A working record,
with explicit limits.

Evaluation practice

Protocols separate the agent’s visible context from evaluator records. Fixed resource ceilings, append-only evidence, execution lineage, and complete outcome counts help distinguish apparatus behavior from scientific effects. Missing, invalid, and failed attempts stay visible; equal resource maxima do not imply equal realized computation.

In forecast protocols that register it, missing responses incur Brier loss 1 and accuracy 0. The calibration published here has no forecast probes or memory score: its success criterion is a validated investigation action within the fixed delivery envelope.

Scripted checks verify engineering. Live model samples, retrospective audits, targeted diagnostics, and prospective comparisons have different evidential status. Seeds are recorded without promising deterministic inference.

Limits on interpretation

Small trials, delivery constraints, controller scaffolding, tools, extra context, prior training, and resource imbalance can explain apparent effects. Source fidelity is distinct from source truth. Construction is distinct from learned utility. No aggregate metric is a consciousness probability.

Private original evidence is retained locally. Public downloads are explicitly selected and reviewed; derived tables are labeled as such. Where originals cannot be safely redistributed, reproducibility limits are stated.

Publication policy

Every completed experiment or substantive milestone requires a dated public-safe report, including failed, invalidated, negative and inconclusive outcomes. Publication is a post-experiment activity and never changes a running evaluation or consumes its budget.

Execution status and publication status are recorded separately. Closeout requires either a verified public URL and revision or an explicit publication blocker with a retained bundle for retry. Historical backfill is reviewed before release; no favorable-result filter is used.

Local registration files or commits do not establish public preregistration. Experiment dates and first publication dates are separate. Corrections receive dated revision notes, stable URLs and preserved substantive versions. Necessary privacy or security removals are documented without restoring sensitive material.

Only reviewed, redistributable outputs are public. Raw conversations, reasoning transcripts, credentials, private source, personal material, unreviewed logs and internal infrastructure are excluded by default. Questionable material is held separately and the report explains the limitation.

The public file manifest records paths, byte sizes and SHA-256 hashes, excluding the manifest’s own recursive hash. Artifact manifests identify original, redacted, or derived content. No DOI, external repository, analytics, login or model backend is required.