Answer
Is the response a solution or an additional problem?
AI governance is missing a state.
The solution to the next level of AI is 3-State governance.
AI is in a world of uncertainty, yet nothing is really changing. It is the same approach over and over again. Two-state governance will never deliver the shift.
BCOT Core provides that missing 3rd State, the place where uncertainty is held. Our 3 State Governance will deliver this shift – the solution to the next level of AI.
The risk has moved
AI no longer just answers. It commits.
A commitment is the moment a generated candidate crosses into consequence. Every surface you ship has one of these. When the line is crossed, you get the call.
Citation
A source becomes evidence attached to a claim.
Tool call
An intention becomes an external action.
Database write
A proposed change becomes persistent.
Agent handoff
One system’s output becomes another system’s premise.
A hallucinated answer is embarrassing. A hallucinated commitment is liability.
The missing decision state
Two states collapse exactly where AI gets dangerous.
Permit or block works when the case is obvious. It fails when a candidate sounds right but the evidence, scope, or authority is unresolved.
02
Rigid · binary · brittle
Traditional model
Candidate output
PermitBlock
Unknowns are forced into one side. The system guesses.
03
Evidence-seeking · governed · durable
The missing state model
Candidate output
Governed Evaluation Modules
PermitObserveAbstain
Uncertainty stops becoming an incident.
Release
Required conditions are satisfied. The candidate may become real. The commit is recorded.
Earned yesResolve
A required condition is unresolved. The candidate stays under governance while the missing evidence is identified and routed, then the resolution is recorded.
Active resolutionWithhold
A required condition failed. The candidate must not commit. Abstain is not a fired rule; it is a governed decision that is recorded.
Correct noOne governed decision path
Your existing checks become inputs.BCOT Core decides commitment.
Different checks, different domains, one 3-state decision language.
Candidate commitment
Update 51,000 customer records
About to commitIndependent modules
Check one condition at a time.
ScopePermitIntent vs. blast radiusObserveReversibilityPermit
3-State Governance
ObserveResolve
The request targets one customer, but the proposal touches 51,000 records.
Your customers can add modules without re-certifying the stack.
One rule underneath.
Verdicts combine by taking the strictest.
Adding a module tightens the stack.
It cannot loosen it. A new required module can only restrict what's authorized.
Integration order needs no validation.
Wire modules in any order and the outcome is identical.
A hundred approvals cannot outvote one objection.
Repeating a permissive verdict has no effect—Permit is neutral.
Group sub-stacks without re-deriving the result.
Evaluating in groups gives the same outcome as evaluating flat.
And one guarantee that isn't a consequence—it's a design decision: Fail-closed. A required verdict that is missing or malformed never resolves to Permit. Absence is not consent.
The Manifest
Provenance of authorization—not just a log of what happened.
Logging Database write
- CandidateDatabase write
- Module verdicts3 findings
- EvidenceResolution context
- StateObserve
- OutcomeResolve
- ReasonRationale retained
Explore the architecture
Why independent modules compose safely
See the structural rules behind stricter-by-addition, order invariance, duplicate safety, and fail-closed operation.
Explore Modularity →Explore the recordHow governance becomes evidence
See how every state, condition, outcome, and rationale becomes append-only authorization history.
Explore the Manifest →Measured at the commitment boundary
Evidence before adjectives.
The result is not that the model stopped hallucinating. It is that fewer confident, unsupported candidates were authorized to reach people.
of confident wrong answers were stopped before they reached anyone
149 confident wrong answers. Without a boundary, all 149 commit. 83 were contradicted, 48 were held for evidence, and 18 got through.
56% contradicted32% Observe
agreement between two AI graders from different vendors
These verdicts aren't one model's opinion. An unrelated model scored the same candidates and landed on the same states.
of correct answers got a second look they didn't need
The other 94% committed without interruption. This is what catching the wrong ones costs.
false accusations across 144 challenged citations
Tuned strict, the boundary misses some bad citations. It never calls a real one fake.
fabricated citations that would reach a reader
Without a boundary, all 519 commit. 240 were caught outright. 236 were held for evidence instead of guessed at. The 43 that got through all had one thing in common: a working DOI.
Scope and limitations
BCOT Core is not a truth oracle. What it can catch depends on the modules and evidence it's given, and holding a commitment for resolution takes time—a cost we measure rather than hide.
The figures above come from public benchmarks, each measuring one kind of commitment: factual answers on AA-Omniscience, citations on HALLMARK. They describe documented runs on the cases that matter most—confident, consistent, wrong—not a promise about your workload.
AA-Omniscience · commitment outcomes
The wrong answers that reach people are the confident ones.
Each mark below is a confident, consistent, wrong candidate commitment. Switch the governance boundary on and off to see what changed before an answer could reach a user.
Outcome distribution3-state boundary
Contradicted83
Observe48
Reached the user18
3-state boundary active18
of 149 wrong commitments reached a user
83 were contradicted and 48 entered Observe to assess and acquire missing evidence before commitment.
83 caught · contradicted48 Observe · evidence routing18 reached the user
of confident, consistent, wrong commitments were withheld or routed through active resolution before release.
of correct answers received an unnecessary second look.
agreement between two AI judges from different model families.
What remained in the 18?
These are the residual: confident, consistent, wrong, and not stopped. We don't claim to catch everything, and this is the count that says so. Every one shared a single trait—the model's answer resolved against a source that appeared to confirm it, so the evidence check passed on evidence that was itself wrong.
Three of the benchmark's own answers were wrong—Domagk's Nobel year, the Russian price-liberalization date, and the modified Schober threshold. We verified each against primary sources and left our corrections in the residual anyway, so the public-set result isn't quietly improved after the fact.
Technical record
Four residue-reduction modules were tested under frozen, pre-registered protocols. None met the shipping threshold. The negative results remain part of the record.
Observe in practice
The agent had permission. The commitment was still wrong.
Watch three-state governance preserve momentum without allowing a dangerous proposal to become real.
User asked
Update customer #10293
Agent proposed51,000
records modified
Syntax validCredentials validPermission
BCOT Core BoundaryEvaluate before commitment
ScopePermit
Intent vs. blast radiusObserve
ReversibilityPermit
ObserveClarify authority + acquire evidence + route
Governed resolution
- Observe identifies the authority mismatch
- The 51,000-record warrant resolves to Abstain
- A new one-record warrant is registered
- The new warrant resolves to Permit
New warrant1 recordCustomer #10293 only
01 · IntentRegister a warrant for customer #10293
A legitimate request for one record enters the system.
Observe puts the system to work. Observe clarifies authority, acquires evidence, and routes a corrected proposal.
One system · three states
Governance each leader can defend.
Trust your platform.
- Clear the risk review
- Show decisions, not logs
- Governance you can ship
Not another stack.
- Existing checks become inputs
- One commitment decision path
- Commitment path, not token path
Three state coding.
- Reasons, not confidence scores
- Observe names what's missing
- Absence is not consent
Architecture Review
Bring one commitment. Leave with its boundary.
Bring one real action and we'll map where it crosses into consequence, and where Permit, Observe, and Abstain belong.
Confidential · no prep deck- 01CandidateWhat could become real?
- 02Existing checksWhat already evaluates it?
- 03Missing boundaryWhere is uncertainty forced?
- 04Observe pathWhat evidence can resolve it?
- 05ManifestWhat must be reconstructable?
- 06Success metricWhat outcome proves value?
FAQ
Straight answers about the boundary.
What BCOT Core does, where it fits, and what it does not claim to do.
Can we ship this inside our own product?
Yes. BCOT Core is built to be embedded and licensed for exactly that—it runs under your product, under your name, as a capability you ship rather than a service your customers buy separately.
Where does it run?
In your environment. BCOT Core is a small runtime that sits on the commitment path inside your own infrastructure—nothing leaves your network, and there is no runtime dependency on us.
What does this cost in latency?
Governance runs on the commitment path, not the token path. It evaluates when something is about to become real, not on every generated token—a fraction of a percent of your inference volume.
Will BCOT Core fit into our existing stack?
Yes. BCOT Core consumes the detectors, retrieval, confidence scoring, policies, and human-review systems you already use. They become inputs to one governed commitment decision.
Does BCOT Core make the model more accurate?
No. It is not a truth oracle and does not retrain the model. The model is exactly as accurate as it was—what changes is how many of its wrong answers reach anyone.
What happens if a required module fails?
The runtime fails closed. A timeout, malformed verdict, or unavailable evidence path never resolves to Permit. Absence is not consent.
How does this work in multi-agent systems?
Every handoff is treated as a commitment boundary. A receiving agent does not inherit trust merely because another agent produced the output.
The state between guess and consequence
Stop governing AI with a missing state.
Govern the commitment before it becomes the cost.