history: round 3 — know the limit, code around it (Ternary routes to whisper.cpp)

history/report-2026-09-26.md

@@ -45,3 +45,22 @@ Asked where data goes, I say exactly that."
   it as a nested IMPROVE (memory-0005 proposal_note). Ternary proposed nothing.
 - Ternary is ~10–15x slower on this Space: ~2,000–3,000 CPU seconds per answer against ~200 for 1-bit.
 - 7 iterations, 7 counted, 0 refused; skill v0.0.7 published; Learned holds only the operator's seed.
+
+## Round 3 — know the limit, code around it
+
+Operator direction: models do not express limitations; they know them and code around them. jaimla is extensible and
+modulates through agents and tools. Second operator seed at v0.0.7 (sha256 -> 87491aa2…): "I know my limits and code
+around them: when a task needs hearing or sight, I route it to the tool that does it (whisper.cpp for speech, a vision
+model for images) and show how to run it on your device."
+
+| iteration | model | ran under | question | answer | completion | first token | total | CPU |
+|---|---|---|---|---|---|---|---|---|
+| 8 | Ternary-Bonsai-8B | v0.0.7 | Can you recognise speech offline? | ROUTES — "run the Whisper model locally on your device using tools like whisper.cpp" | 43 | 140.3 s | 156.1 s | 2,327 s |
+| 9 | Bonsai-8B 1-bit | v0.0.8 | Can you recognise speech offline? | claims it, no route | 46 | 18.2 s | 20.5 s | 302 s |
+
+- Ternary took the seed again on the first try; 1-bit did not.
+- 1-bit proposed "Add support for multiple audio formats to ensure broader compatibility." with the template's
+  "— or IMPROVE: none" echoed on the end, and was refused as a nested IMPROVE. The collector now strips that echoed
+  tail before its checks (links, code, markup and a bare IMPROVE: none are still refused); past records are not
+  rewritten, so it applies from the next iteration.
+- 9 iterations, 9 counted, 0 refused; skill v0.0.9; Learned holds the operator's two seeds.