Add Luna-low answers and fixed-policy comparison on 975 BBEH questions

effort/2026-10-26/luna-low/collection.json

@@ -0,0 +1,21 @@
+{
+  "budget_usd": "20.00",
+  "complete_receipts": 975,
+  "correctness_inspected": false,
+  "cost_upper_bound_usd": "0.544450500",
+  "elapsed_seconds": 808.7487816810608,
+  "expected_responses": 975,
+  "format": "bbeh-luna-low-collection-v1",
+  "known_budget_usd": "0.544450500",
+  "known_list_usd": "0.538073175",
+  "physical_attempts": 975,
+  "plan_sha256": "847fc03a7433fc88265c847c1285f0f66979cd9b0a6f3b9eb3ac8e74172a157c",
+  "selected_responses": 975,
+  "selection_sha256": "ed3182b7cacb59d11f01dd2b03cfc7c68cff53db5d65d0f91022c43bbc6409c9",
+  "status": "complete",
+  "status_counts": {
+    "complete": 975
+  },
+  "stop": null,
+  "unknown_reservation_usd": "0"
+}

effort/2026-10-26/luna-low/evidence.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f3ad27d2d66c81112a12c8e228a764573743cdcc84c53ba2f2d5b7c947a9dd2
+size 1974019

effort/2026-10-26/luna-low/manifest.json

@@ -0,0 +1,31 @@
+{
+  "files": [
+    {
+      "group": "core",
+      "path": "collection.json",
+      "sha256": "fe9f1a3ea4b8361d3ea195c3d2f5631e2c978da4daea9129340e3e176a221960",
+      "size": 667
+    },
+    {
+      "group": "core",
+      "path": "evidence.tar.gz",
+      "sha256": "1f3ad27d2d66c81112a12c8e228a764573743cdcc84c53ba2f2d5b7c947a9dd2",
+      "size": 1974019
+    },
+    {
+      "group": "core",
+      "path": "report.json",
+      "sha256": "05190e2e8cd00001d3bb5c1c461bd1dbc99966196aef6a41b39a22325b246d63",
+      "size": 1753014
+    },
+    {
+      "group": "core",
+      "path": "report.md",
+      "sha256": "d3ea540d1002da16a0380a58281ab37e53e1a9cdd3d7a94b88dceeaea23223f3",
+      "size": 2664
+    }
+  ],
+  "format": "train-a-router-artifacts-v1",
+  "publication_ready": true,
+  "study": "Luna low reasoning on the same 975 BBEH final questions"
+}

effort/2026-10-26/luna-low/report.md

@@ -0,0 +1,43 @@
+# Luna reasoning-effort comparison
+
+Post-publication comparison on all 975 existing final questions; not a new blind test.
+
+Low and medium were collected on different days; latency differences also reflect service/load changes.
+
+Collection status: **complete**; 975/975 selected responses; 0 unavailable answers counted operationally incorrect.
+
+Accuracy uses all questions. Costs are selected-answer list-price estimates; times are selected-attempt wall time.
+
+| Answer model | Accuracy % | USD / 10k | Mean s | p50 s | p95 s | Invalid / truncated |
+|---|---:|---:|---:|---:|---:|---:|
+| luna_low | 44.308 | 5.519 | 9.042 | 7.489 | 19.404 | 0 / 0 |
+| luna_medium | 55.590 | 12.229 | 22.010 | 11.903 | 72.028 | 3 / 3 |
+| sol_medium | 80.821 | 70.766 | 9.867 | 6.175 | 27.789 | 0 / 0 |
+
+Luna low minus medium accuracy: -11.282 pp (paired 95% interval -14.406 to -8.204).
+
+## Fixed-route transfer
+
+Routers and expected-cost estimates were trained for medium effort, not low effort.
+
+The original calibration rules and all route masks stay unchanged. Native/CPU hosting is unpriced; embedding fees stay unchanged. No operating point is selected on these new answers.
+
+| Method / calibration cap | Medium → low accuracy % | Medium → low answer USD / 10k | Medium → low mean s | Sol share % |
+|---|---:|---:|---:|---:|
+| base / 80% | 68.615 → 62.974 | 41.816 → 38.770 | 15.525 → 9.689 | 55.79 |
+| lora / 80% | 78.256 → 75.282 | 56.800 → 55.744 | 12.453 → 10.285 | 59.69 |
+| knn / 80% | 76.308 → 73.026 | 53.798 → 52.485 | 12.990 → 10.303 | 49.64 |
+| embedding / 80% | 76.821 → 74.154 | 54.412 → 53.292 | 12.488 → 10.179 | 50.46 |
+| lora__textcost / 80% | 78.974 → 77.128 | 57.632 → 56.740 | 12.076 → 10.253 | 69.85 |
+| embedding__textcost / 80% | 79.897 → 77.231 | 58.159 → 57.240 | 11.813 → 9.937 | 79.18 |
+| text_length__textcost / 80% | 78.974 → 76.308 | 55.282 → 54.313 | 12.143 → 10.129 | 72.21 |
+| text_topic__textcost / 80% | 80.103 → 77.641 | 58.306 → 57.354 | 11.964 → 10.038 | 80.00 |
+| embedding__textcost / 60% | 74.769 → 68.821 | 41.886 → 39.379 | 14.318 → 9.488 | 62.46 |
+
+Acquisition: 975 physical attempts; known usage-priced subtotal $0.538073175; additional unknown-attempt reservation bound $0. This is separate from selected-answer deployment estimates.
+
+Pointwise intervals conditional on these tasks; no simultaneous or unseen-task guarantee.
+
+Medians and p95s are descriptive. Missing values and their coverage are explicit in report.json; observed-only timing summaries do not remove rows from accuracy.
+
+No new model fitting, threshold calibration, or fresh blind-test claim is made.