Open-Source Evidence-First Financial Research Platform
Composite research signals tied to SEC filings, time-machine replay across a 90-day evidence window, and a public git-anchored Verified Research Ledger for tamper evidence. Research and education only — not investment advice. Signal labels are research classifications, not buy/sell recommendations.
Live Dashboard · Quickstart · Methodology · Disclaimer · API Terms · PyPI
pip install yuclaw-evidence
python3 -m v3.cli why NVDASample output:
NVDA composite score: +0.299 (signal label: NEUTRAL)
Components (score × weight × confidence):
C1 Momentum +0.46 (weight 0.12)
C2 Volume +0.00 (weight 0.08)
C3 Sector -0.15 (weight 0.12)
C4 Macro +0.60 (weight 0.15)
C5 Oil/Rates/FX -0.47 (weight 0.05)
C6 Event Impact +0.16 (weight 0.18)
C7 Peer Corr +0.95 (weight 0.10)
C8 Cascade +0.00 (weight 0.12)
C9 Model Trust +0.00 (weight 0.08)
Top contributing events (last 7 days):
↑ +0.02 2026-05-14 M_AND_A_CLOSE (d1 cascade)
CASCADE d1 via HPE→NVDA(supply,w=0.15) from HPE: H3C divestiture
source: https://www.sec.gov/Archives/edgar/data/1645590/...
Compliance: Research only. Not financial advice. Not a registered investment advisor.
python3 -m v3.cli why TICKER # Composite signal + ranked evidence w/ SEC source URLs
python3 -m v3.cli replay TICKER --date DATE # Point-in-time signal at end of date
python3 -m v3.cli validation # In-sample event validation + forward tracking ledger
python3 -m v3.cli brief # Personalized digest (uses ~/.yuclaw/profile.json)
python3 -m v3.cli watch add TICKER # Manage local watchlist
python3 -m v3.cli verify TICKER --date DATE # Verified Research Ledger integrity check
python3 -m v3.cli profile show # Local preferencesPublic signal vocabulary: STRONG_BULLISH, BULLISH, NEUTRAL, WATCH, WEAKENING, NEGATIVE_EVENT, BEARISH_WATCH, RISK_ALERT. There is no SELL or SHORT label.
yuclaw l2: real iceberg detection requires a Level-2 data feed. Without one, the command returnsN/Ainstead of fabricated microstructure.
yuclawlab.github.io/yuclaw-brain — refreshes every 30 minutes from cron.
- Evidence-first composite signals. Every YUCLAW signal traces back to a verifiable SEC filing or deterministic supply-chain cascade — no opaque "model said so". The 9-component composite (momentum, volume, sector velocity, macro regime, oil/rates/FX, event impact, peer correlation, supply-chain cascade, model trust) is confidence-weighted; C6 event impact carries the highest single weight (0.18), by design.
- SEC EDGAR ingestion + SourceLock Guard. Form 4 / 8-K / 10-Q / 10-K / 6-K filings ingested via local Llama 3.1 70B (Ollama). A deterministic SourceLock Guard validates every LLM extraction against the source text before any signal sees it.
- Time-machine replay. Any signal can be recomputed as of a past date with point-in-time filtering (
available_as_of <= as_of). Leak-audited; reproducible via theyuclaw replayCLI / REST/replay/ MCPyuclaw_replay. - In-Sample Event Validation + Forward Tracking Ledger. Two clearly separated panels: in-sample is replay-reconstructed (~1,000 snapshots over a 90-day window), forward is live-emitted from launch onward. Hit rates always reported alongside their
n— never a headline percentage alone. - Verified Research Ledger. Each day's signal hashes are committed to a public git repo (yuclaw-trust). Anyone can
yuclaw verify TICKER --date DATEto confirm a signal hasn't been edited since publication. This verifies record integrity and timing — not investment merit. - Multi-surface access. Python SDK (
pip install yuclaw-evidence, import asyuclaw_py), REST API, FastMCP stdio server (7 tools), CLI (yuclaw why / replay / validation / brief / watch / verify / profile). - Local LLM inference. Llama 3.1 70B (Q4_K_M, ~42 GB) via Ollama on NVIDIA DGX Spark GB10. Zero cloud LLM dependency for extraction. SEC EDGAR is the only external data source for the evidence layer.
- ~80-ticker universe. Equities + sector ETFs + broad ETFs + macro instruments.
Full methodology lives in docs/methodology/backfill.md. The honest limits at launch:
- In-sample is replay reconstruction, not a live backtest. The In-Sample Event Validation panel was materialized after the fact by the replay engine — not emitted live.
- 5 of 9 components are point-in-time approximations. C1 momentum, C3 sector velocity, C4 macro regime, C5 oil/rates/FX, C7 peer correlation read a market-data cache that holds only the latest snapshot. On historical replays they self-degrade to confidence 0.3 with an explicit warning. C6 event impact, C8 cascade, and C9 model trust are point-in-time exact in-sample. v3.1 will land historical market data so the full composite runs point-in-time.
- Forward Tracking Ledger starts at n=0. Launch is Day 0. 1-day outcomes mature next trading day; 5-day a week later; 20-day a month later. The forward panel looks sparse for the first few weeks — correct, not a bug.
- Extreme labels are rare by construction. STRONG_BULLISH and BEARISH_WATCH require broad component agreement plus at least one material non-insider event. Day-0 OOS 99th percentile sits at +0.531, just below the +0.55 STRONG_BULLISH floor. See
docs/methodology/backfill.md§8 for the full reachability analysis.
No table of headline % returns appears in this README. Hit rates in both panels are reported alongside their n; small-n panels are tagged "preliminary". See yuclawlab.github.io/yuclaw-brain/validation.html for the live numbers.
graph TD
A[SEC EDGAR — Form 4 / 8-K / 10-Q / 10-K / 6-K] --> B[Llama 3.1 70B via Ollama + SourceLock Guard]
B --> C[events table — evidence layer]
C --> D[9-component composite C1..C9 — C6 event impact weight 0.18]
D --> E[signal_snapshots — content-hashed]
E --> F[Verified Research Ledger — git-anchored, public]
E --> G[Forward Tracking Ledger — outcomes vs SPY at 1/5/20d]
E --> H[Live landing + validation pages]
E --> I[SDK / REST API / MCP server]
v3/
signal/ 9-component composite (C1..C9), supply-chain graph, cascade engine
sources/ SEC EDGAR poller + backfill + Form 4 deterministic parser
extract/ LLM extraction + SourceLock Guard
replay/ Time-machine replay engine
track/ price_history + outcome_updater + In-Sample Validation panels
proof/ Verified Research Ledger writer + verifier
radar/ Change detector + Telegram/Email/Slack adapters
api/ FastAPI REST server
mcp/ FastMCP stdio server (7 tools)
cli/ why / replay / validation / brief / watch / verify / profile
signal/healthcheck.py Daily pipeline gate
sdk/ yuclaw-evidence — public SDK (pip install yuclaw-evidence)
docs/methodology/backfill.md v3.0 methodology + limitations + leak audit
This is the live cron table as of v3.0.0. Frequencies are read from crontab -l, not aspirational.
| Engine | Frequency | What it does |
|---|---|---|
| v3.0 daily pipeline | weekdays 17:00 MDT | healthcheck → snapshot_writer → outcome_updater → radar → proof.ledger → refresh_v3_pages — single chained pipeline, && short-circuits on failure |
| Ollama check | every 30 min | sanity ping to local Ollama |
| Health monitor | every 30 min | /tmp/yuclaw_health.log |
| Sentiment archive | every 4 hours | output/sentiment/*.json (research-side, orthogonal to v3.0) |
| Oil intelligence | hourly | output/oil/YYYY-MM-DD_brief.json (research-side) |
| Oil brief | nightly 23:00 MDT | LLM oil synthesis (research-side) |
| Swarm debate | nightly 23:00 MDT | Bull/Bear/Oracle LLM debate (research-side) |
| ATROS daemon | daily 18:15 MDT | alert + AutoDream summary (research-side, pre-v3.0) |
| PyTorch check | daily 22:00 MDT | dependency sanity |
v3.0 retired the v2.3.0 refresh_dashboard.sh, nightly_score_refresh.sh, and yuclaw.telegram.broadcast_bot cron lines. The single signal pipeline is now the daily 17:00 MDT chain above; Telegram broadcasts go through v3.radar.run to @yuclaw_signals when material changes are detected.
- GPU: NVIDIA Grace Blackwell GB10 (128 GB unified memory)
- LLM: Llama 3.1 70B (Q4_K_M, ~42 GB on GPU, 80 layers) served via
Ollama. Exposed locally as the
nemotron-3-super-localOllama tag with a financial-analyst system prompt. The real Nemotron 3 Super 120B is wired inyuclaw/core/router.pyas a dormant OpenRouter fallback (sm_121a-blocked on the vLLM path); the active production path uses Llama 3.1 70B locally. - Measured generation speed: ~2.2–2.7 tok/s on 50-token completions
(rendered live in the dashboard's TOK/S stat card —
output/inference_stats.jsonis rewritten by every nightly cron run). - Signal cycle: ~39 s end-to-end for the score-regeneration pipeline.
# As an OpenClaw skill
bash <(curl -s https://raw.githubusercontent.com/YuClawLab/yuclaw-brain/main/yuclaw/openclaw/install.sh)
# Or as MCP server
python3 yuclaw/openclaw/mcp_server.py # listens on port 8002| Dashboard | yuclawlab.github.io/yuclaw-brain |
| @Vincenzhang2026 | |
| GitHub | YuClawLab |
| PyPI (v3.0) | pypi.org/project/yuclaw-evidence |
| Methodology (v3.0) | docs/methodology/backfill.md |
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. MIT Licensed.
See docs/methodology/backfill.md and
DISCLAIMER.md for the long-form versions.
Released under the MIT License — free for everyone.
Built on NVIDIA DGX Spark GB10 · Llama 3.1 70B via Ollama · Local inference · Git-anchored Verified Research Ledger