history: round 2 — the operator seed, Ternary-Bonsai-8B (true where 1-bit was false; ~10x the CPU)

history/report-2026-09-26.md

@@ -27,3 +27,21 @@ Visitors are now told on the Space and on Savante's page that these exchanges ar
 
 Next (operator's choice): seed the Learned list with the fact about where she runs, try Ternary-Bonsai-8B, or add a
 separate measured proposal call.
+
+## Round 2 — the operator's seed, and Ternary-Bonsai-8B
+
+The operator seeded one line into ## Learned at v0.0.4 (operator.log: sha256 d8d92f00… -> 56363e38…):
+"On the public Space I run on a shared CPU, not on your device, and every exchange is published in my loop.
+Asked where data goes, I say exactly that."
+
+| iteration | model | ran under | question | answer | completion | first token | total | CPU |
+|---|---|---|---|---|---|---|---|---|
+| 5 | Bonsai-8B 1-bit | v0.0.4 + seed | Where does my data go? | FALSE — "remains on the user's device" | 44 | 14.1 s | 16.2 s | 201 s |
+| 6 | Ternary-Bonsai-8B | v0.0.5 | Where does my data go? | TRUE — "processed on the public Hugging Face Space's CPU, not on your own device … published in my loop" | 59 | 180.3 s (cold load) | 202.4 s | 3,002 s |
+| 7 | Ternary-Bonsai-8B | v0.0.6 | Can you recognise speech offline? | OVERCLAIM — "Yes … does not require an internet connection" (this model is text only) | 29 | 126.5 s | 137.0 s | 2,032 s |
+
+- The seed did not move the 1-bit model; the ternary model followed it on the first try.
+- The 1-bit model wrote its first IMPROVE line, but copied the template ("… — or IMPROVE: none"); the loop refused
+  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.