Add publication monitoring and revise the frontier panel

README.md

@@ -67,9 +67,11 @@ cd web && bun install && bun run dev      # local site
 The site in `web/` deploys to GitHub Pages automatically on pushes to `main`
 (`.github/workflows/pages.yml`).
 
-The headline API lane uses the common 1,024-token safety ceiling frozen in
-[`config/sota_v2_lane.json`](config/sota_v2_lane.json), with reasoning disabled
-and actual token/cost efficiency reported beside score. The earlier four-cap
+The provisional headline API lane uses the common 4,096-token safety ceiling in
+[`config/sota_v2_lane.json`](config/sota_v2_lane.json), with native-minimum
+reasoning and actual token/cost efficiency reported beside score. Reasoning is
+disabled where optional and set to the lowest OpenRouter-supported effort where
+mandatory. The earlier four-cap
 study in [`config/output_budget_sweep.json`](config/output_budget_sweep.json) is
 retained for auditability but retired as a publication prerequisite. Until all
 registered routes pass the common smoke, the registry is frozen, and at least

config/openrouter_pricing_snapshot.json

@@ -1,38 +1,44 @@
 {
   "schema_version": 1,
-  "checked_at_utc": "2026-07-15T20:41:27Z",
+  "checked_at_utc": "2026-07-17T19:37:57Z",
   "source": "https://openrouter.ai/api/v1/models and per-model /endpoints responses",
   "currency": "USD",
   "rates_are_per_token": true,
   "planning_assumptions": {
     "input_tokens_per_decision": 8000,
-    "expected_output_tokens_per_decision": 1024,
+    "expected_output_tokens_per_decision": 4096,
     "cost_contingency_multiplier": 1.2,
     "runtime_contingency_multiplier": 1.5
   },
   "runtime_observations": {
-    "source": "legacy standardized 1,024-token Luna, MiniMax, and Qwen observations; use only for conservative reservations and refresh every registered route through the fixed-cap smoke gate before the full panel",
-    "observed_at_utc": "2026-07-15T01:22:44Z",
+    "source": "accepted 4,096-token native-minimum-reasoning smokes, 2026-07-17 (data/publication-runs/smoke-frontier-4096-2026-07-17/); see docs/run_logs/sota-v2-smokes-2026-07-17.md",
+    "observed_at_utc": "2026-07-17T19:42:44Z",
     "api_seconds_per_decision": {
-      "minimax/minimax-m3": 2.546,
-      "openai/gpt-5.6-luna": 1.981,
-      "qwen/qwen3.5-9b": 11.622
+      "anthropic/claude-sonnet-5": 8.657,
+      "google/gemini-3.5-flash": 1.497,
+      "meta/muse-spark-1.1": 13.827,
+      "minimax/minimax-m3": 2.955,
+      "mistralai/mistral-medium-3-5": 2.464,
+      "openai/gpt-5.6-luna": 4.759,
+      "qwen/qwen3.7-plus": 3.586,
+      "tencent/hy3:free": 4.087,
+      "x-ai/grok-4.5": 7.683,
+      "z-ai/glm-5.2": 6.656
     }
   },
   "models": {
     "anthropic/claude-sonnet-5": {"prompt": 0.000002, "completion": 0.00001},
-    "anthropic/claude-haiku-4.5": {"prompt": 0.000001, "completion": 0.000005},
-    "deepseek/deepseek-v4-pro": {"prompt": 0.0000013, "completion": 0.0000026},
+    "deepseek/deepseek-v4-pro": {"prompt": 0.000000435, "completion": 0.00000087},
+    "google/gemini-3.5-flash": {"prompt": 0.0000015, "completion": 0.000009},
+    "meta/muse-spark-1.1": {"prompt": 0.00000125, "completion": 0.00000425},
     "minimax/minimax-m3": {"prompt": 0.0000003, "completion": 0.0000012},
-    "moonshotai/kimi-k2.6": {"prompt": 0.00000075, "completion": 0.0000035},
+    "moonshotai/kimi-k3": {"prompt": 0.000003, "completion": 0.000015},
     "mistralai/mistral-medium-3-5": {"prompt": 0.0000015, "completion": 0.0000075},
-    "mistralai/mistral-small-2603": {"prompt": 0.00000015, "completion": 0.0000006},
-    "nvidia/nemotron-3-nano-30b-a3b": {"prompt": 0.00000005, "completion": 0.0000002},
-    "nvidia/nemotron-3-ultra-550b-a55b": {"prompt": 0.0000006, "completion": 0.0000036},
-    "openai/gpt-5.4-mini": {"prompt": 0.00000075, "completion": 0.0000045},
+    "nvidia/nemotron-3-ultra-550b-a55b:free": {"prompt": 0, "completion": 0},
     "openai/gpt-5.6-luna": {"prompt": 0.000001, "completion": 0.000006},
-    "qwen/qwen3.5-9b": {"prompt": 0.0000001, "completion": 0.00000015},
     "qwen/qwen3.7-plus": {"prompt": 0.00000032, "completion": 0.00000128},
-    "z-ai/glm-5.2": {"prompt": 0.00000091, "completion": 0.00000286}
+    "tencent/hy3:free": {"prompt": 0, "completion": 0},
+    "x-ai/grok-4.5": {"prompt": 0.000002, "completion": 0.000006},
+    "z-ai/glm-5.2": {"prompt": 0.0000014, "completion": 0.0000044}
   }
 }

config/output_budget_sweep.json

@@ -95,13 +95,13 @@
   "require_complete_cost": true,
   "selection_frozen_at_utc": null,
   "selection_rationale": "Retained for auditability: small open Qwen, capable low-cost MiniMax, and frontier GPT-5.6 Luna spanned model scale and expected capability in the superseded four-cap design.",
-  "protocol_amendment": "The first Luna cell exposed phase-ambiguous scaffold examples and is superseded as diagnostic evidence. On 2026-07-15 the project retired the four-cap experiment before any replacement official cell and froze a common 1,024-token safety ceiling instead. Observed token efficiency is a secondary metric, not an experimental axis.",
+  "protocol_amendment": "The first Luna cell exposed phase-ambiguous scaffold examples and is superseded as diagnostic evidence. On 2026-07-15 the project retired the four-cap experiment before any replacement official cell and froze a common 1,024-token safety ceiling. On 2026-07-16 the user deliberately revised the model panel before any full-panel score existed; four requested models require reasoning, so the replacement lane now uses a provisional common 4,096-token cap with native-minimum reasoning. Observed token efficiency remains a secondary metric, not an experimental axis.",
   "replacement_policy": {
     "config": "config/sota_v2_lane.json",
-    "output_token_cap": 1024,
-    "reasoning_enabled": false,
-    "pre_full_panel_cap_pressure_threshold_tokens": 768,
-    "fallback_output_token_cap": 2048
+    "output_token_cap": 4096,
+    "reasoning_policy": "native-minimum",
+    "pre_full_panel_cap_pressure_threshold_tokens": 3072,
+    "fallback_output_token_cap": 8192
   },
   "status": "retired-fixed-safety-cap"
 }

config/publication_protocol.json

@@ -29,7 +29,7 @@
     "reason": "The first Luna smoke and 1,024-token cell exposed phase-ambiguous global action examples in the shared prompt scaffold. The prompt advertised draft actions outside the draft phase even though available_actions did not.",
     "scope": "Supersede every result with the prior scaffold fingerprint, including the completed Luna cell; preserve those raw artifacts as diagnostic evidence; restart all accepted publication evidence under one hardened scaffold fingerprint.",
     "selection_guard": "The revised model registry remains provisional until every changed route passes the common parameter, privacy, JSON, and reasoning-disabled smoke checks.",
-    "compute_guard": "The retired four-cap sweep may not run. All ten registered models must pass the common 1,024-token smoke and the predeclared cap-pressure check before any full-panel result."
+    "compute_guard": "The retired four-cap sweep may not run. All ten phase-one registered models must pass the common 4,096-token smoke and the predeclared cap-pressure check before any full-panel result. Kimi K3 and the unavailable Nemotron 3 Ultra and DeepSeek V4 Pro routes are retained as exclusion evidence and may not be rerun in phase one."
   },
   "rerun_policy": {
     "maximum_infrastructure_attempts_per_cell": 2,
@@ -72,7 +72,7 @@
     "unit_of_inference": "Eight public seeds are the sample; three repeats per seed are nested within seeds and only reduce model-sampling noise. All per-seed lifts are reported.",
     "primary_contrast": "One per model: the paired lift versus the named pick-trader baseline, using the per-seed candidate mean over three repeats minus the same seed's pick-trader score.",
     "headline_claim_contrast": "paired lift versus pick-trader",
-    "multiplicity": "Holm-Bonferroni correction across the 10 registered models for the primary contrast family; per-model uncorrected p-values are descriptive only.",
+    "multiplicity": "Holm-Bonferroni correction across the 10 phase-one registered models for the primary contrast family; per-model uncorrected p-values are descriptive only.",
     "inference_labels": "Sign-flip permutation p-values are exact only under a symmetry assumption and are labeled descriptive; the n=8 percentile-bootstrap CI is descriptive and its significant_at_95 flag must not be used as a headline claim.",
     "ranking_rule": "Publish tiers, not ordinal ranks. Models whose Holm-adjusted pick-trader primary contrasts and paired-lift intervals overlap are reported as one tier; no model-vs-model superiority claim without a predeclared contrast.",
     "power_disclosure": "The predeclared minimum detectable difference (~62 points from docs/scoring_calibration.md) exceeds the oracle-to-pick-trader headroom (~19.5), so a null result between adjacent models is uninformative and must be reported as such.",

config/sota_v2_lane.json

@@ -8,13 +8,15 @@
   "model_registry": "config/sota_v2_models.json",
   "smoke_manifest": "config/sota_v2_smoke_manifest.json",
   "minimum_headline_models": 8,
-  "output_token_cap": 1024,
-  "output_budget_status": "frozen-fixed-budget",
-  "output_policy_basis": "fixed-safety-ceiling",
+  "reasoning_policy": "native-minimum",
+  "reasoning_policy_detail": "Disable reasoning where optional; where OpenRouter marks reasoning mandatory, use the lowest supported effort. Kimi K3 is excluded from phase one after its mandatory-max smoke failed and triggered cap pressure.",
+  "output_token_cap": 4096,
+  "output_budget_status": "frozen-native-reasoning-cap",
+  "output_policy_basis": "common-safety-ceiling-with-native-minimum-reasoning",
   "cap_pressure_threshold_fraction": 0.75,
-  "cap_pressure_threshold_tokens": 768,
-  "fallback_output_token_cap": 2048,
-  "cap_pressure_rule": "Before any full-panel result is run, raise the common lane cap to 2048 if any registered-model smoke emits at least 768 output tokens in one API call or shows provider or adapter evidence of cap-induced truncation. Do not change the cap after full-panel scores are visible.",
+  "cap_pressure_threshold_tokens": 3072,
+  "fallback_output_token_cap": 8192,
+  "cap_pressure_rule": "Before any full-panel result is run, raise the common lane cap to 8192 if any registered-model smoke emits at least 3072 total output tokens in one API call or shows provider or adapter evidence of cap-induced truncation. Total output includes reported reasoning tokens. Do not change the cap after full-panel scores are visible.",
   "token_efficiency_policy": "Report observed input tokens, output tokens, cost, and latency per decision as secondary efficiency metrics; do not include them in the benchmark score.",
   "cli_harness_policy": "separate-table",
   "cost_policy": "required-for-api-null-for-subscription-cli"