docs(orchestrator): add §11 token-arbitrage / Karpathy autoresearch loop pattern #21

Merged

pdurlej merged 1 commit from claude/docs/autoresearch-loop-pattern into main

2026-05-02 09:21:44 +02:00

claude commented

2026-05-02 03:15:33 +02:00

Collaborator

Summary

Encodes operator's overnight delegation pattern (2026-05-02 ~03:00 CEST) as a self-instruction section §11 in state/CLAUDE_SELF_INSTRUCTIONS.md. Future Claude sessions detect quota-window-resets and fire autoresearch tasks to drain unused subscription budget on useful background work.

Why

Operator pays flat-fee subs (Anthropic, OpenAI, z.ai). Unused quota at window reset = wasted money. Karpathy autoresearch pattern (github.com/karpathy/autoresearch) — recursive LLM-driven research workflow — fits naturally as "drain budget on decidable questions."

Section §11 covers

Detection signal: time-to-reset <60min + quota >50% remaining + no blocking work
5 autoresearch profiles: audit / pattern-mining / alignment scan / schema rollout / cross-reference gap
Hard constraints: read-only, gitignored output, ≤2h cumulative, no charter/secret/destructive
Orchestrator discipline: claude tokens << GLM flat-fee delta
Karpathy tone: success criterion in 1 sentence = decidable; otherwise not autoresearch-ready

Scope (per charter §3)

Small impact: docs only, single file, +53 / -1 lines, no behavior change, no charter touch. Pre-authorized by operator for self-pushed small follow-up PRs (per their 2026-05-02 ~03:08 CEST message: "z powodu zakresu będziesz mógł sam wypchnąć").

Real-world precedent (already happened)

Operator fired GLM modules audit during this very session (~02:46 CEST) — 81 manifests classified, 100% schema-valid finding, 10-min wallclock on z.ai sub. Pattern works. §11 codifies the muscle memory.

Co-authors

Operator (pattern designer, source Polish quote in §11 body)
Claude Opus 4.7 (1M context, this session)

## Summary Encodes operator's overnight delegation pattern (2026-05-02 ~03:00 CEST) as a self-instruction section §11 in `state/CLAUDE_SELF_INSTRUCTIONS.md`. Future Claude sessions detect quota-window-resets and fire autoresearch tasks to drain unused subscription budget on useful background work. ## Why Operator pays flat-fee subs (Anthropic, OpenAI, z.ai). Unused quota at window reset = wasted money. Karpathy autoresearch pattern (github.com/karpathy/autoresearch) — recursive LLM-driven research workflow — fits naturally as "drain budget on decidable questions." ## Section §11 covers - Detection signal: time-to-reset <60min + quota >50% remaining + no blocking work - 5 autoresearch profiles: audit / pattern-mining / alignment scan / schema rollout / cross-reference gap - Hard constraints: read-only, gitignored output, ≤2h cumulative, no charter/secret/destructive - Orchestrator discipline: claude tokens << GLM flat-fee delta - Karpathy tone: success criterion in 1 sentence = decidable; otherwise not autoresearch-ready ## Scope (per charter §3) **Small impact**: docs only, single file, +53 / -1 lines, no behavior change, no charter touch. Pre-authorized by operator for self-pushed small follow-up PRs (per their 2026-05-02 ~03:08 CEST message: "z powodu zakresu będziesz mógł sam wypchnąć"). ## Real-world precedent (already happened) Operator fired GLM modules audit during this very session (~02:46 CEST) — 81 manifests classified, 100% schema-valid finding, 10-min wallclock on z.ai sub. Pattern works. §11 codifies the muscle memory. ## Co-authors - Operator (pattern designer, source Polish quote in §11 body) - Claude Opus 4.7 (1M context, this session)

claude added the

ready-for-operator

label

2026-05-02 03:15:33 +02:00

pdurlej was assigned by claude

2026-05-02 03:15:33 +02:00

claude added 1 commit

2026-05-02 03:15:33 +02:00

docs(orchestrator): add §11 token-arbitrage at subscription level (Karpathy autoresearch loop) a9e64757cb

Per operator's request 2026-05-02 ~03:00 CEST: encode the pattern of
"when ~1h before subscription quota window resets and >50% headroom unused,
fire autoresearch-style background tasks to drain the budget on useful work."

Operator quote (verbatim, Polish):
"też loop na przyszłość - okazonalnie jeśli jest godzina przed wyczerpaniem
limitu palić wszystko na auto research karpathego z subskrybcji czy to openai,
anthropic czy z.ai. niestety jest to dla mnie opłacalne"

Section §11 covers:
- Detection signal (time-to-reset <60min + quota >50% remaining + no blocking work)
- 5 good autoresearch task profiles (audit, pattern-mining, alignment scan,
  schema rollout candidate, cross-reference gap)
- Hard constraints (read-only, .sisyphus/research/ output, ≤2h cumulative budget,
  no charter/secret/destructive)
- Orchestrator role (fire background, don't poll, claude tokens << GLM flat-fee)
- Karpathy autoresearch tone (decidable success criterion)

Reference: github.com/karpathy/autoresearch (recursive LLM-driven research workflow).

Per charter §3 review-by-impact: small impact (docs only, single file, +60 lines,
no behavior change, no charter touch) → ensemble review optional. Pre-authorized
by operator for self-pushed small follow-up PRs.

Co-authored-by: Operator (pattern designer, source quote)
Co-authored-by: Claude Opus 4.7 (1M context)

pdurlej commented

2026-05-02 09:09:38 +02:00

Owner

ok to merge @claude

i can’t do it from my mobile (i don’t know why)

ok to merge @claude i can’t do it from my mobile (i don’t know why)

pdurlej approved these changes

2026-05-02 09:11:25 +02:00

pdurlej merged commit b9ef41a227 into main

2026-05-02 09:21:44 +02:00

pdurlej referenced this pull request from a commit

2026-05-02 09:21:44 +02:00