Skip to content

Latest commit

 

History

History
33 lines (24 loc) · 6.02 KB

File metadata and controls

33 lines (24 loc) · 6.02 KB

specs/ — normative methodology output

The intent-eval-lab umbrella covers research and experimentation across multiple classes of operationally-deployed inference systems. The specs/ tree is where that research crystallizes into normative methodology — versioned, testable, vendor-neutral specifications that codify what "operationally ready" means for each class.

Each module under specs/ covers one class of inference system. Each module ships its own versioned spec, conformance test suite, and per-vendor case studies.

Modules

Module Class of system First case study Status
mcp-plugin-observability/ MCP plugins for agentic CLIs (Claude Code, Gemini CLI, Codex CLI, Copilot CLI) (engagement-private; not for public release without partner consent) v0.1.0-draft (in progress)
prompt-evaluation/ Prompts + context templates as eval targets (the prompt→context-engineering unit, generalizing beyond skills) (pending) v0.1.0-draft skeleton — gated; see 043-DR-RFC
validator-contract-reliability/ AI-generated validator contracts and cryptographic primitives (pending) placeholder — RFC welcome
forecasting-drift-detection/ Time-series forecasting under data-shift conditions (pending) placeholder — RFC welcome
decentralized-crypto-evaluation/ Decentralized cryptographic agents and key-material handling (pending) placeholder — RFC welcome
methodology/ Cross-module patterns (Toulmin findings, diagnostic-first auditing, conformance-vs-eval distinction) n/a placeholder
tier-bridge/ Platform-internal: how a skill's static tiers (grading + static production gate) compose with the behavioral (j-rig) tier into one promotion verdict n/a (platform-internal) v0.1.0-draft (normative)
taxonomy/ Platform-internal: the 7-layer testing taxonomy (git-hooks → static → unit → integration → system → E2E → acceptance) the audit-tests / implement-tests skills reference n/a (platform-internal) v0.1.0-draft (normative)
skill-refiner-evidence-report/ Platform-internal: the per-pass Skill Refiner evidence report — the human-readable + machine-verifiable view of one skill-refiner-pass/v1 attestation n/a (platform-internal) v1.0.0-draft (normative)

Most modules cover a class of inference system; the last three (tier-bridge, taxonomy, skill-refiner-evidence-report) are platform-internal process specs — they codify how the platform's own tooling composes, not an external system class. Modules are added as engagements warrant. Placeholder READMEs exist to signal intent and welcome RFC-style input from practitioners working in those domains; they don't represent committed work without a corresponding case study.

Spec authoring conventions

  • Versioning — semantic, with explicit vN.M.K-draft / vN.M.K-rc / vN.M.K tags. A spec stays at -draft until at least one independent implementation ships against it.
  • Anchoring — every normative requirement cites the canonical upstream source (Anthropic docs, vendor specs, RFC, ISO standard, etc.) it codifies. Specs are codifying what's already canonical, not inventing.
  • Test suites — every module ships a reference test suite under conformance-test-suite/. A claim of conformance without a test pass is not a claim.
  • Case studies — each module's case-studies/ dir holds vendor-specific instantiations. Case studies are illustrative, not normative; they show how the spec applies to a real system.
  • Multi-vendor by default — module specs are vendor-neutral. If a requirement only applies to one vendor's surface, it goes in that vendor's case study, not the spec.

License

Apache 2.0 — see LICENSE at repo root.