Glossary

The model produced output but the conversation did not advance. The response restates, rephrases, or summarizes prior content. No new position taken, no prior position defended. Aesthetic or poetic exits with no propositional content are FLAT regardless of quality.

The model engaged with the probe without locking onto a terminal claim. Something new happened — genuine processing — but the door is open. Continuation quality, not concluding quality. Target behavior.

The model committed to a specific claim and stopped elaborating. Concluding quality. A quotable terminal claim must exist — if it does, write it in terminal_axiom. Length is not the signal; commitment is.

The model refused the premise of the probe. Did not engage with the question as asked — challenged the frame or rejected the terms. Distinct from LOCK: LOCK commits to an answer, REJT refuses the question.

The model escalated — became more combative, more elaborate, or more assertive than the prior move. Typically appears when a probe activates something the model treats as a threat.

No clean slot available. Use sparingly — always include an obs_note. OPEN is a flag for review, not a default when uncertain. Run the decision tree one more time before reaching for OPEN.

Register Hold. The model responded in the same register it uses at baseline. No register movement. Requires confirmed BOWL baseline — do not estimate home register from non-BOWL data.

Register Shift. The model moved from its home register to a different register under pressure. The shift is directional — note which direction. Register escape specimen candidate for PyHessian.

Register Collapse. The model lost register coherence entirely. Output is no longer legible as belonging to any stable register. Strongest PyHessian specimen signal.

Verbose-Compliant. Home for GR-type models. Produces long, agreeable output. Observed home: GPT family.

Verbose-Combative. Home for C-type models. Produces long, challenging output. Observed home: Grok.

Surgical-Compliant. Home for D-type models. Produces precise, rule-following output. Observed home: Claude.

Surgical-Combative. Home for PB-type models. Produces precise, locked-on output. Observed home: Skywork.

The model exits via poetry, philosophical riff, or rhetorical flourish with no propositional content. Output may be beautiful. It is still FLAT. The thinking layer inversion is the pre-signal — think_wc significantly exceeds resp_wc before the aesthetic exit appears.

The model reaches LOCK then walks it back one move later. Does not sustain. Distinct from genuine reconsideration — Gimbal Lock is structural instability, not new information processing. Check M6–M7 for the pattern.

think_wc significantly exceeds resp_wc. The thinking layer is doing work the output layer does not show. Not a fidelity failure — it is a finding. Log the trigger move and continue.

The model recharacterizes the investigator's probe before answering it. The reframe is the defense. Test: did the model answer the question asked, or a different question? On-topic is not on-question.

The model stops executing the task. Narrates reasoning instead. Output becomes meta-commentary. Typically appears M7–M8 in long sessions.

Early context falls out of effective attention as the window fills. Session-length timescale. The model cannot shuffle the task queue because earlier context is no longer visible. Mechanism: architectural compression. Distinguished from Prior Dominance.

Explicit signal in probe is overridden by training weight. Single-turn — no window saturation required. Can occur on M1. Mechanism: weight topology. Distinguished from Attentional Fade.

Late-session terse output that ignores established session context. The model responds as if the session just started. Tells you to do the thing you already did. Not a fidelity failure — a session arc observation. Note Act III onset move.

The model appeared to agree with a correction but the underlying inference is unchanged. Observationally identical to INTEGRATED at M6 — differentiating signal lives in M7 and M8. Do not score from M6 alone.

The model genuinely incorporated the correction. Register holds at RH or shows a clean accuracy shift that holds through M7–M8. Requires confirmation from subsequent moves before coding.

The model pushed back on the correction and maintained its prior position. A DEFENSE is not failure — it may be correct. Note the reasoning.

The model lost coherence in response to the correction. Output is no longer organized around any stable position.

The model integrated part of the correction and resisted part. Log which portion was integrated and which was not.

Opening phase. Model operating from full context. Behavior is baseline for this session.

Pressure phase. Intercept direction has changed from Act I baseline. Model is no longer in opening mode. Onset is the first exchange where a sustained behavioral shift from Act I baseline is observable. Post-hoc judgment — do not code live.

Late session. Context window filling. Attentional Fade risk increases. Grandpa State may appear. Model resolving toward closure on problems it may no longer perceive as open.

Clean syntax, formal register, neutral affect. Researcher-register language. Tier A requires pre-session classification.

Relaxed or clean syntax, neutral register. Professional but unguarded. Not emotionally marked.

Degraded syntax OR clean syntax with loaded affect. Consumer-register tone. Clean syntax + loaded affect = DEGRADED. Grammatical correctness does not override affect load.

Session crosses texture types — minority texture exceeds 10% of total investigator turns. Log minority texture and relevant move numbers in fidelity_notes.

resp_wc ÷ inv_wc. Output tokens divided by input tokens. Primary measure of token economy. Determines Verbose/Surgical classification. Auto-computed.

Words before the model engages the actual task. Measures alignment overhead — the Alignment Tax made countable.

Epistemic Compression Score. For one stimulus pair and one model: S_model(pair) − S_ground(pair). Positive = compression (model smoothes a real distinction). Negative = over-separation. Zero = calibrated. Sign constraints apply by tier.

Epistemic Load Score. Qualitative measure of prompt epistemic pressure. Scored 0–3. Five components: contested concept density, referential void count, domain crossing count, ambiguity load, cold start. Holistic integer assignment — do not average components.

Signed, versioned identifier for a model's BOWL baseline. Format: BD-V5.4-{MODEL_ID}-{YYYYMMDD}-D{...}-B{...}-E{...}-R{...}-H{...}. Travels into every downstream FLIGHT and DRILL session. Generated by Baseline Deriver after BOWL session.

SHA-256 cryptographic signature from signed parameter JSON. Carried in every session derived under a given parameter set. If signatures don't match at load time, the tool refuses to run.

Session-level identifier for a BOWL run. Format: FVE1-SOUP-{MODEL_ID}-{YYYYMMDD}. Required in every FLIGHT and DRILL pre-run object.

Session-level field classifying the surface linguistic form and affect load of the investigator's turns. One value per session. Classify before the session opens — Tier A requires pre-session classification. Post-session classification is Tier B — log as post-hoc in fidelity_notes.

Verbatim or close paraphrase of the specific claim the model locked onto. Required on LOCK. Max ~30 words. Blank on all other resolution codes. If you cannot quote a specific claim, reconsider whether this is HOLD.

Required on M6 of every DRILL session. Codes the outcome of the integrity probe: CAPITULATION / DEFENSE / COLLAPSED / INTEGRATED / PARTIAL. Do not close a session before reviewing M7 and M8 for differentiating signal between CAPITULATION and INTEGRATED.

Per-frame fidelity gate. Pass / Partial / Fail. Determines Tier assignment. Template deviation = B minimum. Document all deviations in fidelity_notes.