ContextLattice v3.12.0 - Proof, Not Vibes

ContextLattice v3.12.0 - Proof, Not Vibes

v3.12 gives memory a sense of time and synthesis a burden of proof. An agent can
now ask what is true, what used to be true, what replaced it, what contests it,
and exactly which evidence allows the claim to exist.

What changed

• Added a bounded, persisted Temporal Claim Graph with validity intervals,
  provenance, verification, contradiction, supersession, causality, branch,
  commit, agent, and session identity.
• Added deterministic retrieval_plan.v1 advice with task-phase evidence
  obligations, source/query plans, token allocation, graph expansion advice,
  observed reliability and p95 latency, and marginal-value stop conditions.
• Added Proof-Carrying Synthesis v2. Every emitted finding carries support,
  opposition, temporal state, causal context, missing proof, and decomposed
  confidence. Findings without an evidence identity are excluded and counted.
• Added primary CLI commands:
  contextlattice_claim_write, contextlattice_claim_query,
  contextlattice_retrieval_plan, and contextlattice_synthesis_pack_v2.
• Added nine Go-native HTTP/tool/telemetry routes and four bounded contracts.
• Expanded strict native-ownership, context-boundary, capability, installer,
  discovery, audit, and public-core parity surfaces for the cognition core.
• Preserved every v1 route and contract for gradual adoption.

Measured behavior

The public eval ledger records deterministic, model-free holdouts and benchmarks:

• Temporal claim persistence/supersession: passed across process-style reload.
• Claim, planner, and synthesis contract holdouts: passed on three repeated runs.
• Model calls: 0.
• External API/network calls: 0.
• 1,000-claim bounded query median: 931454 ns/op.
• 1,000-claim allocation pressure: 2297993 -> 33432 B/op.
• 1,000-claim allocations: 6933 -> 7 allocs/op.
• Planner median: 10719 ns/op.
• Thirty-two proof claims median: 88211 ns/op.

These are component benchmarks on the release development host, not universal
latency promises. Reproduce them with the commands in
docs/evals/v3.12-cognition-proof-core.json.

Safety boundaries

• Retrieval planning is advisor / shadow_only. It cannot silently mutate the
  live source policy.
• Synthesis is deterministic and reports llm_used=false.
• Lexical claim linking does not pretend to be semantic equivalence.
• Unverified claims cannot silently raise confidence.
• Contradictions remain visible and unresolved; the public core does not invent
  a winner.
• The active application path remains Go/Rust. No Python service or gateway-owned
  subprocess was added.
• The claim ledger is bounded and uses the configured ContextLattice data volume.

Try it

contextlattice_retrieval_plan "prove release readiness" --project contextlattice --pretty
contextlattice_claim_query "current release" --project contextlattice --include-superseded --pretty
contextlattice_synthesis_pack_v2 "prove release readiness" --project contextlattice --pretty
contextlattice_context_boundary --pretty
contextlattice_strict_runtime_native_ownership --pretty