Skip to content

ContextLattice v3.11.0 - Cognition Core

Choose a tag to compare

View all tags
@sheawinkler sheawinkler released this 11 Jul 22:25
v3.11.0
35254e0
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

ContextLattice v3.11.0 - Cognition Core

v3.11 turns memory from stored material into a bounded decision surface. Agents can ask for a Synthesis Pack, receive the evidence that matters most, keep slow retrieval warming in the background, and close the loop with observed outcomes instead of guesses.

What changed

  • Added deterministic synthesis_pack.v1 responses through contextlattice_synthesis_pack, POST /memory/synthesis-pack, and POST /tools/synthesis_pack.
  • Preserved ranked evidence, prompt sections, and compiler metadata even when context must collapse to the sparse boundary.
  • Added a bounded async inbox with dedupe and acknowledgement state. Pending or running continuations are warming; only terminal failures are degraded.
  • Added idempotent context-pack outcome capture from task-worker completion paths.
  • Separated all observed outcomes from calibration-eligible outcomes so infrastructure blocks remain visible without poisoning retrieval-quality rates.
  • Added a public-core capability manifest covering 14 contracts and surfaces shared by every ContextLattice distribution.
  • Added saved graph/recall eval artifacts with explicit baseline, intervention, holdout, latency, cost, failures, and reproduction commands.
  • Patched golang.org/x/crypto to v0.52.0, with compatible x/sync and x/text upgrades.
  • Brought Pi, Droid, OMP, and Mercury integration to profile parity across public, paid, and private distributions.

Measured memory activation

The release ledger records the first bounded project graph activation:

  • Graph edges: 0 -> 3 (2 explicit, 1 inferred).
  • Connected documents: 0 -> 4 of 5.
  • Graph quality score: 20 -> 74.
  • Saved recall holdout: 5/5 cases passed on two runs.
  • Recall@5: 1.0.
  • MRR: 0.9.
  • Numeric exactness: 1.0.
  • Citation coverage: 1.0.
  • Graph lift: 0.0 because direct retrieval already found every expected file.

That last number matters. Graph expansion stays disabled until a holdout proves nonzero lift. ContextLattice records the truth, not the marketing fantasy.

Safety boundaries

  • The CLI remains the primary operator and agent interface.
  • ContextLattice does not auto-install Pi, Droid, OMP, Mercury, or any other third-party harness.
  • Runner quality remains advisory; it does not auto-dispatch work.
  • Pi and Droid process execution remains in repo-local adapters, never in gateway-go or Docker Compose.
  • No Python service is part of the active public application runtime. Compact hook glue and optional development scripts remain separate from the Go/Rust hot path.
  • Release artifacts are sanitized for secrets, local absolute paths, and user memory.

Reproduce

contextlattice_synthesis_pack "what matters before the next decision?" --project contextlattice --pretty
contextlattice_async_inbox_drain --session-id <session-id>
contextlattice_runner_quality --pretty
scripts/agent/audit-capability-parity --pretty
scripts/agent/recall-quality-eval --include-debug --tuning --pretty
Assets 5
Loading