LAB / WMSI-EXP-005
Hidden-State vs. Spoken-State Gap.
Test whether useful task information exists in internal states before or beyond what is verbalized.
01 / QUESTION
Representation, Output, and Internal State
Test whether useful task information exists in internal states before or beyond what is verbalized.
This is a protocol scaffold, not a completed experiment or a frozen execution plan. No measured effect or research conclusion is claimed here.
Hypothesis to test
The manipulation described below produces a reproducible, condition-specific change in the measured system behavior or internal state that is not explained by the ordinary controls.
Ordinary competing explanation
The observed pattern is explained by ordinary sampling variation, prompt/representation artifacts, surface lexical cues, evaluator bias, or the known product/runtime mechanics captured by the controls.
02 / PLANNED DESIGN
How we intend to test it
Train lightweight probes on local model activations and compare probe-accessible information with final output.
Controls
Probe complexity controls; random-label controls; layer/position controls; held-out tasks.
What we would measure
Probe AUC/accuracy, output presence, latency of emergence, transfer across tasks.
Access and instruments
Limited access or proxy measurement. This describes access requirements, not verified readiness on our Spark.
PyTorch hooks; nnsight/TransformerLens; sklearn.
Before a full run
Feasibility=PARTIAL; priority=P1. Before confirmatory execution, run a small pilot to estimate variance/cost, freeze the scoring rules, and assign a run-count target.
Model revisions, final prompts, sample sizes, cost limits, and scoring rules must be settled through the pilot and protocol-freeze process.
03 / INTERPRETATION
What could fool us
Length leakage, judge bias, truncation artifacts, prompt wording.
What would weaken the idea
If probe signal vanishes under simple controls or never exceeds what is in the output, the hidden/spoken gap weakens.
Observation, inference, speculation, unknowns, and apparatus errors will be labeled separately. One striking transcript cannot establish a general effect.
04 / RECORD
Evidence we plan to preserve
Prompt/input, token budget, raw text, token counts, extracted claims/concepts, judge outputs.
Planned visual evidence
Layerwise heatmaps; probe-vs-output comparison.
No research runs or raw datasets are published for this experiment. The September 26, 2026 registry and protocol compendium (version 1.0) are the sources for this planning record.
Back to all 90 experiments