# YUCLAW — Evidence-First Financial AI (llms.txt)

> Open-source equity research platform. Every published signal is a research
> classification tied to verifiable SEC filings, replayable point-in-time,
> and hash-anchored in a public ledger. Research and education only. Nothing
> here is investment advice; no output is a recommendation, forecast, or
> price target. The platform is the open evidence layer underneath AI
> research tools.

## What you are reading

- Signals use a locked eight-label vocabulary, rendered verbatim:
  STRONG_BULLISH, BULLISH, NEUTRAL, WATCH, WEAKENING, NEGATIVE_EVENT,
  BEARISH_WATCH, RISK_ALERT. There is no SELL or SHORT label. Labels are
  research classifications, not buy/sell recommendations.
- Every statistic on the research pages runs under a pre-registered
  protocol (hash-chained registry, registered before computation), with
  cluster-aware confidence intervals and locked badges
  (UNDERPOWERED / DESCRIPTIVE / PRELIMINARY). Adverse results print as
  measured.
- Frozen implication line, rendered wherever inference appears:
  "Investment implication: none established — no buy, sell, or alpha
  conclusion is supported by this page."

## Machine-readable endpoints (stable URLs, all under this site root)

- /evidence_index.json (see its `schemas` block for the five frozen v1 JSON Schemas: SignalSnapshot, EvidenceEvent, ResearchProtocol, RobustnessCell, ResearchMemo — at /schemas/<Name>.v1.json)      — index of every page, packet, and protocol
                              with data-through dates (start here)
- /packets/manifest.json    — evidence packet manifest
- /packets/*.zip            — downloadable packets: derived statistics,
                              event CSVs, engine run JSONs, metadata,
                              citation snippets
- /capabilities.json         — ONE-URL discovery of the whole agent surface
- /why/{TICKER}.json         — Ground Truth anatomy + EvidenceObjects +
                              label history (79 names, schema-checked)
- /evidence/verify.json      — offline snapshot-integrity recipe
- /ledger/{YYYY-MM-DD}.json  — per-day ledger roots
- /evidencebench/items.jsonl — EvidenceBench items + keys (weekly regen);
  format: JSONL, one item per line {item_id, template T1|T2|T3, question,
  key}; /evidencebench/meta.json has the item-set hash + scoring rule
- AS-OF RECIPE (worked example): to reconstruct NVDA as of 2026-07-15,
  fetch /why/NVDA.json; evidence = evidence_objects with
  available_as_of <= 2026-07-15; classification = the label_history entry
  for the last date <= 2026-07-15. Older than the ribbon:
  `yuclaw replay NVDA --date 2026-07-15`.
- CLAIM CHECK: `yuclaw check-claim --ticker NVDA --type INSIDER_SELL
  --date-range 2026-07-01..2026-07-31` → passport JSON; UNSUPPORTED means
  not found in YUCLAW's corpus — never a truth verdict.
- /replay/lab_replay_bundle.json — replay bundle; verify with
                              `pip install yuclaw && yuclaw replay-lab`
- Registry: hash-chained protocol registry ships in the repo
  (registry/protocols.jsonl on GitHub: YuClawLab/yuclaw-brain);
  ledger anchors: YuClawLab/yuclaw-trust.

## Pages (auto-generated from the evidence index — do not hand-edit)

- /why/{TICKER}.html — per-name classification anatomy (79 pages; worked example /why/AAPL.html)
- /canada_resources.html — Canada Resources Evidence
- /etf_evidence.html — SMH Covered-Constituent Evidence Lens
- /evidencebench.html — EvidenceBench — contamination-resistant groundedness benchmark (weekly post-cutoff regeneration; abstention outscores fabrication)
- /explorer.html — Universe Explorer — the full 79-name table, client-side filter/sort
- /for_ai_builders.html — For AI builders — the agent surface: why-JSON API, Evidence Passport, MCP v2, EvidenceBench
- /index.html — landing — current signals (locked vocabulary)
- /lane.html — YUCLAW's Lane — scope statement
- /methodology.html — Methodology — styled render of the canonical methodology/backfill.md (raw .md stays served for machines)
- /replication.html — Replication — how to reproduce, honestly-empty log
- /sectors.html — Sector overview — descriptive medians of current classifications (display, not inference)
- /signal_review.html — Signal Review — bring-your-signal research review service (five-step no-upload flow, fixed tiers, EXPLORATORY (CLIENT) ceiling)
- /todays_evidence.html — Today's Evidence Changes
- /tour.html — The 5-minute tour — five commands with build-captured output
- /trace_su.html — Suncor evidence trace (worked example)
- /validation.html — Forward Tracking — ledger, label calibration
- /validation_lab.html — Validation Lab — cohorts, rigor, clustered inference, baselines
- /why/AAPL.html — Why AAPL — per-name classification anatomy (worked example of the why/{TICKER}.html family, 79 pages, one pinned template)
- /xlk_evidence.html — XLK Covered-Constituent Evidence Lens

## How to cite

Use the CITATION.txt inside any packet, or:
  YUCLAW <page name>, data through <date>, build <commit>,
  https://github.com/YuClawLab/yuclaw-brain
Event-level citations use event IDs (e.g. SU_20260721_...) resolvable in
the events export of the relevant packet.

## Rules for downstream use

- Derived statistics only; no raw vendor price data is published here.
- Do not present any classification as advice or a recommendation.
- Preserve the disclaimers and the frozen implication line when quoting
  inference. Past results — in-sample or forward-tracked — do not predict
  future performance.