ContextLattice v3.21.0

Retrieval Receipts

ContextLattice can now show not only what an agent received, but why every
piece of evidence was allowed to matter.

The CLI remains the primary interface. HTTP remains the integration fallback.

What changed

• Every Context Pack carries one canonical memory_trust_assessment.v1 and one
  retrieval_decision_trace.v1, with compact references instead of repeated
  payloads.
• Prompt overrides, secret exfiltration, encoded instructions, destructive
  claims, self-awarded trust, and coordinated risky paraphrases fail closed.
• Legitimate bounded runbooks remain usable; quarantined material receives zero
  ranking and prompt influence.
• Same-snapshot ablation reports exact result, rank, and citation deltas without
  inventing utility.
• Independently verified outcomes can produce immutable, review-only regression
  proposals and advisory evidence reputation.
• Cross-project causal explanations require typed current evidence and
  citations; lexical or associative similarity alone always abstains.
• The Go-native CLI exposes contextlattice retrieval-governance for entitled
  receipt, bridge, counterfactual, reputation, regression, and defense policy.

Product boundary

The public core includes local receipts, trust isolation, quarantine, ablation,
regression proposals, advisory reputation, and causal explanations. Team adds
receipt and causal-bridge governance. Operator and Enterprise add continuous
counterfactual, reputation, regression, and defense operations.

Paid governance controls policy, bounded retention, scheduling, and incident
review only. It does not execute retrieval, mutate ordinary memory, delete
immutable evidence, or disable public defenses.

Rollback

Set optional ranking influence to zero and deactivate paid T4 governance
policies. Keep the last-known-good public trust rules fail closed. Security
defenses are not a rollback target and cannot be disabled through an
entitlement.