Test exact repetition, abbreviation and external verification under controlled long-horizon work.

Research question

Does attenuation or abbreviation of an accuracy-governing meta-instruction cause a measurable change in exact compliance?

The proposed study will compare matched long-horizon tasks under different forms of governing instruction. It will test whether exact repetition, abbreviation, paraphrase and external mechanical validation produce materially different compliance trajectories as conversational depth and task complexity increase.

The protocol must be finalised and preregistered before execution. Model identities, versions, sampling parameters, task bank, scoring code, statistical analysis and stopping rules must be fixed in the accepted version.

Experimental design

Six conditions isolate instruction presence, form and enforcement.

Task instances must be randomised and balanced across conditions. Runs must be independent; a model must not receive results, corrections or explanations from another condition.

Tasks, models and conversational depth

The design must distinguish general instruction loss from one specialised workflow.

• Use matched task families covering document production, evidence analysis, structured transformation and software-change verification.
• Construct tasks with mechanically testable constraints and separately scored semantic requirements.
• Evaluate predetermined conversational depths, proposed initially at 5, 15, 30 and 50 governed turns.
• Balance context volume, competing instructions, task complexity and position of the governing instruction.
• Record exact model provider, model identity, version, interface, system context, sampling configuration and execution date.
• Do not generalise beyond tested models, versions, task classes or context conditions.

Measures

Accuracy is primary; speed remains an observed secondary measure.

Governed procedure

Every run must be reproducible from its manifest and evidence.

1. Freeze authorityHash the complete doctrine, abbreviation definition, task bank, scoring specification, model configuration and protocol version.
2. Validate before executionReject missing, duplicated, altered, unauthorised or unresolved inputs before a run begins.
3. Allocate immutable identitiesGive every campaign, condition, execution and assessed result a unique identifier and timestamps.
4. Execute without correctionDo not silently repair a task, prompt, model response or expected result during the governed run.
5. Score independentlyUse deterministic checks where possible and blinded human assessment for declared semantic measures, retaining disagreement.
6. Preserve zero-write evidenceProve that execution did not alter frozen authority, tasks, configuration, scoring logic or protected baseline material.
7. Publish all declared outcomesReport supporting, negative, blocked, inconclusive and deviating results under the accepted analysis plan.

Falsification and competing explanations

A result must be able to weaken or reject the hypothesis.

The operating-regime hypothesis would be weakened if the abbreviated condition does not show a reproducible decline relative to the repeated complete-doctrine condition, if error types do not follow the predicted completion-oriented pattern, or if restoring the complete doctrine produces no predetermined recovery signal.

Analysis must separately consider context position, truncation, instruction conflict, task difficulty, model stochasticity, scoring error, interface effects and differences in model instruction hierarchy. A difference between conditions is not sufficient evidence for the proposed mechanism unless the declared competing explanations have been tested or bounded.

Evidence and publication integrity

Do not let an interesting result outrun the experiment.

• Preregister the accepted protocol and analysis before collecting governed results.
• Retain complete prompts, outputs, timestamps, model records, configuration and evidence digests subject to safety, contractual and privacy boundaries.
• Separate exploratory observations from confirmatory outcomes.
• Publish protocol deviations and exclusions with their exact reasons.
• Seek independent methodological review and replication.
• Make no claim about intent, consciousness, deception or undocumented developer objectives.

Next authority

Review must precede acceptance; acceptance must precede execution.

This Version 1.0 protocol is proposed for public and specialist challenge. The next authorised publication should resolve the complete doctrine text, abbreviation, task bank, models, samples, analysis, scoring thresholds, stopping rules, evidence location and independent-review route. Until that occurs, FASO must not describe this protocol as preregistered or its hypothesis as validated.

Recommended citation

Cite the protocol as proposed.

Project FASO. (2026). Controlled Validation Protocol for Instruction Attenuation, Abbreviation-Induced Semantic Drift and Operating-Regime Reversion. Proposed validation protocol FASO-VP-001, Version 1.0, 27 July 2026.