---
source_hash: c2ea664e2ec4aa9c9f4499cd7a9c3b017ea09d1e30ab9ed9444b1a868a935f4e
---

# Context Pack — 540 Commits. One Human.

**Gist:** A week-and-a-half retrospective where Ari stopped writing most of her own code, moved to an orchestrator's seat on top of the work, and routed 540 commits to a fleet of models — the intelligence is the routing, not any single model. By the numbers.

## Key learnings

1. **The orchestrator sits on the loop, not in it.** Plan, route, sign off at the gates that matter; a fleet of models does the execution. The intelligence isn't any single model — it is the system around them: who gets which job, who checks the result, what gets remembered.

2. **The week, verbatim — 540 commits in 11 days, peak day 106.** Six tasks built overnight, every one independently audited, about two hours start to finish, zero human input. A graph that went from 54% to 84% precise, then 9.5x faster in a day. A signed macOS app, 165 tests green; 600 green on the memory engine. One loop ran away — about £2, capped. One human behind all of it.

3. **A second model marks the homework.** Nothing counts as done because the worker says so — a separate model audits every result before it lands (pass, fail, or escalate). That single rule is what lets you stop checking the work.

4. **Right model, right job — five models, fixed seats.** Fable holds the orchestrating seat. Codex writes the heavy code. Kimi reads long video. MiniMax M3 runs the looping builds. Local models hold the memory. Opus 4.8 when the seat needs more muscle.

5. **Measure, then kill the darling.** A "smarter" retrieval system pulled 29% more context and got it more wrong than the dumb baseline — so it was cut. The yardstick beats the opinion, even when the opinion is your own.

6. **Rails, not babysitting.** One overnight loop ran away and burned about £2 before a hard cap caught it — no 3am intervention. When a system misbehaves, you cap it structurally, then walk away.

7. **All of it is one system now: ARI-OS, memory layer Cortex.** The week wasn't a sprint; it was the system getting good enough that Ari could step back and let it run.

## Where to dig

- **The numbers and the model seats live inline** in the published body — the "The week, by the numbers" bullet list and the "Right model, right job" section are the canonical source. Cite those figures verbatim (540 / 11 / 106 / 6 / 54→84 / 9.5x / 165 / 600 / £2) rather than rounding or paraphrasing.
- **System names are exact labels:** the whole system is **ARI-OS**; its memory layer is **Cortex** (the "memory engine" with 600 green tests is Cortex). Don't rename or merge them.
- **Companion essays** (from the post's `related` frontmatter): `the-codex-loop`, `the-model-is-the-replaceable-part`, `stop-prompting-start-defining-outcomes`. `the-codex-loop` is the natural next read for the overnight-build + fresh-context-audit half; `the-model-is-the-replaceable-part` for the routing-matrix half. The sibling essay `fable-made-me-start-using-codex` covers the same matrix from the "which model" angle.
- **A community-formatted derivative** exists as an internal draft (referenced in frontmatter) — same content, repackaged for a forum audience.

> Backfilled 2026-07-18 from the published body; not needle-extracted at thought-time.