Running the House — a day in the life of a brigade

Part 3 of 3 — operations, for delivery teams. Assumes Parts 1-2.

The operator lifecycle

mise            → readiness table; fix every FAIL (each row prints its remedy)
service start   → lock → mise gate → walk the rail until it's dry
  (tickets flow: pull → Gate A → phase-0 → stations → critics advise → expo decides → ack → file)
service end     → stand down between tickets; work-logs hold the resume state
fire <order>    → out-of-band "now" — the expo wraps the order in a ticket itself; same gates, same record, no queue
close-out sweep → steward scans filed tickets for unsigned terminals, delivers, signs

Two habits that keep the house honest:

• Never skip mise. It exists because "it worked on my machine" is a statement about your machine. Readiness is declared per brigade and checked deterministically — a FAIL means the brigade refuses service, by contract.
• End between tickets. service end is graceful by design; killing a walker mid-ticket leaves a leased ticket for the lease-reclaim path to mop up. The work-log means a resumed walk knows exactly where the last one stood.

The nine steps (the mental model)

Every order, regardless of domain, walks the same arc:

1. A requester brings a need to the steward
2. The steward gathers context from the cellar and pairs the need to a brigade's menu
3. The steward writes a ticket and hangs it on the rail
4. The expo pulls it (lease), runs Gate A and the phase-0 sufficiency judgment
5. Stations execute, marking up the ticket as they go
6. Critics evaluate and advise
7. The expo decides the exit
8. Approved artifacts land in the cellar; the ticket files to its subject
9. Close-out: completion flows back to the requester, and the record is signed

Status honesty, from the live audits: steps 2-8 have run repeatedly on real engagements. Step 1 currently enters as free text — the structured intake surface (the registrar's Jira/form front door) is designed, not built. Step 9's sweep is written and mechanized but has the youngest live track record. Teach the arc; know where it's still soft.

A real run, end to end

One recent engagement exercised the entire arc in a single day (identifying details omitted; the filed tickets are the full record):

• Intake & research. A company-research ticket fanned out scraping and filings stations; when the target's job platform wasn't covered by an existing scraper, the station built one mid-serve rather than shipping a coverage hole — and recorded the partial-coverage caveat honestly on the ticket.
• Assessment. The assessment brigade walked its phase chain with per-phase contract gates. Inside one phase, use-cases became micro-tickets on an inner rail — parallel evaluation at swarm scale (~90 sub-agent lens executions) that never touched the deciding agent's context window.
• Governance, visibly working. At the highest-stakes gate, the multi-judge council vetoed the draft twice for overselling — refires consumed exactly the budget, fixes were independently re-verified, and only then did the manifest advance. Later, in collateral production, an expectation-review gate vetoed a draft for dodging the client's actual signal with boilerplate; the refire produced substantively better output. This is the pattern's core claim demonstrated: the critics could not be argued with, and the expo could not be bypassed.
• Human gate, recorded. The one human approval in the chain was recorded on the ticket with explicit delegation provenance — the walk never stalled, and the record shows exactly who approved what, under what authority.
• Landing. The committed manifest ingested into the marketplace website with zero code changes; collateral kinds passed their deterministic invariant gates on the first fire; every ticket filed, every close-out signed, rail left empty.

Failure lessons worth teaching (they're the curriculum)

The pattern's best arguments are its recorded failures:

• The units bug. A schema-valid manifest sailed through every gate with dollar figures 1000× oversold — because "no existing gate asserts anything about units." Caught by a human eyeball on a rendered page. The fix was not "look harder": it was a new deterministic cross-check gate comparing the manifest against its upstream value model, plus a runtime-verification critic axis (run the artifact, look at the rendered output — type-checks can't see a wrong number). Lesson: every gate list is a hypothesis; every miss becomes a new gate.
• Grader fragility. An execution-eval run initially showed a skill failing — the grader was string-matching answer names too strictly. The false negative was caught by reading the transcripts before reporting, and the grader was rebuilt on fixed output schemas. Lesson: verify the verifier; never report a grade you haven't traced to the artifact.
• Aggregate lift lies. A skill that fixed one real failure mode showed ~0 aggregate lift because three already-easy fixtures diluted it. The expo now consumes per-fixture lift. Lesson: averages hide exactly the signal you built the eval to find.
• The self-review trap. Early runs let the working agent apply critic axes to its own output. The independent zero-context critic that replaced it immediately found a design smell nobody else had caught. Lesson: independence is the ingredient; a self-applied checklist is not a critic.

Where it runs

A brigade ships as a self-contained plugin — vendored adapters hash-stamped against canon (drift is detectable by mise), a mise declaration, a service wrapper, its stations. The reference deployment is local-first — one operator, filesystem rail and cellar, everything on one machine — and the same house has already published its outputs into Snowflake (content ingest + an App Runtime-hosted marketplace site) without the brigades themselves leaving the box. Multi-operator and client-governed profiles change only the adapters behind the rail/cellar ports; the brigades, the commands, and the contracts don't change at all. That is the ports-and-adapters ("hexagonal") bet paying rent.

Provenance: distilled from the working house's own contract docs (BRIGADE-INTERFACE, TICKET-CONTRACT, RAIL/CELLAR/MENU specs), its plugin READMEs, and the audited receipts of live runs. Where this pack says "designed, not proven," that grading comes from those audits, not modesty.