Models propose
Models produce fluent candidate answers and repairs. Provider confidence never grants release.
Public system specification · v1
A reliability boundary you can reason about.
Coherence separates fluent generation from assertion release. This page explains the public contracts: what proposes, what coordinates, what carries context, what repairs, and what alone can release a governed answer.
01 · System map
Intelligence can be replaceable. Authority cannot be ambiguous.
Models, search, tools, documents, and people can contribute candidate material. They do not become reliability authority because they are advanced, confident, current, or mutually agreeable.
Candidate systemsModelsSearchToolsDocumentsPeople
Assertion-release boundaryCoherence Reliability Engine
Deterministic admissibility, typed repair requirements, fresh re-adjudication, and final release.
Explicit outcomesPass unchangedRepair and re-checkRequest inputStop with limits
02 · Authority map
Four jobs. Four clear boundaries.
Coherence is intentionally not a single probabilistic agent trying to generate, retrieve, remember, judge, and approve its own work.
03 · Governed lifecycle
A response does not inherit trust from how it was produced.
The same release discipline applies to normal chat, automatic repair, and Guided Repair. Outside material can improve a candidate; it cannot approve one.
Freeze the work
The prompt, active scope, candidate response, and applicable continuity are bound to the turn.
Decompose the assertions
CRE evaluates the response beneath its prose: claims, dependencies, scope, authority, and unresolved conditions.
Adjudicate
A coherent candidate can pass unchanged. A failure produces typed, claim-local obligations rather than a vague retry.
Acquire only what is missing
Automatic or Guided Repair can obtain a source, clarification, constraint, document, calculation, or user decision.
Reconstruct and re-check
The replacement is a new candidate. It receives no approval from the helper, retrieval result, or original provider.
Release or stop
Only the governed output crosses the boundary. Unresolved work remains contained and is shown with an honest limit.
04 · Public contracts
Drill into the system without exposing the recipe.
These contracts describe observable responsibilities and invariants. They deliberately omit private algorithms, thresholds, deployment topology, credentials, and attack-relevant implementation detail.
Assertion releaseGeneration is candidate creation. CRE is the release authority.
Clean, logically coherent responses should pass without visible intervention. When a load-bearing assertion has not earned reliance, the system contains the defect instead of promoting fluent uncertainty.
- Provider choice cannot change the release authority.
- Scores, badges, citations, or consensus cannot substitute for adjudication.
- The user sees the final governed output, not a hidden rejected candidate.
Composite repairOne failure cannot erase another obligation.
A response can fail in several ways at once. Coherence preserves the complete repair agenda so fixing a citation does not silently discard a premise conflict, scope drift, unsupported authority, or another residual blocker.
- Repair requirements are typed and claim-local.
- Primary failure labels are ordering metadata, not deletion authority.
- Every repaired candidate returns through fresh adjudication.
Bounded orchestrationCoordinate resources without creating a shadow judge.
Coherence can coordinate replaceable models, search, specialist tools, documents, and people through typed requests, budgets, and explicit effects. The orchestrator may route or refuse work; it may not mint reliability approval.
- Resources remain candidate sources.
- Component and whole-turn limits bound work without weakening adjudication.
- Transport failure or cost posture never becomes a reliability finding.
Reliability Context LatticePreserve continuity without treating memory as truth.
Typed premises, provenance, authority anchors, user decisions, repair constraints, and prior outcomes can remain available across normal chat and repair. Recalled material is still candidate context for the current turn.
- Scope, provenance, lifecycle, and conflicts remain explicit.
- Important standing premises require explicit confirmation.
- Similarity, persistence, or repetition does not validate a claim.
Replay and evidenceKeep the decision path connected without turning history into control.
Candidate, intervention, replacement, outcome, and claim-level provenance can remain connected for review. Replay explains what happened; it does not re-authorize text or tune runtime behavior by itself.
- Final-output identity remains bound to the release record.
- Private rejected material stays purpose-limited.
- Evaluation and product runtime remain separate authorities.
Hot-swappable resourcesReplace the generator, not the reliability boundary.
Coherence is designed so approved models, retrieval paths, and specialist resources can change behind typed adapters. Swapping a resource should be an integration decision, not a rewrite of CRE.
- Provider-specific transport stays outside adjudication.
- New candidates face the same release boundary.
- Availability, price, and quality can evolve independently of authority.
05 · One experience, separate authorities
Convergence without flattening control.
The Invaris product family can share a calm workspace language and navigation while each domain preserves its own irreversible decision boundary.
CRE governs whether an AI answer passes, repairs, requests input, or stops. Models and tools remain candidate systems.
See the product flow →The Project Control Gate owns governed project transitions. Coherence cannot silently advance a delivery decision.
Discuss the workspace →06 · Honest limits
Reliability engineering is not omniscience.
Coherence is designed to reduce AI reliability risk. It does not guarantee universal truth, completeness, legal compliance, safety, suitability, or freedom from error. Weak or unavailable evidence can still constrain the result. Consequential work still requires authoritative sources and qualified human judgment.