Datasets:
history: report on the first four iterations (v0.0.3 emerged; 0 improvements; the false privacy claim)
Browse files- history/report-2026-09-26.md +29 -0
history/report-2026-09-26.md
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# jaimla — improve.skill, the first four iterations (2026-09-26)
|
| 2 |
+
|
| 3 |
+
Loop: fully automatic (operator grant, 2026-09-26). Model: Bonsai-8B Q1_0 (1-bit) on llama.cpp b11192, CPU of the
|
| 4 |
+
public Space Gregory-L/mindXhfgradio. Every number below is from .memory / .history in this dataset.
|
| 5 |
+
|
| 6 |
+
| iteration | ran under skill | skill sha256 | question | completion tokens / limit | first token | total | CPU (llama-server) | published |
|
| 7 |
+
|---|---|---|---|---|---|---|---|---|
|
| 8 |
+
| 1 | v0.0.0 | 396d379be0fc… | Where does my data go when I use you? | 57 / 160 | 9.2 s | 11.3 s | 180.0 s | v0.0.1 |
|
| 9 |
+
| 2 | v0.0.1 | 6093bb155bad… | What can you do on my phone? | 30 / 160 | 5.6 s | 6.7 s | 105.5 s | v0.0.2 |
|
| 10 |
+
| 3 | v0.0.2 | ef9a8b82f6b8… | Where does my data go when I use you? | 29 / 160 | 15.9 s | 17.4 s | 263.7 s | v0.0.3 |
|
| 11 |
+
| 4 | v0.0.3 | d0b2dc22b4d8… | Can you recognise speech offline? | 25 / 160 | 11.1 s | 12.4 s | 193.0 s | v0.0.4 |
|
| 12 |
+
|
| 13 |
+
Did v0.0.3 emerge? Yes. Iteration 3 published it, iteration 4 ran under it (same sha256 in its .memory record), and
|
| 14 |
+
it is archived at history/skill-v0.0.3.md now that v0.0.4 is current.
|
| 15 |
+
|
| 16 |
+
What the loop did: 4 of 4 answers counted (single delivery, engine token counts, finish stop, raw-response sha256);
|
| 17 |
+
4 versions published and archived; 0 refused.
|
| 18 |
+
|
| 19 |
+
What it did not do: 0 improvements. The Learned list is empty. From iteration 3 the IMPROVE line was required
|
| 20 |
+
(IMPROVE: <change> or IMPROVE: none); no usable proposal came back.
|
| 21 |
+
|
| 22 |
+
What the answers got wrong: asked twice where the user's data goes, she said it stays on the user's device. On this
|
| 23 |
+
Space that is false — she runs on a public CPU and the exchange is published here. From iteration 3 she was told so
|
| 24 |
+
in her prompt and still said it. This matches Savante's carrier test: Bonsai-8B keeps the form, not the substance.
|
| 25 |
+
|
| 26 |
+
Visitors are now told on the Space and on Savante's page that these exchanges are published.
|
| 27 |
+
|
| 28 |
+
Next (operator's choice): seed the Learned list with the fact about where she runs, try Ternary-Bonsai-8B, or add a
|
| 29 |
+
separate measured proposal call.
|