannot, and whether the conclusion survived.

Mission

Make financial AI accountable to evidence.

• Trace every claim to the filing it came from.
• Lock every method before the data arrives.
• State every conclusion’s limits.
• Keep every result — including the ones we wished had come out differently.

A public, hash-linked record, built to be recomputed by anyone.

Vision

Become the Science Trust Layer for Financial AI.

The evidence infrastructure that AI systems, researchers, and institutions use to decide what deserves to be believed.

How we work

What you get

Statistics is one instrument. Evidence is the foundation. Science is the discipline.

AI is the market. Trust is the product. Accountability is the mission.

🍁 Built in Canada

First touch — install, ask for help, check a claim, reproduce the Lab (expected exit code in the comment):

pip install yuclaw                                              # installs the CLI
yuclaw --help                                                   # command list with one-line descriptions · exit 0
yuclaw check-claim --text "NVDA reported an insider sale in May 2026"   # Evidence Passport JSON · exit 0
yuclaw check-claim --ticker NVDA --accession 0001045810-26-000019       # passport for one cited filing · exit 0
yuclaw check-claim --accession 0001045810-26-000019            # accession alone: unique → same passport · exit 0
                                                               # (ambiguous accession → exit 2 with the candidate tickers)
yuclaw replay-lab                                              # rebuilds every published Lab statistic · exit 0 = reproduced

Exit-code contract for every command: 0 = success · 1 = ran, negative result (e.g. a replay mismatch) · 2 = usage or validation error · 3 = environment unsupported. The transcript below is generated from the release-candidate wheel and regenerated every release:

Research with v9: English / Français — start here.

Transcript produced from the release-candidate wheel yuclaw-9.1.0-py3-none-any.whl (yuclaw 9.1.0, Python 3.12) by tools/cli_transcript.py; the replay-lab run uses the documented local-bundle path. Every command ran with YUCLAW_CORPUS=snapshot: check-claim answers from the corpus snapshot bundled in the wheel and never looks for a research node (without that setting, a host that reaches a research node answers from it and its passport has no corpus block). No date: release verification compares this block byte-for-byte with a fresh transcript of the final artifact, made the same way.

$ yuclaw --version
yuclaw 9.1.0
[exit 0]

$ yuclaw --help
yuclaw 9.1.0 — evidence-first financial research CLI (research and education only; not investment advice)

usage: yuclaw <command> [args]   ·   yuclaw <command> --help

commands:
  brief            evidence brief (legacy v3 helper)
  cascade          supply-chain cascade view for a ticker (deterministic, evidence-backed)
  challenge        structured challenge bound to an exact artifact + claim id; dispositions keep the original finding (v7)
  check-claim      Evidence Passport — deterministic claim check (--text, --ticker/--type/--date-range, --accession)
  decision         document-use receipt bound to an exact packet manifest digest; export needs permission (v7)
  demo             3-minute guided offline journey — zero config, no backend
  events           accepted-events export (derived data only)
  evidencebench    score EvidenceBench predictions against a supplied item set (--items; rubric v1 exact, v2 candidate; stdlib-only) (7.0.1)
  export           lens events export (--format csv|json; --page builds the evidence packet)
... (18 more lines)
[exit 0]

$ yuclaw check-claim --text "NVDA reported an insider sale in May 2026"
{
 "status": "SOURCE_MATCHED",
 "claim_as_parsed": {
  "ticker": "NVDA",
  "type": "INSIDER_SELL",
  "accession": null,
  "date_range": null
 },
 "misses": [],
 "matched_evidence": "<5 object(s)>",
 "corpus": {
  "mode": "offline_snapshot",
  "snapshot_generated": "2026-08-06T01:06:08.732897+00:00"
 },
 "...": "<10 fields total; not_advice line present: True>"
}
[exit 0]

$ yuclaw check-claim --ticker NVDA --accession 0001045810-26-000019
{
 "status": "SOURCE_MATCHED",
 "claim_as_parsed": {
  "ticker": "NVDA",
  "type": null,
  "accession": "0001045810-26-000019",
  "date_range": null
 },
 "misses": [],
 "matched_evidence": "<1 object(s)>",
 "corpus": {
  "mode": "offline_snapshot",
  "snapshot_generated": "2026-08-06T01:06:08.732897+00:00"
 },
 "...": "<10 fields total; not_advice line present: True>"
}
[exit 0]

$ yuclaw check-claim --accession 0001045810-26-000019
{
 "status": "SOURCE_MATCHED",
 "claim_as_parsed": {
  "ticker": "NVDA",
  "type": null,
  "accession": "0001045810-26-000019",
  "date_range": null
 },
 "misses": [],
 "matched_evidence": "<1 object(s)>",
 "corpus": {
  "mode": "offline_snapshot",
  "snapshot_generated": "2026-08-06T01:06:08.732897+00:00"
 },
 "...": "<10 fields total; not_advice line present: True>"
}
[exit 0]

$ yuclaw replay-lab docs/replay/lab_replay_bundle.json
Replay bundle built 2026-10-15 02:45 UTC from source commit eda1549ff47e
Ledger repo: https://github.com/YuClawLab/yuclaw-trust @ 925c9ab7cbf6

[forward] 100 rebalance periods, window ['2026-05-20', '2026-10-14']
  spread top_minus_bottom   mean/period -0.00259  p=0.320184  n=100  CI95=(-0.0076570546292354606, 0.0024497829429687896)
  spread top_minus_universe mean/period -0.00162  p=0.270732  n=100  CI95=(-0.004497031658379483, 0.0012196032104730685)
  IC  1d mean -0.008093758771769873 NW-p=0.735482 T=105 dates
  IC  5d mean -0.046099243108509505 NW-p=0.183462 T=101 dates
  IC 20d mean -0.07616357041380668 NW-p=0.044147 T=86 dates
  market-model vs_universe alpha/period -0.00170 beta 1.12 p=0.25183 n=100
  market-model vs_spy alpha/period -0.00142 beta 0.99 p=0.353117 n=100
[in_sample] 13 rebalance periods, window ['2026-02-18', '2026-05-18']
  spread top_minus_bottom   mean/period +0.00553  p=0.685718  n=13  CI95=(-0.019365001701843236, 0.030789287070102277)
  spread top_minus_universe mean/period +0.00586  p=0.496075  n=13  CI95=(-0.009809033927242129, 0.02165172551022834)
... (12 more lines)
[exit 0]

Protocol chain (statistics locked BEFORE computation, tamper-evident): curl -sO https://raw.githubusercontent.com/YuClawLab/yuclaw-brain/main/registry/protocols.jsonl · full clean-environment replication: make replicate.

How we compare → · Take the 5-minute tour →

What you'll find inside: the baseline test our own composite lost at current sample sizes — published under the pre-registered protocol · a label-calibration panel that says "directional meaning not yet demonstrated" · retired hypotheses preserved with their grounds · 500+ filings shown to be 3 evidence stories · robustness grids that print where results break.

Research and education only — not investment advice. Signal labels are research classifications, not buy/sell recommendations.

v8 — the local workbench and its four modules

yuclaw.ca is the research-content site. The v8 workbench is a different thing: an application that runs on your own computer — it binds 127.0.0.1 only, your data stays in the workspace folder you name, there is no hosted service and no account. It traces one financial commitment from an exact source passage through a typed, frozen claim, its revisions, exact calculation, history and review to an export that a fresh workspace re-verifies. Four modules work on the same records:

pip install yuclaw
yuclaw workbench selftest                                  # checks this installation in a temporary fictional workspace (no test tools needed)
yuclaw workbench principals init --workspace ~/yuclaw-ws   # optional: roles for the four modules; prints the first administrator's credential once
yuclaw workbench serve --workspace ~/yuclaw-ws             # then open http://127.0.0.1:8765   (Ctrl-C stops it)
yuclaw workbench guide                                     # the full guide, from the package (the same text as the link below)

In the browser: load a fictional example on the Workspace page, follow its seven steps, build the export, then verify it in a second, fresh workspace (current guide; full manuals: User Guide (EN) · Guide utilisateur (FR)). python -m v8.workbench … is the same program.

Status: experimental, local, owner-operated. Including the modules activates nothing — no role, approval, budget or study exists until you set one up. Protected SHD admission needs Linux with Landlock; on macOS, Windows or a kernel without it that one route stays closed and everything else works. No independent security review has been performed, and there has been no user study: human benefit is PENDING. What has and has not been demonstrated is on the evidence scoreboard, under its own as-of time.

v9 — research briefs with traceable AI assistance

A brief is research text whose individual sentences can be inspected for the exact source passage, the frozen v8 claim version and the registered calculation they rest on, how the text was produced, what changed since, and which checks could or could not be run. It runs locally beside the v8 workbench; the complete journey needs no API key, model download or hosted service. Five independent answers per statement — byte integrity, recorded origin, issuer trust, substantive support, time scope — are never merged into a score. Introduced in 9.0.0 (see the CHANGELOG for the release state).

yuclaw workbench brief example --workspace ~/yuclaw-workspaces/research   # fictional fixture → the acceptance brief (EN; --lang fr)
yuclaw workbench brief show --workspace ~/yuclaw-workspaces/research --brief brf-… --statement 3
yuclaw workbench brief export --workspace ~/yuclaw-workspaces/research --brief brf-…
yuclaw workbench brief verify ~/yuclaw-workspaces/research/exports/bpk-….zip --workspace ~/yuclaw-workspaces/fresh
yuclaw workbench brief guide --lang fr                                     # packaged quick start, both languages

Guides, schema reference and migration note: docs/guide/v9/. Deferred, not in 9.0: live provider generation/detection connectors, a local entropy-controlled sampler, detector calibration campaigns. Detector reports can be imported and inspected; NOT_DETECTED never proves human authorship, and no detector score is a percentage of AI-written text.

v7 — check → reproduce → challenge → document use (outsider verification, offline)

Everything below runs without an account or a hosted service. The outputs below were recorded on the 7.0.0 candidate checkout (operator runs; not outsider receipts); the commands are unchanged since, and the manifest digest is different for every later tree.

$ yuclaw packet build ./yuclaw-packet --source <checkout>
[packet] built ./yuclaw-packet: 10 files, source <sha>; verify with: yuclaw packet verify ./yuclaw-packet
$ yuclaw packet verify ./yuclaw-packet
[packet] SUCCESS — manifest b89617e9735ecd5d…; exact bytes verified and published Lab statistics/ledger roots reproduced from the frozen bundle; not an outsider receipt, not a research interpretation
$ # after one byte of an artifact is changed:
[packet] MISMATCH — manifest b89617e9735ecd5d…; first discrepancy: byte mismatch at docs/capabilities.json: expected 96beff4fa51b4f91…/3341 B, observed 4c2a56369d42a623…/3341 B

• yuclaw check-claim … now returns support_limits — source match, temporal eligibility, replay status and research_interpretation: NONE — plus an explicit unsupported-conclusion sentence ("a coverage statement, not a finding that the claim is false").
• yuclaw challenge --store <private> create … records a challenge bound to an exact artifact (type, sha256, byte length) and a claim id; a RESOLVED disposition must name a tested revised artifact and never erases the original finding.
• yuclaw decision --store <private> record … records a research decision (investigate further / discard / request evidence) bound to a packet manifest digest; export needs explicit permission.
• yuclaw receipts --store <private> … imports an attempt, verifies artifact bytes, records a designated review and derives counts. Primary counts are qualified attempts (failed and inconclusive retained); successful attempts are reported separately; package evidence (wheel, sdist) is separate from site checks; program-wide exact-artifact evidence is separate from a release's exact-target evidence (bound to an independently supplied target manifest, UNBOUND until then). A submission can describe an outcome; it cannot award itself qualification. yuclaw receipts --store <private> explain <receipt id> answers "why did this count / not count?" with reason codes; yuclaw receipts --store <private> history a.json b.json compares two published scoreboard snapshots (additions, supersessions rendered CORRECTED in their original window, disposition changes, per-window and absolute movement).
• The Evidence Scoreboard (evidence_scoreboard.html, /receipts/scoreboard.json, REST /v1/receipts/scoreboard, MCP get_evidence_scoreboard) shows witnesses, pilots, replications, audits, refusals, packet uses and challenges with their counting definitions. Zero and pending are real states; no trust score is computed. Registration of the receipt program is pending (owner decision), so counts are unwindowed.

Why this is different

• Statistics are registered before they are computed. Every published statistic names its protocol in a hash-chained, append-only registry (registry/protocols.jsonl); estimator changes are supersessions, never edits. Check the chain yourself:

  python3 - <<'EOF'
  import sys; sys.path.insert(0, 'tools')
  from yuclaw_protocol_registry import Registry
  Registry('registry/protocols.jsonl').verify_chain(); print('chain OK')
  EOF
  

• Self-audits publish as measured. A pre-registered champion-challenger test found a persistence baseline ahead of our own composite at the primary horizon — that table is on the Lab page, not in a drawer. The label-calibration panel prints where labels carry no demonstrated directional meaning.

• Lenses pass a published admission standard or they don't ship. The XLK lens is live because it passed the same admission standard the SMH lens registered; verdicts and their reasons print on each page.

• The evidence layer is machine-readable. llms.txt and evidence_index.json give AI agents stable URLs for every page, packet, and protocol.

What YUCLAW does

• Every signal traces to a filing. Each composite score decomposes into nine components, and every evidence event links to the SEC document it was extracted from — checked against the source text before any signal sees it.
• Every snapshot is hashed to a public, tamper-evident ledger — git-anchored, never edited. Daily signal sets are content-hashed and committed to yuclaw-trust before pages publish. Outages are disclosed, never backfilled.
• Every Lab chart is reproducible bit-for-bit. yuclaw replay-lab (or a standalone stdlib script) rebuilds the cohorts, recomputes every statistic, and re-derives every ledger hash root from published derived data.

Sixty seconds

pip install yuclaw
yuclaw demo                        # 3-minute guided journey — works offline, zero config
yuclaw why AMD --as-of 2026-05-20  # bundled offline signal, no backend needed

Live signals for all tickers need the local backend (docs/v4/backend_setup.md); the published-data commands work anywhere with no backend. The most recent recorded external-machine replication (see the Replication Log) ran replay-lab from a fresh venv with nothing but pip install yuclaw and reproduced 74 daily ledger roots exactly plus one anchored-subset day (6,165 leaf hashes recomputed) and every published statistic, exit 0. It exits non-zero on any mismatch.

What is in 6.0.x

Current package version: 9.1.0 — the release notes, the frozen wheel and sdist SHA-256 hashes, and the shipped-object list live on the GitHub Release for this version (all releases · CHANGELOG). 7.0.0 adds the receipt engine, the offline verification packet, structured claim-support limits, local challenges and document-use receipts, and the Evidence Scoreboard. Scientific states are unchanged: Gate #15 stays MANUAL_REVIEW, Phase 6 N_eff stays PENDING, C6 stays DESCRIPTIVE, U350 stays shadow-only. Zero outsider receipts exist at release; the scoreboard shows that as a real zero.

Command surface

yuclaw --help                      # Command list with one-line descriptions (also -h, help)
yuclaw why TICKER                  # Composite signal + ranked evidence w/ SEC source URLs
yuclaw check-claim --text "..."    # Evidence Passport — deterministic claim check (also --ticker/--type/--date-range/--accession)
yuclaw replay TICKER --date DATE   # Point-in-time signal at end of date
yuclaw replay-lab                  # Reproduce the Validation Lab from the public bundle
yuclaw validation                  # In-sample event validation + forward tracking ledger
yuclaw events --ticker SU --since 2026-05-01   # Accepted-events export (derived data only)
yuclaw lens canada --lens XEG      # Lens summary-card data as JSON (same numbers the page renders)
yuclaw export --lens GDX --format csv          # Lens events export; --page builds the evidence packet
yuclaw memo --ticker SU --days 30  # Evidence memo — grounded, citation-verified, linted (docs/usage.md)
yuclaw verify TICKER --date DATE   # Verified Research Ledger integrity check

Worked examples with real output: docs/usage.md.

Public signal vocabulary: STRONG_BULLISH, BULLISH, NEUTRAL, WATCH, WEAKENING, NEGATIVE_EVENT, BEARISH_WATCH, RISK_ALERT. There is no SELL or SHORT label — these are research classifications, not trade directions.

How it works

SEC EDGAR (Form 4 / 8-K / 10-Q / 10-K / 6-K / 20-F / 40-F)
  │
  ▼ systemd poller (always-on, 5-min sweep)
  ├──▶ Form 4 → deterministic XML parser (no LLM, zero GPU) → events table
  ▼ prose-first text acquisition (exhibit / MD&A prose; XBRL cover fallback)
  ▼ Llama 3.1 70B extraction + SourceLock Guard (checked against source text)
  ▼ events table — the evidence layer
  ▼ Layer-1 specialist swarm (10 specialists; risk channel kept SEPARATE from direction)
  ▼ 9-component composite (C1..C9)
  ▼ signal_snapshots (content-hashed)
  ├──▶ Verified Research Ledger (git-anchored, public)
  ├──▶ Forward Tracking Ledger (outcomes vs SPY at 1 / 5 / 20 days)
  ├──▶ Live landing + Validation Lab pages (regenerated daily)
  └──▶ SDK / REST / MCP server

132-name coverage: a 79-name scoring universe (equities + sector ETFs + broad ETFs + macro instruments) plus a 53-filer evidence tier (49 Canada Resources issuers + 4 SMH-lens foreign filers: ASML, NXPI, STM, TSM) — ingested and dashboarded, never scored; the boundary is machine-enforced.

Deep dives: system architecture, operations, hardware · OpenClaw / MCP integration · methodology.

Signal Validation Lab

A decile-cohort event study of whether YUCLAW's composite score carries forward information — methodology reviewed by two senior finance/methodology academics, plus a structured multi-AI critique loop:

• Regenerated daily after U.S. market close, freshness-stamped, with a staleness alarm in the health monitor.
• Statistical rigor panel: bootstrap confidence intervals, Newey–West and clustered inference, market-model alpha, and a statistical power meter that quantifies what the current n can and cannot detect.
• Statistics computed per-regime (in-sample vs forward) and never blended across the boundary.
• Reproduce this page: yuclaw replay-lab or the standalone stdlib script.

In the Lab's own words: "No forward alpha has been statistically proven yet."

🔬 Live: Signal Validation Lab · Today's Evidence Digest · Independent Replication Log · Methodology: docs/methodology/validation_lab.md

Hypothetical research illustration — not investment advice, not performance advertising.

Methodology & honest limitations

Full methodology lives in docs/methodology/backfill.md. The honest limits, stated up front:

• The forward record is young. Forward tracking began 2026-05-20; roughly 50 trading days of look-ahead-free history exist as of early August 2026 — enough to display, not enough for statistical significance; the Lab's power meter quantifies this.
• In-sample is replay reconstruction, not a live backtest. The in-sample panel was materialized after the fact by the replay engine; the canonical look-ahead statement below says exactly what the extraction model could and could not have seen, and why in-sample results stay systematically optimistic.
• C6 risk channel is partially confirmed. Rareness confirmed OOS; the sign question remains open — the first read under the registered v2 protocol printed INCONCLUSIVE (2026-07-30) and accrual continues. The sign confirmation is the gate for Layers 2–10, and it has not been met.
• C4 macro regime is temporarily frozen as of 2026-05-18 with a staleness disclosure, pending macro-engine restoration. C1/C3/C5/C7 read live price_history; C6/C8/C9 remain point-in-time exact.
• Jun 26 – Jul 3, 2026 outage — disclosed, not patched. A network outage froze price-derived inputs at Jun 25 closes while snapshots continued point-in-time on-box. No snapshot or ledger row was retroactively edited.
• No table of headline % returns appears in this README. Hit rates are reported alongside their n on the live validation page; small-n panels are tagged.

Look-ahead statement — canonical text, byte-identical on every surface (README, methodology, Validation Lab page; checked by the copy-consistency gate):

In-sample look-ahead statement. The in-sample replay rows (signal dates 2026-02-18 to 2026-05-13, evidence window 2026-02-18 to 2026-05-17) were built from evidence events extracted by one language model, Meta Llama 3.1 70B Instruct (served locally as yuclaw-llm-70b), whose published pretraining cutoff is December 2023 (Meta model card). Form 4 events in the window come from a deterministic XML parser with no language model. The earliest in-sample date is about 26 months after that cutoff, so no filing text in the window could have been seen in training: there is no parametric look-ahead from the filings themselves, and the model's general market knowledge also ends before the window begins. The replay engine flags any as-of date before 2024-07-01 as inside the model's training window; no in-sample date triggers it. In-sample results nonetheless remain a replay reconstruction, not a live record: the scoring design was finalized in May 2026, after the window it is replayed over, market-layer components read approximated inputs, and no in-sample signal was exposed to external challenge in real time. In-sample results are therefore treated as systematically optimistic and educational only; the forward record (signal dates from 2026-05-20) is the look-ahead-free record.

Community

⚠️ Disclaimer

YUCLAW is open-source research and educational software. It is NOT financial advice, investment advice, or a recommendation to buy, sell, or hold any security. All signals, scores, and analyses are generated by automated AI models and may contain errors.

Past performance does not guarantee future results. Trading involves substantial risk of loss. You are solely responsible for your own investment decisions. Consult a licensed financial advisor before making any investment.

YuClawLab, its contributors, and affiliates accept no liability for any losses arising from use of this software.

For educational and research purposes only. See docs/methodology/backfill.md and DISCLAIMER.md for the long-form versions.

For AI agents & researchers

The open evidence layer for financial AI.

Agents citing YUCLAW inherit accession-verified, point-in-time, hash-anchored evidence. Start at capabilities.json — one URL discovers the why-JSON API, the Evidence Passport, the schemas, EvidenceBench, and the MCP tools (full pitch).

• Start here: llms.txt and the machine-readable evidence_index.json (every page, packet, and protocol with stable URLs and data-through dates).
• Objects: the five frozen v1 JSON Schemas — SignalSnapshot, EvidenceEvent, ResearchProtocol, RobustnessCell, ResearchMemo — at /schemas/; today's real outputs validate against them in the daily gate suite.
• Consume: evidence packets (derived statistics, event CSVs, engine run JSONs, metadata + citation snippets) from /packets/; the MCP server exposes why / memo / events / lens / universe / validation / verify as tools with friendly no-backend behavior.
• Cite: use the CITATION.txt inside any packet; event-level citations use event IDs resolvable in the packet CSVs.
• Verify: pip install yuclaw && yuclaw replay-lab recomputes the published Lab statistics from the public bundle; the protocol registry (registry/protocols.jsonl) is hash-chained and append-only.
• Rules: derived statistics only; preserve the disclaimers and the frozen implication line when quoting inference; nothing here is advice or a recommendation.