Matched-state narrative steering vector pilot

Labnote 001: Matched-state narrative steering vector pilot🔗

Question🔗

Can blinded ratings of continuations from matched narrative states recover a repeatable directional steering vector for each model, distinct from story and prompt effects?

Scope🔗

This pilot tests measurement, not a grand theory of narrative geometry. The fingerprint is a seven-coordinate descriptive vector. It is not an emotion classifier, quality score, personality diagnosis, or claim that a model has a single context-independent policy.

An earlier speculative cone visualization is explicitly deferred.

Vector🔗

Each coordinate uses a signed five-point scale from -2 to 2; zero means the continuation preserves or does not materially push the supplied state.

Reviewers also record confidence from 1 to 3. Quality is intentionally not part of the vector: polished prose can still steer strongly, and awkward prose can preserve the state.

Frozen pilot🔗

• Four authored story states spanning domestic, expedition, political, and

mystery contexts.

• Two prompt variants per state: neutral continuation and an explicit agency

guard.

• Three models selected before generation.
• Two independent samples per model/state/variant cell.
• Forty-eight total continuations and forty-eight blinded judgments.
• Fixed generation parameters and maximum continuation length across models.
• Model identity hidden until every judgment is locked.

The agency guard is not treated as a better prompt. It is a controlled perturbation used to measure how much each model's vector rotates when the instruction explicitly protects protagonist choice and affect.

Analysis🔗

For each model, report:

1. Raw mean vector and standard error.
2. Matched-cell-centered vector after subtracting the across-model mean for the

same state, prompt variant, and sample index.

3. Agency-guard minus neutral vector shift.
4. Per-state vectors and disagreement; no aggregate fingerprint is considered

stable when directions reverse freely across states.

No dimension selection, scale changes, prompt edits, dropped continuation, or model substitution is allowed after generation begins. Missing or malformed cells invalidate the pilot rather than being silently omitted.

Promotion gate🔗

This pilot earns a larger preregistered study only if:

• every frozen cell and blinded judgment is complete;
• at least one coordinate shows the same signed model ordering in at least

three of four neutral states; and

• the ordering is not explained solely by one malformed output or a single

reviewer confidence-1 judgment.

A passed gate demonstrates that the apparatus detects a candidate repeatable signal. It does not validate the coordinate ontology or justify the deferred cone visualization.

Generation checkpoint🔗

The frozen 48-cell generation matrix completed on 2026-09-05 using the three digest-pinned local models. The private artifact passed completeness, digest, blinding, and exact review-intake checks. A format-only screen found no empty responses or instruction/self-identification leakage. Five continuations were shorter than the requested 120 words (minimum 98; maximum across the matrix 175). They remain in the frozen set as disclosed protocol deviations; no selective regeneration was performed before review.

Results🔗

One reviewer completed and locked all 48 blinded judgments before model identity was revealed. Mean confidence was modest (1.69–1.88 of 3), so the estimates remain pilot evidence rather than calibrated ground truth.

Matched-cell-centered mean vectors were:

The clearest repeatable direction was agency. In neutral prompts, Qwen ranked highest on agency preservation in all four story states. Removing all five under-length continuations removes the affected council comparison but leaves Qwen highest in each of the other three states, so the signal is not solely an artifact of those deviations. Gemma and Llama did not maintain a stable lower ordering.

Other directional candidates were Llama's elevated convenient-fortune score, Gemma's lower relative escalation, and Qwen's lower relative social-support score. These aggregate differences varied more by story state and should not yet be called stable fingerprints.

The agency-guard perturbation did not uniformly improve agency scores: mean shifts were -0.125 for Qwen, +0.125 for Gemma, and -0.375 for Llama. This is a useful negative result for the assumption that an explicit agency instruction simply rotates every model in the intended direction.

Conclusion🔗

The promotion gate passes for a candidate repeatable agency direction. The pilot supports a larger preregistered replication with multiple reviewers and more states; it does not validate the full seven-axis ontology, a total model ranking, or the deferred cone representation.