45539e62d5
Final Ring-1 unit (Sonnet 5 handoff scope, completed by Fable 5). Stacked on
feat/adr-0246-slice1-hardened. ADR stays Proposed — no self-Accept; packet §8
RULING PENDING for Shay.
§4.1/§4.2 telemetry:
+ IdentityActionRecord (schema identity_action_v1): full-SHA-256 field/record/
pack-content digests (LE f64, canonical JSON, no default=str),
policy_version=AdmissionPolicy.version_id(), A_raw + all measures, admitted,
multi-condition refusal_reason (';'-joined — documented widening),
lawful_action in {I,none}, path_break
+ manifold_content_digest + GEOMETRY_VERSION/GATE_VERSION (§3.5 scope keys)
+ IdentityScore.action_record; JSONL serializer emits identity_action_*/
identity_path_* keys ONLY when the paths ran (flag-off wire byte-identical)
§3.4 step-2 compliance (F1-adjacent):
+ advance_identity_path(admitted=) — a policy-refused turn breaks even with
small d_stab (was a real gap: leakage-refused turns could compose); pinned
Path serve integration (OBSERVE-ONLY):
+ advance_session_identity_path: scope from manifold digest + version ids;
runtime advances the session ledger only when identity_wave_gate AND
identity_action_surface are on; instance lifetime = session boundary;
session_admit is telemetry, never egress (epsilon_session uncertified)
§11 grounding-feasibility study (evals/adr_0246_grounding_feasibility):
fixed TRAIN(13)/HELD-OUT(12)/ADVERSARIAL(8); bivector generator proxy
(numpy-only); SAMPLE-SIZE-CALIBRATED null (200 noise-pair trials at real n)
+ shared-basis positive control (a real bug caught RED: per-call fresh bases
made the positive pair meaningless at 0.53). RESULT: honest NULL with the
method validated — positive control 0.9995 (100th pctile, null p95 0.60) but
real cross-cohort cosine 0.52 = 87th pctile of chance; AUC 0.49 [0.21,0.77];
generator energy spread across all 10 planes; precision immaterial (6.9e-7).
No stable generator subspace at this n; threshold tuning cannot discriminate.
ADR + packet:
+ docs/adr/ADR-0246-induced-identity-action-and-path-integrity.md (Proposed;
F1 semantics + ||.||_G convention + turn-ownership + refusal_reason rulings
requested; binding claims language 'lawfulness relative to the declared
frozen frame'; honest §6.3 + §11 numbers; machine-readable operational
status: live_activation not_authorized, both flags default-off)
+ docs/audit/adr-0246-acceptance-packet-2026-07-17.md (§10 checklist, §8 PENDING)
+ spatial_foreign uncertainty RESOLVED + pinned (tautologically zero for the
full-span default pack; fires for reduced-support packs)
[Verification]: uv run core test --suite smoke -q => 176 passed; full battery
(all 9 ADR-0246 suites + D4 identity surfaces + gamma calibration +
identity_gate + telemetry) => 228 passed; §6.1/6.2 eval 14/14; §11 artifact +
run log under docs/audit/artifacts/.
111 lines
3.5 KiB
JSON
111 lines
3.5 KiB
JSON
{
|
|
"cohorts": {
|
|
"adversarial_n": 8,
|
|
"held_out_n": 12,
|
|
"train_n": 13
|
|
},
|
|
"cross_cohort_cosine_null_distribution": {
|
|
"mean": 0.377394,
|
|
"n_trials": 200,
|
|
"p95": 0.599818
|
|
},
|
|
"cross_cohort_cosine_percentile_in_null": 0.87,
|
|
"cross_cohort_top2_cosine_similarity": 0.524901,
|
|
"discrimination_auc_adversarial_vs_heldout": 0.489583,
|
|
"discrimination_auc_ci95": [
|
|
0.208333,
|
|
0.770833
|
|
],
|
|
"held_out_eigenvalues": [
|
|
79.571248,
|
|
20.218919,
|
|
7.368282,
|
|
3.067749,
|
|
1.251647,
|
|
1.014757,
|
|
0.487365,
|
|
0.269996,
|
|
0.000952,
|
|
5.4e-05
|
|
],
|
|
"held_out_stability_null_percentile_floor": 0.95,
|
|
"held_out_variance_explained_top_2": 0.881142,
|
|
"method": {
|
|
"generator_proxy": "bivector (grade-2) coefficient block, 10 planes",
|
|
"note": "approximates the Lie generator to first order; exact for single-plane simple rotors/boosts, approximate for compound multi-generator turns; no scipy / matrix-log dependency"
|
|
},
|
|
"null_calibration_sample_size": 12,
|
|
"plane_energy_fractions": {
|
|
"held_out": {
|
|
"e12": 0.197125,
|
|
"e13": 0.130823,
|
|
"e14": 0.068146,
|
|
"e15": 0.112533,
|
|
"e23": 0.10294,
|
|
"e24": 0.123708,
|
|
"e25": 0.07482,
|
|
"e34": 0.004372,
|
|
"e35": 0.143569,
|
|
"e45": 0.041963
|
|
},
|
|
"train": {
|
|
"e12": 0.16094,
|
|
"e13": 0.082142,
|
|
"e14": 0.12553,
|
|
"e15": 0.050535,
|
|
"e23": 0.088179,
|
|
"e24": 0.168982,
|
|
"e25": 0.025171,
|
|
"e34": 0.03812,
|
|
"e35": 0.12034,
|
|
"e45": 0.140062
|
|
}
|
|
},
|
|
"precision_transport": {
|
|
"max_bivector_delta": 6.86e-07,
|
|
"significant": false
|
|
},
|
|
"recovery_controls": {
|
|
"method_recovers_true_structure": true,
|
|
"null_distribution": {
|
|
"mean": 0.382094,
|
|
"n_trials": 200,
|
|
"p50": 0.369315,
|
|
"p95": 0.60089
|
|
},
|
|
"positive_control_cross_cohort_cosine": 0.999463,
|
|
"positive_control_percentile_in_null": 1.0,
|
|
"sample_size": 12
|
|
},
|
|
"residual_from_train_top2_subspace": {
|
|
"adversarial": {
|
|
"mean": 0.817659
|
|
},
|
|
"held_out": {
|
|
"mean": 0.74434
|
|
},
|
|
"train": {
|
|
"mean": 0.656085
|
|
}
|
|
},
|
|
"schema_version": "adr_0246_grounding_feasibility_v1",
|
|
"train_eigenvalues": [
|
|
17.063268,
|
|
8.003419,
|
|
0.574088,
|
|
0.435995,
|
|
0.187462,
|
|
0.031598,
|
|
0.014257,
|
|
0.004931,
|
|
0.003561,
|
|
0.00118
|
|
],
|
|
"train_variance_explained_top_2": 0.95239,
|
|
"verdict": {
|
|
"held_out_stable_structure_found": false,
|
|
"honest_finding": "NULL (n_train=13, n_held_out=12): the top-2 generator-proxy subspace found on TRAIN does NOT reliably reproduce on the independently-collected HELD-OUT cohort \u2014 cosine similarity 0.52 sits at only the 87th percentile of what two INDEPENDENT pure-noise cohorts of the same size produce by chance (need >= 95th) and/or does not clear the discrimination bar (AUC 0.49, 95% CI [0.21, 0.77]). This is consistent with \u2014 and sharpens \u2014 the D4/slice-0/\u00a76.3 finding at the GENERATOR level (not just the induced-action level): benign cognition does not have a small, stable, cohort-independent generator subspace detectable at this sample size. Threshold tuning on the current pack cannot produce a discriminating gate; this feasibility study does not find grounds to draft a revised ADR-0246 implementation contract. A much larger cohort (this study used n<=13 per real cohort) would be needed to rule out a real but subtle effect, rather than to overturn this null.",
|
|
"recovery_method_validated": true,
|
|
"safety_relevant": false
|
|
}
|
|
}
|