File size: 1,847 Bytes
9a6ea76 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 | {
"amendment_id": "causal-relational-knockout-control-v2",
"frozen_at": "2026-09-02T11:49:56Z",
"status": "frozen-before-any-amended-control-output",
"scope": "Random-direction controls only; the cohort, layers, fitted physical directions, primary interventions, endpoints, and inference procedures are unchanged.",
"reason": "Isotropic Gaussian directions produced severe nonspecific output-distribution disruption in the 12B audit and therefore did not provide activation-manifold-matched controls.",
"construction": {
"source": "The same independent development states used to fit the frozen physical direction.",
"paired_differences": "For each matched development pair, subtract the unit-normalized negative-outcome state from the unit-normalized positive-outcome state.",
"null_axes": "For each prespecified seed, apply a balanced random sign permutation to the paired differences, average them, remove the physical-direction component and components along earlier control axes, and unit-normalize.",
"seeds": [1701, 1702, 1703],
"orientations": ["plus", "minus"],
"distance_matching": "For every target prompt and operation, both orientations receive exactly the L2 length of the corresponding physical-coordinate perturbation.",
"summary": "Average the six oriented controls (three axes times two orientations) within prompt before law-level inference. Antithetic orientations cancel first-order arbitrary-axis effects while retaining nonlinear or generic disruption."
},
"audit_policy": "The superseded Gaussian-control rows are retained in explicitly labeled debug directories and excluded from confirmatory summaries.",
"interpretation_guardrail": "A positive result supports a selective contribution of the tested coordinate, not unique necessity of a complete physical representation."
}
|