The Pulse

Every number on this page was measured on 2026-08-10 by the stack’s own tooling: git, the scheduler, the logs, the issue tracker. Each figure was then re-derived by an independent second pass before publication. Nothing here is an estimate. The figures are a dated snapshot, and the exact command behind every one of them is kept in the operator registry.

Source: the record

The site renders from a record, not from memory. The record is a Markdown vault of 32,962 files, 29,125 of them canon extracts under a published schema, curated into 30 project entities carrying 237,813 words of canonical truth. The operator registry adds 5,327 documents and 805,164 words of decisions, session logs, and mined findings. Assets live in an object store: 2,751 files, about 2 GB.

Process: the machine

The pipeline is 192 Python scripts, 54 of them tests, organized behind 36 skills, 12 scheduled tasks, two supervised services, 3 MCP servers, and 6 session hooks. The point of the inventory is not size. It is that every piece is versioned, and the whole machine can be rebuilt from the repository.

The stack commits its own work. Of the 718 commits on the core repo in the last thirty days, most were landed by an hourly task that drains the session-mining queue: 675 completed runs so far, and 3,321 mined notes now sitting in the registry. Nobody remembers to persist a decision. The scheduler does.

Work moves through an issue tracker like any other engineering org: 412 issues on the core repo, 382 closed, 30 open. May is zero because May belonged to other repos in the org. August is ten days old at measurement, which is also why it wears an asterisk.

Since late April the primary agent surface has processed 14.6B tokens. The honest decomposition: 13.9B of those were cache reads, which is the price of giving agents long memories, and 103.8M were actually generated. Not all inference leaves the building, either: 17 models totalling 160.6 GB sit on local disk, and session transcripts are mined by a local 14B model at about 31 tokens per second on a 20 GB RTX 4000 Ada, instead of being shipped to a cloud model.

Site: the render

This website is the smallest layer of the system: 645 commits since December 2024, 920 tracked files, 52,560 lines across the five languages that build it. The output is 121 project records and 81,161 published words, every page checked against a published bar at build time.

42 records carry full teardowns: mechanism, constraints, decisions, scars. 54 are deliberately small, a complete record at reduced depth. 25 were cut and stay unpublished. The floor is held to the same bar as the ceiling, by a build gate rather than by memory.

How these numbers are made

• Every figure came from a command, not a recollection: git for commits and line counts, the GitHub API for issue flow, the Windows scheduler for tasks, the drain log for mining runs, the transcript archive for tokens. A second, independent pass re-derived each headline number before this page shipped.
• Commit counts include the machine’s own commits. That is disclosed because it is the design, not an inflation.
• Token totals cover the primary agent surface only. Other surfaces would add to the total, not subtract from it.
• This is a dated snapshot (2026-08-10). The corpus moves daily; the page states its date instead of pretending to be live.

The machine that produces these numbers is documented in EN-OS, and the pipeline that renders this page from the record is the same one being measured.