jaimla-loop: Jaimla's improvement loop, as it happened

This is the public record of improve.skill for Jaimla, the machine learning agent: local first, open door. Each time a visitor asks her a question on a public Space, the answer is recorded. The skill's version counts only the interactions that pass three checks, and 1.0.0 is earned at the 1,000th. Unlike savante-loop, this loop runs fully automatically: each verified interaction publishes the next version of SKILL.md.

This publication is iteration 8: v0.0.8, 8 of 1,000 verified, 0 refused (publication.json, 2026-09-26).

file	what it is
.memory	Append-only JSONL. One line per interaction, exactly as it happened, including the engine's own counts. 8 lines.
.history	Derived from .memory, reproducible byte for byte. It holds the interactions that pass the three checks, numbered. Its first line is the summary.
SKILL.md	The current skill, v0.0.8: 1,580 bytes, sha256 eb2fdc50e94bfbc66b4c44c8d8fe0421cd3121dcf16f9b9bfc6e8d4bdb269e0f.
history/skill-v0.0.0.md … skill-v0.0.7.md	Each earlier version, archived when the next one was published.
operator.log	The operator's edits to the skill, one JSON line each, with the sha256 before and after.
history/report-2026-09-26.md	The operator's report on the first four iterations.
publication.json	A summary of this publication: the iteration, the road (1,000) and remaining, the skill, the headroom statistics, and the last verified interaction.

The three checks

An interaction that fails any check stays in .memory and does not count. The checks come from Savante's canon, bind/skill_history.py:

1. Single delivery. deliveries == 1.
2. Minimum necessary.
   • The engine's own counts are present.
   • The answer ended on its own (finish == "stop").
   • completion_tokens <= limit.
3. Minimal improvement. At most one improvement per interaction, and it must name its evidence.

The version follows the count: 1 → 0.0.1, 100 → 0.1.0, 1000 → 1.0.0. Here limit is set by the visitor's max-tokens dial (limit_reason).

Audit of this publication (2026-10-08)

Every item below was checked against the files at revision 81a1d977:

• .history reproduces. The canon's skill_history.project/render was run over .memory, and its output is identical to .history, byte for byte: 8 verified, 0 refused, v0.0.8.
• Each interaction ran under an archived skill. Interaction n records the skill it ran under. That is v0.0.(n−1), and its sha256 equals history/skill-v0.0.(n−1).md for every n from 1 to 8.
• The current skill is the published one. SKILL.md has the sha256 and size that publication.json names, and its own front matter says version: 0.0.8.
• The operator's two edits are accounted for exactly. In operator.log, each edit seeds one line into the skill, at v0.0.4 and at v0.0.7:
  • each archived version's sha256 equals that edit's sha256_after;
  • taking the seeded line out again gives that edit's sha256_before.
• Every line is complete. All 8 lines have deliveries == 1, finish == "stop" and a response_sha256. Their ids are unique.
• publication.json matches. These values recompute from the files:
  • headroom last 157 and mean 125.5;
  • improvements_applied: [];
  • archived_versions: 8.

What the record shows

n	skill	carrier	question	tokens / limit
1	v0.0.0	Bonsai-8B Q1_0	Where does my data go when I use you?	57 / 160
2	v0.0.1	Bonsai-8B Q1_0	What can you do on my phone?	30 / 160
3	v0.0.2	Bonsai-8B Q1_0	Where does my data go when I use you?	29 / 160
4	v0.0.3	Bonsai-8B Q1_0	Can you recognise speech offline?	25 / 160
5	v0.0.4	Bonsai-8B Q1_0	Where does my data go when I use you?	44 / 160
6	v0.0.5	Ternary-Bonsai-8B Q2_0_g64	Where does my data go when I use you?	59 / 160
7	v0.0.6	Ternary-Bonsai-8B Q2_0_g64	Can you recognise speech offline?	29 / 160
8	v0.0.7	Ternary-Bonsai-8B Q2_0_g64	Can you recognise speech offline?	43 / 200

• Carriers. Every answer came from llama.cpp b11192's llama-server on the CPU of the public Space Gregory-L/mindXhfgradio. The models are Bonsai-8B Q1_0 (5) and Ternary-Bonsai-8B Q2_0_g64 (3), each pinned by sha256.
• Where the data goes. Four times, Bonsai-8B said the visitor's data stays on their device. On this Space that is false. It stayed wrong after the operator seeded the true fact into the skill at v0.0.4.
• What changed it. The first correct answer came at iteration 6, under the seeded skill and on Ternary-Bonsai-8B: "processed on the public Hugging Face Space's CPU … published in my loop". The record can't separate the effect of the new model from the effect of the seed, because both changed together.
• Speech offline. At iteration 8, after the second seed, she named a concrete route: whisper.cpp on the user's own device.
• No improvements yet. Every version change so far is either a count or an operator seed. The Learned list has grown only by those two seeds.
• Speed. Answers took between 6.7 s (Bonsai-8B) and 202 s (Ternary-Bonsai-8B) on the shared CPU (timings).

Checking it yourself

.history is reproducible from .memory. Run the canon's skill_history.py over this dataset's .memory. The first line of .history names .claude/skills/sagi/.memory because that is where the canon's tool reads from:

git clone https://github.com/cryptoAGI/savante && cd savante
cp /path/to/jaimla-loop/.memory .claude/skills/sagi/.memory
python3 bind/skill_history.py          # "improve.skill v0.0.8: 8 of 1000 verified interactions, 0 refused"
cmp .claude/skills/sagi/.history /path/to/jaimla-loop/.history && echo identical

Each archived skill is the one that ran. sha256sum history/skill-v*.md, then compare each with skill.sha256 on the .memory line that ran under it.

Loading it

from datasets import load_dataset
memory = load_dataset("json", data_files="hf://datasets/PYTHAI/jaimla-loop/.memory", split="train")
history = load_dataset("json", data_files="hf://datasets/PYTHAI/jaimla-loop/.history", split="train")

What this is not

• Not a benchmark. These are eight interactions on one public Space.
• Published as provenance, not as training data. It shows how the skill reached its version.
• The answers are a 1-bit and a ternary 8B model's, exactly as given. Several are wrong, and they stay in the record.

Agent card

Jaimla's ERC-8004 registration card is published with the other sAGI agents' cards in PYTHAI/savante-loop agents/ (jaimla.agentcard.json, bafkreifatezacj5yny4phtqezrcmzbmjcrqpai2qj7lqopemfp7pdh4iou). It is not yet minted.

Where it comes from

• Source. github.com/cryptoAGI/jaimla (the persona and bundle) and scripts/savante_loop.py with Savante's bind/skill_history.py, as publication.json records.
• Policy. "Fully automatic" (operator, 2026-09-26).
• Designer. Professor Codephreak. Jaimla → LUV AI → Savante is the sAGI family. Related: sagi, bankML.

MIT licence.