Skip to content

v4.14.0 — Evidence Integrity and Canonical Boundaries

Choose a tag to compare

View all tags
@Patdolitse Patdolitse released this 10 Jul 13:39
v4.14.0
0719bdf

Added

  • Canonical product-boundary contract and guard: public package roots, optional extensions, tool surface, export surfaces, public docs, release evidence, and publish allowlist now share a machine-readable boundary in docs/public-facts.json.
  • LIVE_SMOKE reliability and evidence integrity tooling: run records now preserve failure classes, bounded metadata, validated aggregate evidence, and append-only local history without exposing raw memory bodies, local paths, credentials, or transcripts.
  • Longitudinal real-use evidence evaluator: local collection and report tooling can track first-value evidence over time, while explicitly distinguishing real-use accumulation from synthetic or operational checks.

Changed

  • search_knowledge now routes through a thin application service shared by CLI and MCP adapters, keeping behavior parity while preserving existing protocol boundaries.
  • Provenance and trust semantics are consolidated into canonical helpers for source agent, confirmation source, trust projection, freshness, and recall-visible provenance; unsafe labels and credential-shaped identifiers fail closed.
  • Public facts and test collection now use a deterministic, isolated environment profile, keeping the frozen MCP surface at 57 total tools: 17 core and 40 advanced.

Fixed

  • Product-boundary and release-surface guards now redact private-looking paths and unsafe public terms in diagnostics instead of echoing matched content.
  • Public fact sync now seals host-environment drift from PYTHONPATH, pytest plugin autoload, and ambient Engram stores during collection.
  • Synthetic nested memory-eval validation was tightened so aggregate verdicts cannot overstate readiness.

Documentation

  • Public docs keep remote telemetry and feedback as separate explicit opt-ins, still off by default.
  • First-value and longitudinal evidence wording stays conservative: local first-value collection begins accumulating only after the client is restarted, and synthetic or operational evidence does not verify real-use longitudinal readiness.
Assets 2
Loading