---
source_hash: 8bbb3864361cac430415f6fe028b492e5a839ba7470ae3bd7a6f94e3c1725a90
post: /posts/doctrines/the-output
shape: doctrinal
---

# Context Pack: The Output

Reference for the rewritten July 2026 account. Reported figures are historical; the essay is not an independently reproduced benchmark. Only the quoted excerpts below are verbatim.

## Verbatim excerpts

> Starts are not wins, and the account did not supply a clean total of jobs that landed.

> The typing moved out of the chair; the judgement stayed with me.

> The useful question was what the system could do at the end that it couldn't do at the start.

## Argument and authorship

Output includes useful work advanced before public release. The author disclosed AI-assisted prose and dyslexia: the model helped shape sentences while she held the idea. The strongest seat planned and routed bounded work, with human judgement retained. The reusable investment is contracts, routing and checks that can survive a model replacement, not an assumption that all models are equally capable.

## Keep the numbers distinct

The week reported zero published posts and 715 commits: about 328 feature, 61 tests, 50 fixes and about 135 chores, leaving roughly 141 merges/docs/other. The old approximately 440 feature-floor label actually groups feature, test and fix work (328+61+50=439). Do not call it 440 feature commits. The retained receipt image carries the old label; its caption and alt explain the grouping.

About 640 background jobs STARTED, with retries, startup failures and returned work. A completed-job total is not supplied. The ordering task went from one of five to five of five between Monday and Friday through a self-teaching process; that observed week does not prove permanent competence.

Other receipts: a phone photo editor first booted; a desk cockpit shed a two-second keystroke lag; an 8,943-line depth pipeline merged; local image generation became available; the publishing system was rebuilt as a compiled browser engine. The five per cent public / ninety-five per cent unseen division is a metaphor, not a measured time allocation.

## Reported routing

Claude Code was the harness, UltraCode the parallel-work mechanism, and Fable the initial conductor. Lanes were Claude for prose, Codex for builds, GLM for long cheaper grinds and MiniMax for vision. Local third-party routing is the author's setup, not a statement of universal official provider compatibility. The body links Anthropic's dynamic-workflow documentation for the fan-out feature.

The conductor was later changed to GLM plus a cheaper coding model and work continued. Each substitution still needs evaluation against its task. Cost includes failed and repeated calls. The ConductorFleet widget is an illustration of assignments, not model execution or throughput evidence.

A square means a small useful advance: a commit, a new capability or a removed constraint. Counts need evidence of what changed, not just activity.

## Retained references

- Source: content/posts/doctrines/the-output.mdx.
- Three images under /posts/doctrines/: the-output-receipt.png, the-output-routing-moat.png and the-output-squares-vs-posts.png.
- Two AsciiDoodle instances, conductor-baton and swarm-loop; ConductorFleet exported from apps/web/components/blocks.