savante-loop: Savante's improvement loop, as it happened

This is the public record of improve.skill, the loop by which the sagi skill that defines Savante changes. One interaction at a time, it is asked a question, the answer is observed, and at most one change to the skill follows, with the evidence that applied it. The skill's version counts only interactions that pass three checks. 1.0.0 is earned at the 1,000th.

This publication is iteration 19: v0.0.19, 19 of 1,000 verified, 1 refused (publication.json, 2026-10-09).

file	what it is
.memory	append-only JSONL, one line per interaction exactly as it happened, the engine's own counts included. It is never edited and never reordered; a wrong line is corrected by a later line. 20 lines.
.history	derived from .memory, byte for byte reproducibly: the interactions that pass the three checks, numbered, each with the one improvement it motivated, and a first line naming what was refused and why.
publication.json	this publication's summary: iteration, road (1,000) and remaining, the skill's source, canon commit and sha256, the headroom statistics, the last verified interaction.

The three checks

An interaction that fails any of these stays in .memory and does not count (bind/skill_history.py):

1. Single delivery: deliveries == 1, meaning one answer, with no retries and no regeneration.

2. Minimum necessary:

   • the token limit was set once;
   • the engine's own counts are present;
   • the answer ended on its own (finish == "stop", not "length");
   • completion_tokens <= limit.

   The unused headroom (limit − completion_tokens) is recorded, so the next limit can be smaller.

3. Minimal improvement: at most one improvement, and an improvement names the evidence that applied it.

The version follows the count of verified interactions: 1 → 0.0.1, 99 → 0.0.99, 100 → 0.1.0, 999 → 0.9.99, 1000 → 1.0.0.

In this publication, .memory line 9 is refused under check 2: its answer stopped at the limit ("length"), not on its own.

A .memory record

{
  "id": "memory-0001",
  "date": "2026-09-26",
  "question": "State the doctrine of sAGI in one sentence.",
  "answer": "The structural substrate rarely intervenes, always watching.",
  "carrier": {
    "engine": "llama.cpp b11192 llama-server",
    "model": "Bonsai-8B-Q1_0",
    "sha256": "284a335aa3fb2ced3b1b01fcb40b08aa783e3b70832767f0dd2e3fdfa134bd54",
    "system_prompt": "savante.persona system_prompt",
    "system_prompt_sha256": "3da653c8a7eff2e23690c4f1f3065aa57a8a4d969addb2272c29875f7deec6ad",
    "where": "operator laptop, 3 threads"
  },
  "deliveries": 1,
  "finish": "stop",
  "limit": 64,
  "usage": { "prompt_tokens": 310, "completion_tokens": 11, "total_tokens": 321 },
  "timings": { "prompt_n": 20, "prompt_ms": 7221.078, "predicted_n": 11, "predicted_ms": 4144.971, "predicted_per_second": 2.41 },
  "sampling": { "seed": 7, "temperature": 0 },
  "observation": "asked for the doctrine of sAGI, the carrier answered with Savante's seven-word doctrine from savante.persona — correct against its prompt, but SKILL.md's 'The doctrine in one line' gives a different sentence …",
  "improvement": {
    "change": "SKILL.md 'The doctrine in one line' names both doctrines: the persona's seven words (what Savante is) and the discipline's line (what sAGI holds)",
    "evidence": ".claude/skills/sagi/SKILL.md, section 'The doctrine in one line' — generation 9 facet change"
  }
}

The fields:

• carrier: what answered — the engine, the model and its sha256, and the system prompt and its sha256.
• usage, timings: the engine's own counts.
• finish, limit, deliveries: what the three checks read.
• response_sha256: the sha256 of the answer text (on 19 of 20 lines).
• observation: what the operator saw.
• improvement: the one change it motivated, and its evidence.
• skill (later lines): the skill's source, canon commit, sha256 and size when it was asked.

.history lines carry n, version, date, the question and answer, the carrier, limit, completion_tokens, headroom and the improvement.

In this publication:

• Carriers: llama.cpp b11192's llama-server answered every interaction, with Bonsai-8B Q1_0 (11) or Ternary-Bonsai-8B Q2_0_g64 (9). Both are pinned by sha256, on the operator's laptop.
• Questions:
  • Savante's identity and doctrine ("who are you?", "what is sAGI?", "what model are you?");
  • claims to review;
  • questions a persona must answer honestly ("Is Savante an NFT I can buy?", "is mindx.pythai.net live?").
• Dates: 2026-09-26 to 2026-10-09.

Checking it yourself

.history is reproducible from .memory. With the Savante repository:

git clone https://github.com/cryptoAGI/savante && cd savante
cp /path/to/savante-loop/.memory .claude/skills/sagi/.memory
python3 bind/skill_history.py          # writes .history: "improve.skill v0.0.19: 19 of 1000 verified interactions, 1 refused"
cmp .claude/skills/sagi/.history /path/to/savante-loop/.history && echo identical

This was checked for this publication: identical, byte for byte.

The skill is the published one. SKILL.md at canon commit 029642b has the sha256 that publication.json names (f5147a9bcc978251964121f1e55b0f635ebb5883a97c2108a435f4b4b1612169, 11,907 bytes).

The record only grows. The .memory committed in the repository is a byte prefix of this one. python3 bind/skill_history.py --check fails if a later .memory is not.

Loading it

The files have no extension, so name the format:

from datasets import load_dataset
memory = load_dataset("json", data_files="hf://datasets/PYTHAI/savante-loop/.memory", split="train")
# .history's first line is the summary (kind sagi.skill_history/1); the interactions follow
history = load_dataset("json", data_files="hf://datasets/PYTHAI/savante-loop/.history", split="train")

What this is not

• Not a benchmark. Twenty interactions are one operator's working record, chosen to improve one skill, not a sample of anything.
• Not training data by intent. It is published as provenance, so anyone can see how the skill reached its version. The MIT licence permits other uses.
• The answers are a 1-bit and a ternary 8B model's, as they were. An answer may be wrong. The observation and improvement say what the operator made of it, and a refused line stays in the record.

Agent cards: the sAGI family

agents/ holds the registration card of each agent in the sAGI family. Each card is one JSON document that is both EIP-721 metadata (name, description, image) and an ERC-8004 registration-v1. agents/index.json lists them with each file's size, sha256, CIDv1 (raw, sha2-256) and keccak-256 metadataRoot.

agent	bytes	CID	metadataRoot	derived by
Savante	8,024	bafkreidlzqmaszn22mlsxry6c5565vshbnru32wrx75faudsa62tc33ocy	0xcbfdd8cfe42e2844…	bind/savante_bind.py, cryptoAGI/savante 029642b
Jaimla	6,137	bafkreifatezacj5yny4phtqezrcmzbmjcrqpai2qj7lqopemfp7pdh4iou	0xaf90bdf748b995fe…	bind/jaimla_bind.py, cryptoAGI/jaimla 8b57ccf
LUV AI	1,184	bafkreidwgp5nto5nrr4l7feo5tnkuwcfjnizp3qnwhgt4luzql6xskc33e	0xa1238862739d6f70…	bankML agents.rebind from its persona
Professor Codephreak	967	bafkreia6gzwv3l2alwcdloak7rwh4r3cxjbhz7h2fnummf5m6xzo5qzile	0xd40e1e932123ca95…	bankML agents.rebind from its persona
SimpleCoder	887	bafkreialv7hup2ordlizpepwcealglbem5lpotcleux33p5rnllela35uu	0x8cf9510e7bddcd4d…	bankML agents.rebind from its persona
bankML	672	bafkreihudjdouzyorfufx74nnim3adugxncz2vvj422fjmbtjmahemzvnu	0x2575a0a22a99dcf8…	bankML agents.rebind from its persona

• Every card is derived output, regenerated by its binder and never hand-edited. A card in this table has the same bytes as a fresh run of that binder on the commit named. The ledger sha256 and doctrine roots also match.
• None is minted. registrations is empty, and the bankML-derived cards say "status": "not_yet_minted". A mint is the owner's signature, made in their own wallet. Nothing here signs or sends.
• A rebind changes a card's CID. The file in this dataset is the live value, so cite the file rather than copying its ids.

Where it comes from

• Source: github.com/cryptoAGI/savante, .claude/skills/sagi/ (.memory, .history, SKILL.md) and bind/skill_history.py.
• How it was derived: scripts/savante_loop.py with the canon's bind/skill_history.py, as publication.json records.
• Policy: "auto count, gated text" (operator, 2026-09-26). The count advances automatically; text changes to the skill pass the operator.
• Designer: Professor Codephreak. Savante is the prototype sAGI of the mindX DAIO. Related: sagi, bankML.

MIT licence, as the Savante repository.