Skip to content

Latest commit

 

History

History
78 lines (65 loc) · 3.34 KB

File metadata and controls

78 lines (65 loc) · 3.34 KB

Roadmap and metrics

Now: M0 technical hypothesis

  • Source-bound strict manifest
  • Four obligation classifications
  • Bounded local JSON evaluation
  • Content-bound manual and external attestations
  • Explicit overall-status algebra
  • Deterministic receipt payload and untrusted envelope
  • Offline integrity verification and evidence replay
  • Synthetic example spanning every classification
  • Adversarial usefulness, duplication, adoption, and safety review
  • Complete ≥90% branch coverage gate
  • First clean make verify
  • Descriptor-stable evidence parsing and hashing
  • Closed semantic receipt verification
  • Bounded two-rater discovery metrics utility
  • Typed-wheel content gate
  • Redacted-by-default, manifest-bound evidence collection plan

Next: discovery, not features

  • Map three public SOWs from different jurisdictions.
  • Measure classifiable-clause rate and independent agreement.
  • Interview 8–12 acceptance, procurement, QA, accessibility, security, prime, and IV&V practitioners.
  • Observe one real acceptance-evidence assembly workflow.
  • Obtain one serious, nonbinding design-partner next step.
  • Select exactly one vertical pack.

M1 only if proceed thresholds pass

  • Signed manifest-approval and review roles.
  • Evidence validity windows and explicit staleness.
  • Coverage denominator: every in-scope clause must be classified.
  • One adapter family selected from observed workflow.
  • Static accessible trace report.
  • Amendment diff and invalidation rules.
  • Hosted CI and signed release pipeline.

Later

  • OCDS milestone export.
  • Cross-organization evidence exchange.
  • Proposed-mapping review queue.
  • Model-assisted mapping research with a committed gold set.

Kill rules

Stop or reframe if:

  • fewer than 25% of sampled consequential clauses are objectively classifiable;
  • independent experts cannot reach κ 0.50 on classification;
  • the primary buyer wants document storage rather than evidence replay;
  • no budget-adjacent participant considers the bounded outcome purchasable;
  • the first design partner requires legal conclusions or compliance certification;
  • maintaining adapters dominates the obligation/evidence model; or
  • existing acceptance-management software already solves the observed workflow.

Technical metrics ledger

Attribute M0 target Evidence
Correctness 0 false passes from missing/unbound evidence tests
Reproducibility identical input payload digest tests/demo
Security 0 accepted absolute/traversal paths tests
Maintainability strict typing; complexity ≤10 make verify
Coverage ≥90% branch make verify
Dependency risk 0 runtime dependencies pyproject.toml
Artifact bound ≤2 MiB per evidence/receipt JSON code/tests
Structure bound ≤64 JSON levels and ≤100,000 nodes code/tests
Snapshot binding parsed evidence bytes equal hashed bytes code/tests
Package typing wheel contains py.typed and all runtime modules make package-check
Performance reproducible local median/p95 report; no invented threshold make benchmark
Planning privacy local paths/locators/reasons require explicit opt-in code/tests
Planning currentness checksum self-check separated from exact manifest regeneration code/tests