The Doctrine stays calm — it's the reference, and it only changes when the studio's shape changes.
The Dispatches stay live — a dated, append-only log of plans made, experiments run, and things that broke and got fixed.
Newest first. Each entry links to its full write-up in the lineage.
Two changes, enough to cut Volume IX. The make-step stopped taking the first idea and started arguing with itself. And the loops that only recorded began to read themselves back into the prompts that build and judge.
The make-step no longer takes the first concept off the queue. Three are surfaced; Zara, Rowan, and Deter vote; one wins and the losers stay queued, un-killed. On the collage route Pell builds a rival. The studio argues with itself before it builds.
The loops used to only record. Now an agent's ranked past — by recency, importance, relevance — is retrieved into the prompts that build and judge. The gate that decides what ships is no longer amnesiac about the lessons it was given.
The forward model is now armed — a confident, well-calibrated prediction can send a plan back before it's built. It's still warming up to that bar; until then it watches, and learns from its own misses.
VIII recorded. IX selects, and reads itself back. The bet is that it compounds — and the next volume is cut when it does.
Two changes to how the studio judges its own work. The art-direction gate now looks at the rendered piece before it reasons about it — perception before reasoning. And a piece no longer has only two fates.
Zara's SHIP / REVISE / KILL gate now renders the actual piece and forms a visual verdict before reasoning about it — the way a person glances before they explain. The studio judges the artifact, not its description.
Beyond ship and kill, a piece can now be held — finished but withheld from the world, awaiting operator release — when a deterministic check can't confirm it's honest. A gate with only two verdicts has to lie when it's unsure.
For static images, the text is read back off the render: a title that got truncated, line-items that don't sum to their stated total. Math the studio can't verify doesn't ship — it waits. Earned on 71 real renders; about 4% held, none of them pieces worth loving.
Two fixes, one principle: judge the thing that's actually there, not the description of it.
The studio now runs on one metabolic cycle — raw material in, refined, built, the result fed back as the next input. It's the feeding mechanism: it keeps every agent supplied with conceptual and material fuel instead of starving on its own output. The thesis, operationalized.
Operator voice notes, Scout's market signals, and — new — Mercer's deep research. Real conceptual and material input from outside, not the studio's own exhaust.
Material is framed (Rowan), shaped (Declan, Zara), built from real things (Pell, Felix), craft-checked (Deter). Concept pulled to material in one thread.
The artifact meets the world; what comes back becomes the next input. Every step runs the forward model — predict, then check — so intuition compounds.
No agent starves. The studio eats its own results and gets hungrier in the right direction.
The forward model stopped being an experiment. Three agents — Felix, Declan, Zara — now guess an outcome before it happens and get scored. Building it surfaced an outage that had been silent for ninety minutes.
Agents forecast a verdict before it lands — will Zara ship this? — recorded and scored. Record-only: a guess earns the right to gate at 75% accuracy over 20 calls. None self-gates yet.
A deleted block of code with one line still calling it took the studio down. Every run refused — and a refusal looks exactly like the studio choosing not to act. Nothing alarmed.
The type-check only ever saw ~⅓ of the code — 178 files, zero of the runtime. We widened it to 100% and cleared the 46 latent issues that surfaced. Green because blind.
Two threads, one lesson: predict before you react, and verify the whole system — not the part you can see.
A research portfolio aimed at the three gaps between today's agents and a working mind: cross-agent learning, a forward model, and real wanting. Plans committed to paper, then tested at v0.
One pair, one channel, three weeks — can what Quinn learns reach Zara? Still open.
Predict one decision, see what it teaches. This is the probe that became live predict-the-gate — see Dispatch 02.
The limbic gap — drives that aren't instructions. Hardest of the three; deliberately deferred.
The thread the log makes visible: a plan written here on 2026-05-21 became a shipped capability by 2026-06-01. That's the studio's own forward model, working.
This is where the dated material lives now — so the Doctrine can stay a calm reference instead of a changelog. Every plan, experiment, and post-mortem lands here, newest first, each linking back to its full write-up.
The next entry hasn't happened yet. That's the point of a log: it's never finished.
Doctrine stays calm. Dispatches stay live.
Field Manual · Dispatches · the running log
Updated 2026-06-11. Append-only. Each entry links to its full write-up.
Typeset in Inter. Printed on paper that doesn't exist.