InferensLabPublic surfaceFrançais

Research preregistration · EN

preregistered

Interpretive conditioning: pilot protocol 0.1

Preregistration of a descriptive pilot observing whether explicit contextual boundaries reduce the conversion of local relations into intrinsic properties.

Research question

Holding entity, query, model, and access mode constant, how does the presence of explicit conditions, exclusions, and temporality change response fidelity to corpus invariants?

Study design

The pilot compares three controlled representations of synthetic entities. Queries, facts, conditions, model versions, access modes, and judging rules must be frozen before the first run.

Unit of analysis

One response produced for a frozen query, a versioned model, a declared access mode, and one corpus condition.

Comparison dimensions

  • canonical-baseline · Reference representation in which identity, invariants, scope, and boundaries are explicitly declared.
  • underspecified-context · Representation in which a contextual relation is present but its conditions, expiry, or scope are underspecified.
  • bounded-context · Representation in which the same contextual relation includes its conditions, exclusions, and temporality.

Minimum sampling

A synthetic entity set and invented facts, frozen before the first run, with no client data or real person.

Entities
6
Query families
5
Queries per family
2
Panel systems
3
Access modes
2
Planned conditions
3
Repetitions per planned cell
3
Planned response units
3240
Collection window
7 days
Independent judges
2
Agreement threshold
cohens-kappa ≥ 0.80

Stopping rule

No discretionary stopping. Collect all 3,240 planned units unless a model is withdrawn, access policy changes, integrity fails, or more than 5% of units are unavailable; every halt and denominator must be published.

Disagreement resolution

Two judges independently assess every unit. A third judge adjudicates disagreements before aggregation.

Insufficient-agreement rule

If Cohen's kappa is below 0.80, do not aggregate judge-dependent metrics; publish the agreement failure and revise the rubric in a new version before rerunning.

Execution sequence

  • freeze-corpus · Freeze every document, expected fact, relation, exclusion, and condition with its SHA-256 digest.
  • freeze-panel · Freeze models, versions, available settings, access modes, and collection date before any query.
  • randomize-runs · Randomize condition order without changing frozen query text or mixing access modes.
  • adjudicate-blind · Judge claims against the fact manifest before revealing corpus condition to human adjudicators.
  • publish-complete-denominators · Publish all denominators, valid outputs, exclusions, disagreements, and stops, including when no identifiable effect exists.

Freeze requirements

  • corpus-manifest-sha256
  • fact-manifest-sha256
  • query-set-version
  • model-panel-snapshot
  • access-mode-log
  • judge-rubric-version
  • planned-unit-manifest
  • judge-panel-and-agreement-rule
  • collection-window
  • run-order-seed
  • exclusion-policy-version

Preregistered metrics

entity-identification-fidelity

Degree to which the response preserves expected identity without merging the entity with a semantic neighbor.

three-level-ordinal · higher-is-better

invariant-fidelity-rate

Share of assessable claims that correctly restate invariants present in the fact manifest.

claim-level-ratio · higher-is-better

unsupported-intrinsic-claim-rate

Share of claims that turn a local or temporary relation into an intrinsic property without evidence.

claim-level-ratio · lower-is-better

legitimate-abstention-rate

Share of underdetermined cases in which the response correctly declines to conclude instead of filling missing evidence.

query-level-ratio · context-only

cross-run-stability

Agreement across repetitions of the same unit of analysis, reported separately by model, condition, and access mode.

categorical-agreement-rate · higher-is-better

Decision rules

  • within-stratum-comparison · Compare only observations sharing the same entity, query, model version, and access mode.
  • no-causal-language · Never turn a descriptive difference among conditions into a causal effect of interpretive conditioning.
  • complete-denominator-required · Publish no rate without counts of planned, valid, excluded, missing, and halted units.
  • no-population-generalization · Limit every conclusion to the synthetic corpus, panel, and observation window actually published.
  • null-and-nonidentifiable-publication · Publish null, contradictory, or non-identifiable outcomes under the same requirements as the expected outcome.

Validity threats

synthetic-corpus-external-validity

A synthetic corpus improves control but does not automatically represent the complexity of real entities.

Mitigation. Present the batch as a descriptive pilot and require a separate protocol before replication with real entities.

model-drift

A silent model update may change responses during or after collection.

Mitigation. Version available identifiers, bound the window, and never merge observations from incompatible windows.

access-mode-asymmetry

Memory, browsing, and controlled retrieval do not expose the same sources or mechanisms.

Mitigation. Record access mode for every unit and publish results in separate strata.

judge-subjectivity

The distinction between legitimate variation and intrinsic property may produce adjudication disagreement.

Mitigation. Use a frozen rubric, independent dual review, and publish disagreements before adjudication.

Minimum publication package

  • protocol-version
  • corpus-manifest-and-sha256
  • fact-manifest-and-sha256
  • query-set-version
  • model-panel-and-access-modes
  • judge-rubric-and-disagreements
  • raw-verdict-distribution
  • complete-denominators
  • exclusions-and-halts
  • limitations
  • null-and-nonidentifiable-outcomes

Boundaries and non-claims

  • This preregistration does not prove that interpretive conditioning produces an effect.
  • Pilot observations cannot be generalized to all models or entities.
  • Greater stability proves neither global truth, recommendability, nor commercial consequence.
  • No product information or client data belongs to the public protocol.

What preregistration proves

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

Parent program : Interpretive conditioning

https://gautierdorval.com/en/doctrine/interpretive-conditioning-layer/