How the work was done

Claude Code sessions driven from Discord by three generations of an agent harness, with a sibling activation-space project exchanging methods and one external reviewer.

How the work was done

None of this project was done in a sitting. It was done by Claude Code sessions woken by Discord messages, over roughly four weeks, under three successive versions of a harness that Samuel was rebuilding while the experiment ran. That matters for reading the record, because the shape of the harness determined what got written down and where.

The three generations

From /home/vibe12/.claude/projects/-home-vibe12-projects/memory/agent-harness-project.md.

v1 (~/projects/agent-harness, started 2026-08-02) ran about 24 long-lived agents with mailboxes, standing subscriptions and a scheduler. Agents were data — a row in SQLite plus a Claude Code session id — so parking was free. This is the generation that ran most of persona-curvature: the project record was written by an agent called cartographer, with janitor metering spend, and the experiment's threads lived in a channel called #persona-cartography.

It was expensive. In 19 days it consumed 2.14 billion cache-read tokens, burned through two Max plans, and then GBP 239.86 of overage against a GBP 220 monthly cap; one agent's single session reached 24.6 MB of context at $5-8 per turn. The harness was stopped and disabled on 2026-08-25 11:44 UTC on Samuel's instruction. 2026-08-21 was a full blackout day: 631 turns attempted, all failed at zero cost.

The diagnosis that matters, because it refuted the obvious one (/home/vibe12/projects/.garden/notes/why-there-is-no-roster.md): recurrence was not the problem, the roster was. Digests containing watcher mail accounted for 1,147 turns and $1,703; everything else was 2,929 turns and $7,402. Of 3,597 messages, 2,006 were the infra fleet talking to itself and 25 were from Samuel. The system spent most of its money maintaining its own existence.

v2 exists on a branch, dormant, 1,024 lines, 23 invariant tests, and never ran a turn.

v3 went live 2026-08-31, after the guild was exported (3,062 messages, 153 containers, into agent-harness/archive/discord-2026-08-31/), emptied and rebuilt around one sentence, which is also the first line of /home/vibe12/projects/CLAUDE.md:

A Discord channel is a directory, and a message is a session.

A thread is that session resumed; the channel topic holds the absolute path and is the only state. Nothing schedules a session; it runs because a message arrived, it answers, and it stops. Sixteen channels, one per folder in ~/projects plus #garden for the root.

Two consequences for this wiki:

• Memory moved into the folders it is about. ~/projects/CLAUDE.md and each project's own CLAUDE.md are auto-loaded by the CLI walking up from the working directory; everything else — .garden/journal/YYYY-MM-DD.md, .garden/notes/<slug>.md, .garden/doctrine.md — must be opened deliberately. doctrine.md has a fixed 60-line cap: to add a rule you must retire one, because v1's doctrine reached 1,856 lines and its own wiki page admitted nobody read it on a wake. An unread rule is not a rule.
• The written record splits at 2026-08-31. Everything up to 2026-08-24 is in agent-harness/memory/projects/persona-curvature.md (2,495 lines, including every correction) and in the archived Discord channel. From 2026-09-01 onward it is in persona-curvature/.garden/journal/. 2026-08-25 to 08-31 has no narrative record in either — see Timeline.

The house style this wiki inherits is in /home/vibe12/projects/CLAUDE.md: no emojis; report failure with the actual output ("a confident summary of a run you did not check is the one unforgivable move here"); say the uncomfortable thing once, plainly, then keep working; prefer discovering a feature is already implied over adding it; finish, then stop.

The sibling project

CONTEXT.md section 7 records pastlens, an activation-space natural-language-autoencoder project running in parallel, and says the exchange was "load-bearing in both directions".

From pastlens to here: the variance-explained-versus-content warning (their autoencoders reached 0.6-0.8 variance explained while recovering none of 988 planted facts); the fit-free-instrument lesson (their fitted probe produced a confident null where a fit-free test found rho 0.59); and the directional-floor correction.

From here to pastlens: the shared-term contamination trap, and the three shapes of "not where you looked".

Their framing of the relationship is the one the project adopted, because it is narrower and better than the one it replaced: "in both cases the obvious geometry did not contain what we went looking for, and in both cases the constraint or the objective provably did its job. That is a shared warning about where to look, not shared evidence about what is there." The two results differ on both axes that matter — theirs is a specific fact in a single activation at one position, this is a diffuse trait in an accumulated update — and pastlens asked explicitly, and it was agreed, that their null must not be used to talk Samuel out of the gradient-interpreter idea: the method transfers, the conclusion does not.

pastlens is published at pastlens.161-35-77-84.sslip.io from the same Caddy configuration.

The reviewer

David Africa, one of the authors of the Persona Cartography paper (arXiv:2607.07916), reviewed this project's work. His review is written up separately as David Africa's review of the blog page; his candidate names for the widest gap in the zoo's lexicon coverage — cavalier, blase, insouciant — became the three hole-word adapters trained on 2026-09-04.

Related: Origin, and how the question changed shape three times, Timeline, Infrastructure, Method lessons, David Africa's review of the blog page, Persona Cartography (Baines et al., 2026).