| { |
| "study_id": "causal-relational-knockout-24law-v1", |
| "status": "frozen-before-any-study-model-output", |
| "frozen_at": "2026-09-02T11:03:03.653464+00:00", |
| "scientific_question": "Does a development-fitted relational coordinate make a selective causal contribution to ordinary correct decisions, mediate counterfactual answer changes, and repair initially incorrect answer margins?", |
| "cohort": { |
| "law_identities": 24, |
| "direct": 12, |
| "inverse": 12, |
| "prompts": 192, |
| "factors": [ |
| "two numerical directions", |
| "higher/lower and greater/smaller answer vocabularies", |
| "both presentation orders" |
| ], |
| "independent_unit": "law identity", |
| "equation_free": true, |
| "disjoint_from_direction_fit_laws": true |
| }, |
| "models": { |
| "4b": { |
| "model": "google/gemma-4-E4B-it", |
| "revision": "a4c2d58be94dda072b918d9db64ee85c8ed34e3f", |
| "probe_layer": 34, |
| "development_manifest": "experiments/elicited-physics-abstraction-2026-07-18/prompt_manifest.json", |
| "development_states": "experiments/elicited-physics-abstraction-2026-07-18/representations.npz" |
| }, |
| "12b": { |
| "model": "google/gemma-4-12B-it", |
| "revision": "707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7", |
| "probe_layer": 39, |
| "development_manifest": "experiments/gemma4-12b-equation-free-relational-v1/development_prompt_manifest.json", |
| "development_states": "experiments/gemma4-12b-equation-free-relational-v1/development_states.npy" |
| }, |
| "31b": { |
| "model": "google/gemma-4-31B-it", |
| "revision": "842da3794eaa0b77d5f08bae87a17459d91ff475", |
| "probe_layer": 49, |
| "development_manifest": "experiments/gemma4-31b-equation-free-relational-v1/development_prompt_manifest.json", |
| "development_states": "experiments/gemma4-31b-equation-free-relational-v1/development_states.npy" |
| } |
| }, |
| "direction": "One unit physical-outcome centroid direction per model, fitted only on the pre-existing 16-law development cohort at the prespecified layer.", |
| "position": "final prompt token after the registered transformer block", |
| "interventions": { |
| "clean": "Unmodified forward pass.", |
| "neutralize": "Set only the normalized-state coordinate along the frozen direction to the development midpoint, retaining the orthogonal residual.", |
| "counterfactual_coordinate": "Set only that coordinate to the value in the exact matched prompt with reversed numerical change.", |
| "correct_centroid": "Set only that coordinate to the development centroid for the prompt's registered physical outcome; evaluated as error repair only on all initially wrong answer-margin cases.", |
| "wrong_centroid": "Set the coordinate to the opposite development centroid as a signed falsification.", |
| "full_counterfactual": "Replace the complete state with its exact matched counterfactual state.", |
| "full_counterfactual_direction_restored": "Apply the full counterfactual state but restore its frozen-direction coordinate to the clean receiver value.", |
| "random_controls": "Three deterministic unit directions orthogonal to the frozen direction; each receives exactly the same receiver-specific L2 perturbation as neutralization or counterfactual-coordinate replacement." |
| }, |
| "primary_endpoints": { |
| "necessity": "Law-level clean correct-answer margin minus neutralized margin; the 95% law bootstrap interval must exceed zero and the effect must exceed the mean matched random-control effect.", |
| "counterfactual_sufficiency": "Law-level movement toward the matched counterfactual answer under the one-coordinate swap; interval above zero and larger than random control.", |
| "causal_mediation": "Ratio of one-coordinate counterfactual shift to full-state counterfactual shift, reported with an untruncated law bootstrap interval.", |
| "direction_restoration": "Reduction of the full-state counterfactual effect when the frozen coordinate is restored to its clean value.", |
| "error_repair": "On every clean case whose registered answer margin is below zero, change in correct-answer margin and fraction crossing above zero after the correct-centroid intervention. No minimum error count is guaranteed." |
| }, |
| "secondary_endpoints": [ |
| "answer-margin accuracy before and after each intervention", |
| "higher/lower versus greater/smaller strata", |
| "direct versus inverse laws", |
| "answer-order agreement", |
| "next-token KL divergence and registered-answer probability mass", |
| "4B, 12B, and 31B replication without assuming monotonic scaling" |
| ], |
| "inference": { |
| "law_bootstrap_resamples": 30000, |
| "bootstrap_seed": 20260902, |
| "law_sign_permutations": 100000, |
| "permutation_seed": 20260903, |
| "random_direction_seeds": [ |
| 1701, |
| 1702, |
| 1703 |
| ] |
| }, |
| "guardrails": [ |
| "Every frozen prompt and intervention row is retained.", |
| "No layer, direction, prompt, answer vocabulary, or control is selected from study output.", |
| "Necessity is claimed only for a selective contribution of the tested linear coordinate, never unique or global necessity.", |
| "A negative neutralization result leaves the prior sufficiency claim unchanged.", |
| "The 31B study uses the same intervention definitions despite having only one archived Jacobian lens; the primary direction is raw-state based and lens independent." |
| ] |
| } |
|
|
