test(agent): retain corrected Korean sampling results

Author: chaxus

docs/evaluations/2026-10-04-writing-model-decision-index.md

@@ -109,3 +109,8 @@ Gemma 4 E2B now loads in the matched browser CPU runtime. Its [three remaining o
 The [model-recommended sampling contrast](2026-10-07-gemma4-sampling-analysis.md) retains the explicitly pending budget state on one known Chinese source, while product greedy sampling repeats the omission. Both preserve original product messages/schema and have exact native application/history. This is a joint sampling-configuration comparison with one sample each, about 127–129 seconds per request, not general quality or a new default. Gemma's imported 2.84-GB model remains a higher-resource desktop diagnostic, not the lightweight CPU availability fallback.
 
 The [frozen transfer protocol](2026-10-07-gemma4-sampling-seven-language-protocol.md) starts seven source-language versions across rewrite, summary and translation with the unchanged recommended sampling candidate. Sources are correlated and shared across three tasks; 21 task executions do not mean 21 independent sources. Translation directions cover English/Chinese targets only. At this index update the process is in progress, with no completed-screen acceptance claim. Native-speaker review, broader language directions, usable latency and full device/editor/offline lifecycle requirements remain open. The original sixteen-report table and its byte-bound outcome counts are unchanged.
+
+
+### Completed Korean readiness correction (October 7)
+
+The separately frozen corrected Korean run completed three real inference requests after exact native source readiness. Rewrite and translation retain the checked facts; summary omits Mei Tan and the technical team. Together with the original 18 valid requests, coverage is 21 valid task executions, with the original three zero-inference setup failures preserved separately. Manual narrow reviews total 11 passes, six failures and four uncertain results; six of seven summaries fail. Seven translated versions of one source are correlated and cannot establish independent accuracy. No candidate is adopted. See `2026-10-07-gemma4-sampling-korean-corrected-analysis.md` and its hash-bound receipts.

docs/evaluations/2026-10-07-gemma4-sampling-korean-corrected-analysis.md

@@ -0,0 +1,9 @@
+# Corrected Korean sampling transfer
+
+The first three valid Korean inference requests completed after waiting for exact native source readiness. The original three empty-source failures remain in the separate original receipt. Model, prompts, schema and sampling remain unchanged: temperature 1, top_p 0.95, top_k 64, seed 42 and max_tokens 512.
+
+Rewrite and English translation retain the checked facts. The summary omits Mei Tan and the technical team despite the explicit instruction to retain both. All three outputs were applied through native editing and exact Undo/Redo checks passed. Browser contexts and the process closed; no page or chat errors were recorded. Request durations were 138.6–142.7 seconds, excluding model loading; this is not an isolated performance benchmark.
+
+Combined with the original 18 valid requests, there are 21 valid task executions across seven translated versions of one source: 11 narrow passes, six explicit failures and four uncertain reviews. Six of seven summaries fail the requested factual or language checks; only English passes this narrow screen. There were 24 attempts including the original three setup failures. These correlated sources do not establish an independent accuracy rate. Translation covers two target directions only.
+
+No model adoption or seven-language quality acceptance follows. More independent sources, translation directions, mixed-language text, numbers/dates, injection, latency and physical-device coverage remain required. The receipt and manual reviews are hash-bound separately; the original full-inference gate must continue to reject the original failed attempt.

docs/evaluations/2026-10-07-gemma4-sampling-korean-corrected-bindings.json

@@ -0,0 +1,31 @@
+{
+  "evidenceSHA256": {
+    "2026-10-07-gemma4-sampling-korean-corrected-native.json": "7fb2c9414fe459f81ab68dda128dfc697f0baa99eb8cf39095f8779d8e720ac7",
+    "2026-10-07-gemma4-sampling-korean-corrected-review.json": "cfa5fc26496a113180eef443578fe0ffd2417afb01ddedb6d5650cd036546736",
+    "2026-10-07-gemma4-sampling-korean-corrected-native.mjs": "b033c2cb8aec343d67e1477941bdaf39fdf942c4c39b47af95026c2f62439e0e",
+    "2026-10-07-gemma4-sampling-korean-corrected-protocol.md": "f849d7304d37cf995b24aa6e5f2584ec3e3e642ab69820b77f3bc2e574e6c770",
+    "2026-10-07-gemma4-sampling-seven-language-cases.json": "65e4ffe4ea447ad0c83f24f11bcb49513301aac8eed35d4098a99527df488095"
+  },
+  "initialSnapshot": {
+    "scope": "before corrected process launch; original process terminal exit1",
+    "sha256": {
+      "docs/evaluations/2026-10-07-gemma4-sampling-korean-corrected-native.mjs": "b033c2cb8aec343d67e1477941bdaf39fdf942c4c39b47af95026c2f62439e0e",
+      "docs/evaluations/2026-10-07-gemma4-sampling-korean-corrected-protocol.md": "f849d7304d37cf995b24aa6e5f2584ec3e3e642ab69820b77f3bc2e574e6c770",
+      "docs/evaluations/2026-10-07-gemma4-sampling-seven-language-cases.json": "65e4ffe4ea447ad0c83f24f11bcb49513301aac8eed35d4098a99527df488095",
+      "dist/sw.js": "3010032cc62e0025136352e2c741fc6a7dc9447378bbe41686a5a8f133a8416d",
+      "dist/assets/editor-C3APg_Fa.js": "1e57380c31bfd97cdbb3d35294b3b2a4c6ace2990d6ec31b8b59cff3da065dd7",
+      "dist/assets/client-D5_UYdz1.js": "90a8b6063930ea12fff32dc5795f3c7b580ab67c9f5061dc7d3a080770f45745"
+    }
+  },
+  "unchangedAfterRun": {
+    "docs/evaluations/2026-10-07-gemma4-sampling-korean-corrected-native.mjs": true,
+    "docs/evaluations/2026-10-07-gemma4-sampling-korean-corrected-protocol.md": true,
+    "docs/evaluations/2026-10-07-gemma4-sampling-seven-language-cases.json": true,
+    "dist/sw.js": true,
+    "dist/assets/editor-C3APg_Fa.js": true,
+    "dist/assets/client-D5_UYdz1.js": true
+  },
+  "processExitCode": 0,
+  "validInferenceRows": 3,
+  "adopted": false
+}