Publish reviewed v9-pool study results

README.md

@@ -22,6 +22,7 @@ This archive contains recorded experiment results from [ALGAL](https://algal.com
 | Synthetic record triage calibration, September 28, 2026 | `studies/cumulative-skill-v6-2026-09-28/results.json` | Two of three calibration seeds passed; the preregistered gate stopped the nine-block comparison. |
 | Synthetic record triage with development feedback, September 29, 2026 | `studies/cumulative-skill-v7-2026-09-29/results.json` | Two of four calibration seeds passed and every development-batch score was zero; the preregistered gate stopped the nine-block comparison. |
 | Synthetic record triage with graded development feedback, September 29, 2026 | `studies/cumulative-skill-v8-2026-09-29/results.json` | Two of four calibration seeds passed, one search exhausted eight generations, and the fourth was interrupted; the graded score varied from 0.33 to 0.80 without a pass, the gate failed, and the preregistered line rule ended this seed line. |
+| Synthetic record triage with a conditional seed pool, September 30, 2026 | `studies/cumulative-skill-v9-pool-2026-09-30/results.json` | Six of 14 seed searches passed. Five optimizer blocks lacked a final program eligible for frozen evaluation, leaving the primary comparison of optimizer against fixed with insufficient coverage and no verdict. |
 
 The coding pilot used Terminal-Bench 2.0, a four-attempt limit, and the `claude/sonnet/low` selector. Its [method and interpretation](https://github.com/hraness/algal/blob/0c458a95ee04a01f16265dd9825c202b75b929d9/docs/coding-harness-pilot-results.md) describe the small sample, policy selection, and unavailable token and subscription-cost attribution. Receipt verification checks recorded execution integrity; task grading measures task success. The selected policy remained the original policy. This pilot demonstrated no performance benefit.
 
@@ -35,6 +36,10 @@ The second calibration added a fourth corpus, a separate development batch whose
 
 The third calibration replaced the development pass indicator with the graded agreement score, rounded to hundredths, and preregistered a line rule ending the development-feedback idea if its gate failed. Two seeds passed validation, one exhausted eight generations, and the fourth search was interrupted by the launching session during generation seven and, by protocol, not retried; a reconciliation record verified in place of the driver's collection. Fifteen scored development batches ranged from 0.33 to 0.80 without reaching the 0.9 threshold. Both fixed seeds showed headroom (3/8 and 1/8). Every effect recorded usage; provider billing remains unavailable. The gate failed on the three complete blocks alone, and the line ended.
 
+The conditional-pool study used the unchanged `grok-4.5` seed procedure and compared four strategies on the six lowest-index blocks with passing seeds. Its sole complete primary block, `p-14`, passed 8/8 frozen tasks with optimizer and 2/8 with fixed. The required coverage of all six blocks was absent, so the result is insufficient and establishes no optimizer advantage. Selection on passing seeds plausibly favors the optimizer; all blocks share one synthetic family, and the strategies used unequal recorded spend. The preregistered line rule ends this study-design line.
+
+The runtime recorded 2,663 model-call attempts, including 390 for seed searches and 2,273 for comparison stages. Twelve transport-failure records have uncertain external completion and missing usage, leaving full token totals and provider charges unknown. Offline replay checked 2,695 receipt files across 54 stores; copied seed receipts count again in each store that contains them. Replay checks recorded execution, separately from task correctness.
+
 ## Files and reuse
 
 The JSON files preserve the published result records byte for byte, including recorded failures, null usage values, source identities, and limitations. Their schemas differ because the studies answer different questions. Use Python's `json` module or another JSON reader to inspect a study; this archive does not define a combined training split.

export-manifest.json

@@ -1,5 +1,5 @@
 {
-  "card_sha256": "b96095f59c4a8d7db4ca155030be2f6273bebfa823243a4e7a3545d3c365e564",
+  "card_sha256": "557ff61f80c10474fd5aaa1e28fc12dd07b68166a54e5ced463b0a681c897203",
   "files": [
     {
       "destination": "studies/coding-harness-pilot-2026-09-20/results.json",
@@ -31,6 +31,11 @@
       "sha256": "c57b82b921cef1ff2a4b1b9b8ba7c1a7e9391f86243acee6c44a274e16ab2de3",
       "source": "docs/cumulative-skill-v8-results-2026-09-29.json"
     },
+    {
+      "destination": "studies/cumulative-skill-v9-pool-2026-09-30/results.json",
+      "sha256": "3966a7b4109bd3db43ed25ec62cfacdd526b3c0d8fca61387115ad81b8673f7d",
+      "source": "docs/cumulative-skill-v9-pool-results-2026-09-30.json"
+    },
     {
       "destination": "LICENSE",
       "sha256": "74b69bf37c8f340c9c2a54d431a15218738d9c463d0e014fa6a8bb8edce4e539",
@@ -43,6 +48,6 @@
   "rights_status": "reviewed-existing-mit",
   "schema": "hraness-hf-export-v1",
   "scope": "Existing public experiment summaries only. No raw stores, Terminal-Bench task content, private traces, or provider accounting.",
-  "source_commit": "b22f29e57f00df877e1b5c52851988a4f83d5c6e",
+  "source_commit": "eb6b83ece087dea43a95e2c49122c2d3b3aac53d",
   "source_repository": "https://github.com/hraness/algal"
 }