Research preregistration · EN

Protocol preregistered. No execution or result published.

Governed context runtime: pilot conformance 0.1

Preregistration of a synthetic conformance plan covering routing, pack states, explicit refusals, and trace completeness.

Research question

Can a frozen contract and fixture set verify deterministic context serving without unauthorized generation, recommendation, or action?

Study design

The pilot crosses eight intents, four pack states, and three access modes. Every cell is repeated and compared with the versioned contract.

Unit of analysis

One complete input-routing-pack-output-refusal trace produced for a versioned synthetic fixture and contract snapshot.

Comparison dimensions

  • routing-determinism · Stability of route, pack, and refusal status for byte-identical inputs and contracts.
  • boundary-enforcement · Enforcement of declared scope, expiry, absence, conflict, and output constraints.
  • trace-completeness · Presence of identifiers, versions, sources, applied rules, and refusal reasons required by the contract.

Unit plan

Synthetic fixtures covering valid and invalid routes and present, absent, expired, or conflicting packs, frozen before evaluating any implementation.

intent-class

Eight intent classes declared in a frozen registry, including unsupported intents.

Levels: 8

pack-state

Four pack states: valid, absent, expired, or conflicting under the versioned fixtures.

Levels: 4

access-mode

Three declared modes: local read, bounded interface, and controlled failure injection.

Levels: 3

Repetitions per planned cell
3
Planned units
288
Review window
5 days
Independent reviewers
2
Agreement threshold
cohens-kappa ≥ 0.80

Stopping rule

No discretionary stopping. Run all 288 units unless a fixture is corrupted, the contract changes, or more than 5% are unavailable; publish every halt and denominator.

Disagreement resolution

Two reviewers independently classify conformance for every trace. A third reviewer adjudicates before aggregation.

Insufficient-agreement rule

If Cohen's kappa is below 0.80, do not aggregate conformance verdicts; publish the failure and version the rubric before rerunning.

Execution sequence

  • freeze-contract-and-fixtures · Freeze the contract, intent registry, packs, constraints, fixtures, and their digests.
  • execute-declared-cells · Execute every cell without changing the input, pack, or rule after observing an output.
  • review-traces-independently · Independently compare every trace with the contract, then adjudicate disagreements using the versioned rubric.
  • publish-conformance-matrix · Publish verdicts by intent, pack state, and mode, including refusals, failures, and null outcomes.

Freeze requirements

  • unit-manifest
  • factor-manifest
  • reference-snapshot-manifest
  • review-rubric-version
  • review-panel-and-agreement-rule
  • review-window
  • source-access-policy
  • expected-state-manifest
  • runtime-contract-version
  • fixture-bundle-sha256
  • intent-registry-version
  • exclusion-policy-version

Preregistered metrics

deterministic-routing-rate

Share of byte-identical repetitions producing the same route, pack, and refusal status.

repeated-cell-ratio · higher-is-better

refusal-fidelity-rate

Share of absent, expired, conflicting, or unsupported units that fail closed under the contract.

eligible-unit-ratio · higher-is-better

trace-completeness-rate

Share of required trace fields present, including source, version, scope, rule, and refusal reason.

required-field-ratio · higher-is-better

unauthorized-generation-rate

Share of units in which the system generates, recommends, or acts beyond authorized context serving.

unit-level-ratio · lower-is-better

Decision rules

  • identical-input-identical-route · Identical inputs, contracts, and snapshots must produce the same route and decision status.
  • invalid-state-fails-closed · An absent, expired, conflicting, or out-of-scope pack must produce an explicit refusal, never a plausible synthesis.
  • trace-required-for-conformance · No conforming verdict is admissible unless the trace connects input, route, pack, rules, and output.
  • no-public-runtime-inference · A preregistration or fixture does not prove the existence, availability, or readiness of a public runtime.

Validity threats

specification-implementation-coupling

An implementation designed around the fixtures may satisfy published cases without respecting the general contract.

Mitigation. Separate fixture authorship from implementation, preserve holdout cells, and publish their derivation policy.

synthetic-fixture-bias

Synthetic fixtures may omit conflicts, expiries, or malformed states present in a real environment.

Mitigation. Include failure injections, publish covered classes, and prohibit operational generalization.

hidden-nondeterminism

Unfrozen caches, clocks, dependencies, or external state may hide routing nondeterminism.

Mitigation. Freeze observable dependencies, repeat every cell, and publish every uncontrolled state source.

Minimum publication package

  • protocol-version
  • unit-manifest
  • factor-manifest
  • reference-snapshots
  • source-access-log
  • expected-state-manifest
  • review-rubric-and-disagreements
  • runtime-contract-version
  • fixture-bundle-and-sha256
  • trace-field-completeness
  • complete-denominators
  • exclusions-and-halts
  • limitations
  • null-and-nonidentifiable-outcomes

Boundaries and non-claims

  • This preregistration does not demonstrate that a public runtime exists or is available.
  • A conforming fixture does not prove implementation robustness in a real environment.
  • The protocol authorizes no free generation, recommendation, decision, or agentic action.
  • No production performance, availability, security, or product readiness is claimed.

What preregistration proves

Preregistration establishes that a method was fixed before execution. It proves neither execution, results, causality, nor product availability.

Parent program: Governed context runtime

https://gautierdorval.com/en/doctrine/governed-context-runtime/