Amend sota-v3 design: ten-model cohort, 16x1 allocation

config/sota_v3_lane.json

@@ -6,23 +6,25 @@
   "mechanics_change_policy": "Any score-, action-, observation-, or simulator-semantic change after this preregistration requires a new contract fingerprint, invalidates all sota-v3 smoke evidence, and re-locks provider spend.",
   "preregistration_status": "provisional-pre-smoke",
   "preregistered_at_utc": "2026-07-27T16:58:41Z",
-  "design_amended_at_utc": "2026-07-28T00:00:00Z",
+  "design_amended_at_utc": "2026-08-03T00:00:00Z",
   "design_amendment": {
-    "amendment_id": "sota-v3-design-amendment-1",
+    "amendment_id": "sota-v3-design-amendment-2",
     "status": "pre-data",
-    "supersedes": "docs/run_logs/sota-v3-statistical-design-audit-2026-07-28.md",
-    "record": "docs/run_logs/sota-v3-design-amendment-2026-07-28.md",
-    "rationale": "The superseded design powered a superiority claim (+40 above pick-trader) that the frozen sota-v2 evidence contradicts: all eight eligible models trailed the reference by 180-282 score points with a 0.0 seed win rate on every model. No allocation in the 9-20 seed by 1-3 repeat grid reached 0.80 sensitivity power for that claim, and the 20-seed ceiling caps sensitivity power near 0.36 even at unbounded repeats, so the block was structural rather than a budget shortfall.",
+    "supersedes": "docs/run_logs/sota-v3-design-amendment-2026-07-28.md",
+    "record": "docs/run_logs/sota-v3-design-amendment-2026-08-03.md",
+    "rationale": "Owner-directed pre-data cohort update before any smoke or panel evidence exists: the OpenAI anchor moves from GPT-5.6 Luna Pro to plain GPT-5.6 Luna (the originally intended route, now healthy, without the Pro variant's reasoning.mode inconsistency), and DeepSeek V4 Flash 0731 plus Tencent Hy3 join as open-weight anchors on first-party FP8 routes; Thinking Machines Inkling Small was evaluated and found ineligible because no healthy route advertises response_format under the lane's frozen JSON-mode and require-parameters options. A family of ten tightens the Holm first step from 0.00625 to 0.005, so the allocation was reselected with the identical frozen power machinery; 15x1 no longer holds the 0.80 sensitivity floor at family ten and the smallest qualifying allocation is 16 seeds x 1 repeat.",
     "changes": [
-      "Primary claim restated directionally: each registered model trails pick-trader, replicating the frozen sota-v2 finding under the sota-v3 contract.",
-      "Planning effect changed from +40 to -100 score points, conservative against the -180 to -282 observed on sota-v2.",
-      "Panel repeats changed from 3 to 1: at a fixed episode budget the candidate-minus-reference lift keeps its full seed component, so repeats shrink only the within-seed noise term and seeds dominate for discrimination.",
-      "Evaluated seed grid narrowed from 9-20 to 9-16, since the qualifying allocation falls inside it."
+      "Registered model family grows from eight to ten; Holm family size is now 10 and the first-step threshold is 0.05/10 = 0.005.",
+      "GPT-5.6 Luna Pro is replaced by plain GPT-5.6 Luna on the first-party OpenAI route, removing the unresolved Pro reasoning.mode inconsistency.",
+      "DeepSeek V4 Flash 0731 and Tencent Hy3 are added as open-weight anchors, both on first-party FP8 routes.",
+      "Selected allocation moves from 15 seeds x 1 repeat to 16 seeds x 1 repeat (16 episodes/model): sensitivity power 0.8488, Wilson 95% CI [0.841645, 0.855688]; base power 0.9527, Wilson 95% CI [0.948363, 0.95669].",
+      "The pending private seed panel is now 16 seeds; its identity remains unfrozen pending separately authorized generation."
     ],
     "unchanged": [
-      "alpha 0.05, Holm-Bonferroni across a fixed family of eight",
+      "alpha 0.05, Holm-Bonferroni across the fixed registered family",
       "exact two-sided enumeration sign-flip test with the seed as the unit of inference",
-      "0.80 conservative-sensitivity familywise all-reject power target",
+      "planning effect -100 score points and the 0.80 conservative-sensitivity familywise all-reject power target",
+      "historical variance components, sensitivity multipliers, 10000 trials, simulation seed 2026072800",
       "no model-to-model tiers or ordinal ranking",
       "every execution, spend, and publication authorization remains false"
     ]
@@ -40,9 +42,15 @@
     "status": "frozen",
     "claim_direction": "trails-reference",
     "primary_claim": "Every registered model-plus-compact-scaffold system trails deterministic pick-trader on seed-paired mean lift under the frozen sota-v3 mechanics.",
-    "evaluated_seed_range": [9, 16],
-    "evaluated_repeat_range": [1, 3],
-    "holm_family_size": 8,
+    "evaluated_seed_range": [
+      9,
+      16
+    ],
+    "evaluated_repeat_range": [
+      1,
+      3
+    ],
+    "holm_family_size": 10,
     "target_effect_score_points": -100,
     "target_effect_basis": "Conservative against the frozen sota-v2 observed lifts of -180 to -282 score points at a 0.0 seed win rate for all eight eligible models.",
     "target_familywise_all_reject_power": 0.8,
@@ -60,24 +68,30 @@
       "interaction_variance_floor_as_historical_noise_fraction": 0.1
     },
     "selected_allocation": {
-      "seed_count": 15,
+      "seed_count": 16,
       "repeats": 1,
-      "episodes_per_model": 15,
-      "minimum_exact_two_sided_sign_flip_p_value": 6.103515625e-05,
-      "holm_first_step_threshold_at_alpha_0_05": 0.00625,
+      "episodes_per_model": 16,
+      "minimum_exact_two_sided_sign_flip_p_value": 3.0517578125e-05,
+      "holm_first_step_threshold_at_alpha_0_05": 0.005,
       "exact_sign_flip_holm_feasible": true,
-      "base_power_estimate": 0.9461,
-      "base_power_wilson_ci95": [0.9415, 0.950357],
-      "sensitivity_power_estimate": 0.8357,
-      "sensitivity_power_wilson_ci95": [0.828309, 0.842834]
+      "base_power_estimate": 0.9527,
+      "base_power_wilson_ci95": [
+        0.948363,
+        0.95669
+      ],
+      "sensitivity_power_estimate": 0.8488,
+      "sensitivity_power_wilson_ci95": [
+        0.841645,
+        0.855688
+      ]
     },
     "selection_rule": "Smallest allocation with 9-16 seeds and 1-3 repeats whose conservative-sensitivity familywise-power Wilson 95% lower bound is at least 0.80.",
     "decision_required": null
   },
   "seed_panel": {
     "status": "pending-authorized-generation",
     "name": null,
-    "count": 15,
+    "count": 16,
     "sha256": null
   },
   "reference_agent": "pick-trader",
@@ -87,7 +101,7 @@
   "smoke_manifest": "config/sota_v3_smoke_manifest.json",
   "publication_protocol": "config/sota_v3_publication_protocol.json",
   "pricing_snapshot": "config/sota_v3_pricing_snapshot.json",
-  "minimum_headline_models": 8,
+  "minimum_headline_models": 10,
   "reasoning_policy": "catalog-pinned-pending-live-route-verification",
   "output_token_cap": 4096,
   "output_budget_status": "provisional-pre-smoke-validation",
@@ -101,7 +115,7 @@
   "panel_execution_authorized": false,
   "publication_authorized": false,
   "blockers": [
-    "Generate and commit the 15-seed private panel under separate owner authorization; seed identity is unfrozen until its salted commitment hash is recorded.",
+    "Generate and commit the 16-seed private panel under separate owner authorization; seed identity is unfrozen until its salted commitment hash is recorded.",
     "Complete authenticated exact-route and privacy verification for the catalog-selected model cohort before any provider call.",
     "Live-verify and pin each exact provider route, endpoint name, supported parameters, privacy policy, and reasoning requirements.",
     "Validate the provisional common 4,096-token safety ceiling on one strict smoke per registered route; any predeclared cap-pressure trigger invalidates all v3 smokes and requires one symmetric pre-panel amendment.",