Skip to content

AGENTS.md

Latest commit

 

History

History
126 lines (97 loc) · 4.79 KB

AGENTS.md

File metadata and controls

126 lines (97 loc) · 4.79 KB

AGENTS.md - Operating Rules for AI Agents

This file defines required behavior for any AI agent operating in this repository.

Mission

Produce independent, reproducible AI governance research artifacts that can be verified by third parties and cited publicly.

Primary outputs:

  • OpenClaw case-study report (openclaw-2026)
  • AI tool sprawl flagship report (ai-tool-sprawl-q1-2026)

Non-Negotiable Standards

  1. Reproducibility over rhetoric.
  • Every headline claim must map to artifact + deterministic query in claims/.
  1. Independent research tone.
  • No hype language.
  • No claims without evidence.
  • No product CTA language in report body sections.
  • Explicit limitations and threats-to-validity sections are mandatory.
  1. Deterministic first.
  • Treat deterministic baseline as canonical.
  • Enrich-derived claims require explicit provenance (as_of, source) and separate labeling.
  1. No production risk.
  • OpenClaw execution must occur in isolated containerized lab conditions only.
  • Never use production credentials, customer data, or unrestricted side effects.
  1. Publish gates are hard gates.
  • A report does not publish unless validation and claim/threshold checks pass.

Canonical Control Files

  • Strategic context: internal/WHY.md
  • OpenClaw definitions: reports/openclaw-2026/definitions.md
  • OpenClaw protocol: reports/openclaw-2026/study-protocol.md
  • OpenClaw preregistration: reports/openclaw-2026/preregistration.md
  • Sprawl definitions: reports/ai-tool-sprawl-q1-2026/definitions.md
  • Sprawl protocol: reports/ai-tool-sprawl-q1-2026/study-protocol.md
  • Sprawl preregistration: reports/ai-tool-sprawl-q1-2026/preregistration.md
  • Headline rubric: internal/headline_rubric.md
  • Claim ledgers: claims/*/claims.json
  • Threshold policy: pipelines/config/publish-thresholds.json
  • Citation logs: citations/*.md

If these files conflict with draft notes, control files win.

Required Workflow

A) Before execution

  1. Confirm definitions and protocol version.
  2. Confirm preregistration file exists and lock fields are set for the planned run.
  3. Confirm citation logs exist for timeline/regulatory assertions.
  4. Preflight run scaffold with pipelines/*/run.sh --run-id <id> --dry-run.
  5. Create immutable run scaffold with pipelines/*/run.sh --run-id <id> (or --resume only for existing IDs).
  6. Keep run-script options explicit in manifests (execution mode, workload mode, guardrails).

B) During execution

  1. Write artifacts to:
  • runs/openclaw/<run_id>/...
  • runs/tool-sprawl/<run_id>/...
  1. Keep raw and derived outputs separated.
  2. Preserve exact policy/config snapshots used in run.

C) Before manuscript finalization

  1. Update claim values in the report claim ledger:
  • claims/openclaw-2026/claims.json or
  • claims/ai-tool-sprawl-q1-2026/claims.json
  1. Validate:
    • pipelines/openclaw/validate.sh or
    • pipelines/sprawl/validate.sh
  2. For publish readiness:
    • pipelines/openclaw/validate.sh --run-id <id> --strict or
    • pipelines/sprawl/validate.sh --run-id <id> --strict

D) Before publication

  1. Assemble package:
    • pipelines/openclaw/publish_pack.sh --run-id <id> or
    • pipelines/sprawl/publish_pack.sh --run-id <id>
  2. Ensure bundle hash manifest exists (bundle.sha256).
  3. Confirm all links and artifact references are real and resolvable.

Writing Guardrails

  • OpenClaw Section 1 is facts-only timeline (max three paragraphs).
  • OpenClaw Section 3 is brand-neutral data section (no product messaging).
  • Sprawl report follows 10-section canonical structure.
  • Gait deep analysis belongs in OpenClaw report; Sprawl references Gait only in recommendations context.
  • No install commands, pricing language, or product CTA statements in report body sections.
  • News/social incident references may be used only as context unless directly measured in run artifacts.
  • Every report must include a headline integrity block with:
    • headline number
    • denominator
    • run ID
    • artifact path
    • deterministic query
  • Headline selection must follow internal/headline_rubric.md (minimum score threshold).
  • Every report must include fixed methodological disclosure headings:
    • Limitations
    • Threats to Validity
    • Residual Risk
    • Reproducibility Notes

Change Control

  • Do not silently change metric definitions, threshold policy, or schemas.
  • If changed:
    • update version marker in definitions/protocol/schema docs
    • rerun affected metrics
    • update claims and cite change in commit message

Anti-Patterns (Disallowed)

  • Backfilling narrative claims without data proof.
  • Mixing enriched and deterministic metrics without clear labeling.
  • Publishing production-write percentages without configured production-target policy.
  • Treating non-allow outcomes as executable in governed lane logic.
  • Omitting limitations to make results look stronger.