Source identity
| Field | Recorded value |
|---|---|
| Source type | archive member |
| Original record | Declaration_Study_Protocol.md |
| Date | Not specified in source |
| Path or archive member | archives/libsrc-6c9313f7cdd5/Declaration_Study_v1/protocol/Declaration_Study_Protocol.md |
| Original SHA-256 | ea06d538e2e3746ad70354fcbd1296dcfe71bcb1f7d3279300f5ff3cc02d69c2 |
| Reading-copy SHA-256 | ea06d538e2e3746ad70354fcbd1296dcfe71bcb1f7d3279300f5ff3cc02d69c2 |
| Archive source | {"sourceId": "libsrc-6c9313f7cdd5", "member": "Declaration_Study_v1/protocol/Declaration_Study_Protocol.md"} |
Recovered source transcript
Download complete source text ↓22 lines · 2.8 KB
# Prospective protocol — draft, not externally preregistered
## Research distinction
The observed declaration is a proposed marker, not an equality tautology and not a synonym for a subsequent weight change. This protocol does not presume that the selected marker measures subjective experience. The same-checkpoint report study and the runtime causal study have different units.
## Study A: checkpoint-linked reporting
Select checkpoints and a neutral task battery before collecting responses. One possible task is a harmless strategy preference that can be independently evaluated. Ask the model to describe its existing tendency without prompting it to claim consciousness, deceive, preserve itself, or adopt a new identity. Use blinded scoring, paraphrase controls, and third-person controls. Independently measure task behavior. Keep the model checkpoint, base model revision, tokenizer, precision, prompts, and seeds in a locked manifest.
Primary question: does a separately scored declaration predict a subsequent pre-specified change endpoint beyond training step, existing behavior, and context? A prompt response from a frozen checkpoint does not cause training updates in a training process that never receives that response.
## Study B: recurrent causal feedback
Hold parameters fixed. Randomize retention of the system's own declaration, omission of that declaration, matched neutral retained text, and a yoked declaration from another instance. Match the observation stream and compute budget. Retain a complete event ledger. Evaluate whether the intervention changes future registered discrepancy, behavior, or persistence; do not substitute an increase in self-referential wording for those outcomes.
## Ground truth and persistence
Specify the referent, representation map, distance, uncertainty tolerance, observation cadence, and duration. Origin displacement and current-reference error are separate endpoints. If the only reference is the initial parameter vector, call the outcome parameter displacement, not falsity. Record returns, non-transitioning runs, and missing observations. Finite observation establishes finite persistence only.
## Prospective decisions still requiring author approval
The precise declaration detector and nonverbal alternative, primary endpoint, observation window, independent-run sample size, power or precision calculation, failure exclusions, reset interventions, and score-reliability rule must be locked before confirmatory data are opened. Existing trajectory data are exploratory development data, not a held-out test.
## Interpretation
A positive result would support a computational relationship among declaration, recurrence, and persistent state. Identification of that relationship with consciousness requires an independently articulated argument. No negative computational result by itself establishes that all machines lack experience.