Skip to content

Latest commit

 

History

History
89 lines (65 loc) · 3.08 KB

File metadata and controls

89 lines (65 loc) · 3.08 KB

Research Command Chain

ZERO Research is the public, fixture-backed version of the private research loop. It gives operators and coding agents the missing analysis layer between local memory, genesis proposals, and evolve gates.

The report artifact uses zero.research.report.v1. The API snapshot uses zero.research.snapshot.v1. Source trust decisions use zero.research.source_classification.v1; aggregate source-quality records use zero.research.source_quality.v1.

Commands

The public chain runs seven deterministic commands:

  • hunt: market scan from public paper candle fixtures;
  • edge: accepted/rejected decision analysis from local paper journals;
  • convergence: feedback-loop drift and lockstep detection;
  • thesis: seven-day operating hypothesis, anti-thesis, and scorecard;
  • score: prior-judgment scoring against public paper outcomes;
  • meta: command usefulness audit;
  • sharpen: system-improvement backlog.

Safety Boundary

Research is paper-only and read-only.

  • It does not place orders.
  • It does not mutate the checkout.
  • It does not open branches, commits, pull requests, or deployments.
  • It does not claim live PnL or live edge.
  • It removes private keys, wallet material, venue order ids, raw venue payloads, prices, sizes, quantities, and notionals from public artifacts.
  • It classifies source trust before a source can influence paper research. Prompt-injected, untrusted, unsupported performance, or risk-increasing claims are rejected into quarantine metadata without echoing raw source text.

Live research conclusions require signed operator evidence before they can be used as public proof.

Source Quality

Every research report includes source_quality:

  • accepted: fixture, signed-operator, or repository-document sources that can be used for paper research;
  • rejected: untrusted or adversarial sources that can only be used for quarantine records and safety review;
  • raw_content_included=false: public reports never carry raw source text.

The adversarial fixture coverage lives in engine/tests/test_research.py::test_research_source_classifier_rejects_prompt_injection_without_echoing_raw_text.

Local Example

PYTHONPATH="$PWD/engine/src" python3 -m zero_engine.research run \
  --repo-root "$PWD" \
  --output artifacts/research/research.json \
  --now 2026-05-01T00:00:00Z

PYTHONPATH="$PWD/engine/src" python3 -m zero_engine.research status \
  --report artifacts/research/research.json \
  --now 2026-05-01T00:00:00Z

Installed entrypoint:

zero-research run --repo-root . --output artifacts/research/research.json

API And MCP

The local paper API exposes:

curl -fsS http://127.0.0.1:8765/research

The MCP server exposes zero_get_research_report and the zero://research/report resource. Both are read-only and fixture-backed.

Relationship To Evolve

Research is not /evolve. Research explains what the system should inspect next. Genesis turns evidence into reviewable proposals. Evolve runs guarded paper-only builder, red-team, canary, and calibration gates. Promotion still requires human review.