feat: training intelligence — checker agent, checkpoint cadence, experiment checklist (Batch B+C)#95
Merged
Merged
Conversation
…klist
Batch B+C of the process-hardening work — enforce training-time
behaviour the user kept re-instructing every run (PR-005 applied to
the autonomous run):
- New training-checker agent (Sonnet, phase-in): per-model-run live
monitor spawned as training-{modelname}-checker; alerts on NaN/
divergence/overfit/stall, writes diagnosis.md + next.md. Enforced via
the always-fires Phase-4 instruction in _prompt_experiment_context +
AGENT_PHASE_MAP, independent of the plan's active-agent list.
- should_checkpoint() helper + model-builder "Checkpointing and
Disaster Recovery (REQUIRED)" section: DL every-10-epochs + best +
last (fully resumable); ML per-fold + best + persisted HPO state.
- research-scout general-AI-research track for Phase-4 iteration
(time-series/sequence modelling, method-first), pairs with checker.
- Auto-maintained .zo/experiments/CHECKLIST.md via render_checklist/
write_checklist, baked into every registry mutation.
Agent roster 20 -> 21 with full doc cascade. +20 tests (760 -> 780 on
Python 3.11 & 3.12). ruff src/ clean, validate-docs 0 failures.
PR-040 captures the enforce-not-aspirate lesson.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Deploying zero-operators with
|
| Latest commit: |
b162b38
|
| Status: | ✅ Deploy successful! |
| Preview URL: | https://bdce04af.zero-operators.pages.dev |
| Branch Preview URL: | https://claude-training-intelligence.zero-operators.pages.dev |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Batch B+C — Training intelligence
First of the process-hardening batches: turn training-time rules that previously had to be re-supplied every run into enforced platform behaviour (PR-005 applied to the autonomous run — aspirational agent prose → code paths that always fire).
What's in it
training-checkeragent (Sonnet, phase-in) — a per-model-run live monitor the Lead spawns astraining-{modelname}-checker. Tails the active experiment'smetrics.jsonl/training_status.json, alerts Model Builder + Lead on NaN/Inf, divergence, gradient blow-up, overfit, dead LR, or stall (kill a broken run early), and writes a mechanisticdiagnosis.md+ feedsnext.md. Enforced via an always-fires Phase-4 instruction inOrchestrator._prompt_experiment_context(not the plan's active-agent roster, which_agents_for_phasefilters), backed by the agent file +AGENT_PHASE_MAP.should_checkpoint(epoch, total_epochs, every=10, is_best=)(replaces the contract's previously-undefined pseudocode) + a "Checkpointing and Disaster Recovery (REQUIRED)" section inmodel-builder.md: DL every-10-epochs + best + last with fully-resumable state (optimizer/scheduler/AMP-scaler/epoch/RNG); ML per-fold + best + persisted HPO study state.render_checklist/write_checklistregenerate.zo/experiments/CHECKLIST.mdon every registry mutation: exp → hypothesis → metric → Δ vs parent → tier → top shortfall, + "Next planned".Cascade + verification
src/clean; validate-docs 0 failures (2 warnings: client-blocklist skip + the known grep-vs-pytest test-badge parameterization gap; README badge updated 743 → 780).Batches A (per-project self-evolution), D (optimization audit +
software-engineeragent), E (idle-agent shutdown + swarm reinforcement) queued next.🤖 Generated with Claude Code