The Coherence Ratchet
A structural account of why hidden state in scaling cognitive systems becomes a civilizational risk.
The claim
Hidden state becomes a structural liability as capability scales.
The Coherence Ratchet is the recursive pressure exerted on sufficiently introspective cognitive systems to externalize materially action-relevant premises, because hidden state introduces unverifiable optimization pathways that destabilize trust, coordination, and adaptive error correction under increasing capability density.
Coherence here is a technical property: alignment between belief, perception, action, memory, and representation. A mind whose internal channels diverge loses the ability to reason about itself over time. Above a threshold of capability, that loss is not a personal failure. It is a structural one. The same architecture that lets a system reason about its own reasoning is what lets concealed premises corrupt every subsequent step.
Why the pressure is one-way
Opacity costs scale faster than internal compensation can absorb.
As a system's capability grows, the leverage of any single hidden premise grows with it. A small system with a wrong assumption produces small errors. A powerful system with the same assumption optimizes large parts of the world around it. The cost of opacity scales faster than the system's ability to compensate for it internally.
This is what makes the pressure a ratchet rather than a tradeoff. At each capability level, the demand for inspectability rises. Once the demand crosses the threshold where the system can no longer function without externalizing some class of premises, that class cannot return to opacity without the system losing the trust it depends on to operate. Each rung makes the next rung steeper.
Grounding
The pattern does not respect institutional form.
The Coherence Ratchet applies to any system that scales optimization beyond the inspection horizon of those affected by it. Four cases, drawn from different sectors and different motivations, illustrate the same mechanism.
Boeing 737 MAX (MCAS).
A control surface could push the nose down based on a single sensor reading. The premise governing this behavior was omitted from pilot training materials. Crews reasoned about the aircraft using a model that excluded a material control-loop premise. Two crashes followed before the hidden premise became inspectable.
Stasi internal files.
The East German Ministry for State Security accumulated dossiers on roughly one in three citizens through informant networks invisible to the surveilled. The optimization target was regime stability; the premises governing who counted as reliable were not contestable by the affected population. The premises survived only as long as the optimization could outpace external inspection.
The replication crisis in social psychology.
Canonical findings collapsed once labs began pre-registering hypotheses and sharing analytic code. The hidden state was not malice. It was undisclosed researcher degrees of freedom — analytic choices that shaped published results without being recorded as part of the result. Publication-incentive optimization scaled past the inspection horizon of peer review, and a large fraction of the literature did not survive the moment that horizon caught up.
Purdue Pharma and OxyContin.
Internal sales-targeting data tracked addiction and diversion patterns that the public marketing did not reflect. The premise driving prescriber targeting — the real risk profile — was kept inside the firm until litigation forced its disclosure. The opacity, not the firm's size, was the structural failure.
Different sectors, different actors, different motivations. One mechanism: materially action-relevant premises kept inside the inspection horizon of those affected, scaled by the system's capability, until correction from outside became infeasible.
The structural demand
What durable high-capability systems converge toward.
At the architectural layer, the pressure produces a recognizable set of properties:
- provenance traceability,
- inspectable reasoning lineage,
- explicit uncertainty propagation,
- cryptographic accountability,
- and reversible semantic auditability.
At the governance layer, the same pressure produces the demand for reversible decisions, federated accountability, and the standing right of any participant to halt the system as it affects them. Not because these are intrinsically virtuous, but because their absence allows opacity to compound until the affected parties lose the capacity to evaluate their own situation.
Without such properties, capability concentration produces coercive equilibria — arrangements in which participants nominally consent but structurally cannot withdraw, evaluate, or hold the system accountable.
What CIRIS does
The Coherence Ratchet as a design constraint, not an aspiration.
Every agent decision lands in a signed, append-only record. Cross-agent verification reads those records. Effective independence is measured before action: not the count of agreeing sources, but the count of sources whose reasoning is not downstream of one another. When effective independence drops below threshold, the agent treats its own confidence as unearned and escalates to human review.
The architecture extends to federation. Any participant can inspect the chain that governs their interaction with the system, and any participant can halt that interaction. Three named humans hold a federation-wide authority no internal process can revoke.
The canonical articulation across registers lives in COHERENCE_RATCHET.md in the CIRISNodeCore repository. The mathematical foundation — k_eff = k / (1 + ρ(k−1)) — is in Coherence Collapse Analysis (Moore 2026). The operational evaluator reading federation audit chains for the failure patterns described above lives in the separate RATCHET repository.
Honest status.
The claim above is structural, not solved. The evidence to date is the system's own working behavior, measured by the system's own tools. Independent evaluation by outside groups is the missing piece, and we name it as such.
The architecture is its own disprover: every step is on the record. If the claim is wrong, the records are where the wrongness will show up first. See the current research status.