ArcaQ V1.2: Knowledge-to-Action Contracts — When Knowledge Changes, Execution Follows

"What do we know?" is the wrong final question

Enterprise AI has invested a decade in knowing: search, RAG, knowledge graphs, semantic layers. All of it answers one question — what do we know about X? But the moment an agent approves a supplier, suspends a reflex, or blocks a transaction, a different question decides whether the organization is safe: what is this knowledge allowed to authorize?

Those are not the same question. A supplier assessment can be 95 % credible and still be legally insufficient to automate an approval — because the authority is expired, a jurisdiction envelope does not apply, or a second legitimate source disagrees. Confidence is a property of information. Authority is a property of institutions. ArcaQ V1.2 separates them, permanently, in code.

Knowledge-to-Action Contracts: five classes, one rule

A Knowledge-to-Action Contract (KAC) is a governed, hash-sealed, time-bounded artefact issued for a category of action. It references the governed state of each knowledge it relies on — it never copies the knowledge itself — and assigns exactly one of five eligibility classes:

• EXECUTE — the foundation is sufficient to authorize the action.
• EXECUTE_WITH_CONSTRAINTS — authorized, under explicit conditions.
• RECOMMEND_ONLY — usable for reasoning, insufficient to act.
• HUMAN_REVIEW_REQUIRED — an arbitration or additional authority is needed.
• BLOCKED — cannot ground this action at all.

The rule is deliberately strict: the weakest knowledge decides, weighted by materiality and role — never an average. If a decision rests on four authorized facts and one contested assumption rated D5 (critical materiality), the answer is HUMAN_REVIEW_REQUIRED. Not 82 % confidence. Not "probably fine". The materiality of what is doubted — not the average of what is believed — governs the action.

And because authority decays, a KAC carries validFrom, validUntil, revocation conditions and a sealed policy version. Authorized once never means authorized forever.

The Epistemic Circuit Breaker

The part we are proudest of is what happens next. Knowledge does not stand still — outcomes arrive, sources disagree, regulations shift. In V1.2, every significant epistemic transition (a knowledge becomes contested, restricted, superseded or reaffirmed) automatically triggers a re-evaluation of every active contract that relies on it:

1. A governed disagreement transitions to contested in the Disagreement Ledger.
2. ArcaQ detects the affected contracts and re-runs deterministic eligibility.
3. A contract moves from EXECUTE to HUMAN_REVIEW_REQUIRED — and a KnowledgeActionEligibilityChanged event is persisted with reason and provenance.
4. Downstream enforcement (ArcaX) consumes the event stream: automation is blocked, reflexes suspended, tool calls denied, or a human is routed in — without re-interpreting anything.

ArcaQ decides whether the foundation of an action is still institutionally justifiable. ArcaX enforces what may actually execute. That separation — understanding on one side, enforcement on the other — is the moat.

Standing on governed belief revision (CAP-08)

None of this works without a trustworthy account of how knowledge evolves. ArcaQ's governed belief revision pipeline closes the loop that most AI systems leave open:

• Outcome → Qualification → Attribution: an outcome is admitted only after deterministic policy qualification; attribution to a source requires independence checks — no single bad outcome may smear a whole provenance.
• Revision candidates, not rewrites: propagation produces quarantined candidates with compensation plans, sealed by policy hash. The probabilistic path never becomes the path of truth.
• Institutional doubt as a first-class state: contested knowledge is preserved, not overwritten — with applicability envelopes, epistemic debt scoring, and counterfactual memory that keeps rejected alternatives reopenable.
• Impact simulation before anything changes: a deterministic Epistemic Blast Radius score (criticality × dependency × authority × usage × temporal reach) shows reviewers what a revision would touch — nothing is modified during simulation.

From conversational analytics to governed agents

V1.2 also closes the day-to-day loop between analysis and automation:

• Query Builder → chat: any analysis (SQL, result set, full investigation trajectory) pushes into a governed chat conversation with one click — the conversation continues there and feeds the organizational second brain.
• Conversation → agent blueprint: any chat conversation converts into an editable agent blueprint — name, description, system prompt, tools — generated from the entire thread, not just the first question.
• Explain-and-build everywhere: every result carries a one-line reading with a "Details" funnel, so capitalizing knowledge into reusable agents takes one click, not one data-science ticket.

The result: the trajectory of an analysis — the questions, the doubts, the explanations — becomes a durable, governable asset instead of disappearing into a chat window.

Availability

ArcaQ Enterprise V1.2 is deployed in production and available for new on-premise installations. All processing stays inside customer infrastructure; air-gap deployments remain supported without GPU requirements, and the full UI is localized across 17 languages.

Explore the concepts: Explainability, Decision Intelligence, Operational Ontology Contracts, and the Enterprise Second Brain.