8000 Feat/interop selfdir by azzizzybud · Pull Request #24 · volotat/mini-AGI · GitHub
[go: up one dir, main page]

Skip to content

Feat/interop selfdir - #24

Open
azzizzybud wants to merge 24 commits into
volotat:mainfrom
azzizzybud:feat/interop-selfdir
Open

azzizzybud wants to merge 24 commits into
volotat:mainfrom
azzizzybud:feat/interop-selfdir

Conversation

@azzizzybud
Copy link
Copy Markdown

No description provided.

@volotat
volotat commented Sep 27, 2026
Copy link
Copy Markdown
Owner

Please explain with as much details as you can what you are trying to do here and why? Otherwise I will have to reject this pull request.

@azzizzybud
Copy link
Copy Markdown
Author

Hi Volotat,

Thanks for asking — I should have led with this. Here's exactly what the PR is and why.

What it does (all additive and reversible):

Tool-calling (interop). A literal-ASCII … protocol the model can emit during generation in serve.py, executed against a small registry, with injected back into the stream. No tokenizer change — the markers are ordinary bytes, so vocab stays 265 and every existing checkpoint loads unchanged. The registry ships with (a) a local port of the four Ma'at governance gates, and (b) a generic HTTP/JSON bridge (stdlib urllib) so the model can call an external endpoint — an MCP server or another agent.
Self-directed learning. An off-by-default (--selfdir) channel where, at the growth cadence, the model reads a telemetry block (loss, LR scale, pool stats) and may emit a directive — lr xN.NN (clamped [0.5, 2.0]) or grow 0|1. It's applied as an advisory multiplier on top of your plasticity controller, never past its FLOOR/CEIL, and nothing the model writes can raise or stop a run.
Governance receipts. A hashed, append-only ledger (minagi/ledger.py) recording the agent's actions — tool calls, directives, health events — so an unattended run leaves a tamper-evident trail. It's the governance counterpart to git's commit log.
Health certificate. A discrete complete-monotonicity diagnostic over held-out loss (minagi/health.py): "oscillation = regression risk," with the criterion taken from a machine-checked Lean theorem (exp(-μx) is completely monotone, cos(bx) is not). It's advisory only — it flags, never acts.
Why:

My goal is a model that can run locally and independently — keep learning, act on tools, and be accountable — without a human watching it. Each piece is one leg of that: act (tool-calling), steer (self-direction), prove (receipts + health cert). It extends the README's own motivation — "a model that is genuinely yours… that keeps learning… nobody else can take it away" — by giving that model agency and an audit trail rather than just a chat window.

What I'm explicitly not claiming:

The self-directed learning and health cert build the channel and the diagnostic, not a capability result. The model is toy-level; whether it learns to use these well is open. They're structured so they can't hurt the run either way.
The health cert cites a complete-monotonicity theorem — not the Riemann Hypothesis. RH is stipulated (not proven) in my broader research and is not applied here; nothing in this PR depends on it.
How it respects the project's design:

No tokenizer/vocab change; existing weights load unchanged.
Advisory and never-raising; self-direction off by default.
stdlib-only, no new dependencies.
Self-contained modules with minimal wiring in serve.py/train.py, so any piece can be dropped without touching core training.
I'm happy to split this into smaller PRs (interop / self-direction / governance each standalone) or drop any part you think is outside the project's scope. Tell me what you'd want to see changed

GroveMinting added a commit to GroveMinting/mini-AGI that referenced this pull request Oct 14, 2026
…t#24

- HANDOFF.md: state for moving the work to a GPU box (same-path clone,
  env, opencode session, provider endpoint, next inventory).
- CHANGELOG.md: v0.1.0-mirror snapshot.
- requirements.txt / requirements-corpora.txt; make setup now installs
  from them (CPU torch still pinned separately).
- UPSTREAM.md: pin the synced upstream commit; the sync script now tags
  each sync upstream-YYYYMMDD, matching what the doc promised.
- HARDENING.md: record patch 0008 and a "deliberately not carried" note
  for PR volotat#24.
GroveMinting added a commit to GroveMinting/mini-AGI that referenced this pull request Oct 14, 2026
…t#24

- HANDOFF.md: state for moving the work to a GPU box (same-path clone,
  env, opencode session, provider endpoint, next inventory).
- CHANGELOG.md: v0.1.0-mirror snapshot.
- requirements.txt / requirements-corpora.txt; make setup now installs
  from them (CPU torch still pinned separately).
- UPSTREAM.md: pin the synced upstream commit; the sync script now tags
  each sync upstream-YYYYMMDD, matching what the doc promised.
- HARDENING.md: record patch 0008 and a "deliberately not carried" note
  for PR volotat#24.
GroveMinting added a commit to GroveMinting/mini-AGI that referenced this pull request Oct 14, 2026
…t#24

- HANDOFF.md: state for moving the work to a GPU box (same-path clone,
  env, opencode session, provider endpoint, next inventory).
- CHANGELOG.md: v0.1.0-mirror snapshot.
- requirements.txt / requirements-corpora.txt; make setup now installs
  from them (CPU torch still pinned separately).
- UPSTREAM.md: pin the synced upstream commit; the sync script now tags
  each sync upstream-YYYYMMDD, matching what the doc promised.
- HARDENING.md: record patch 0008 and a "deliberately not carried" note
  for PR volotat#24.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants

0