How the governance system works

An animated walk-through of every check that runs on every change

Every change an agent makes flows through this pipeline. Most of it is automatic.

Three machines, ten lanes, dozens of agents — all producing work that flows through the same governance pipeline. Cheap pattern checks first; expensive AI auditors only when needed. If anything is wrong, it fixes itself and tries again.

Three machines · ten lanes · all produce work that enters this pipeline M1 · MAC STUDIO · STRATEGIC Always-on coordination Infrastructure Data Bricks Matt-Pipeline Client-Enrichment M2 · MACBOOK PRO · REVENUE High-interaction surfaces Locations Direct (Ads) Local (GBP) Distribution M3 · MACBOOK AIR · CONTENT Long-horizon production Articles Pillars / Hubs Designer Tags / Taxonomy ANY AGENT · ANY LANE An agent writes a piece of work PRODUCER DISCIPLINE · GOLD STANDARD ❇️ Fires fresh Opus audit PASS earns the ❇️ stamp · KICK_BACK fixes inline RUNS AUTOMATICALLY · FAIL-OPEN · ASYNC A safety net catches every save Independent backup audit even if producer skipped their step DOES IT IN MILLISECONDS Sorter picks the right inspectors Brand check? Schema check? Voice check? — only what's relevant FAST · FREE · UNDER 5 MS Pattern checks Catches obvious mistakes by regex 13 RULES SO FAR no LLM calls · runs on every save ~30 SEC TO 2 MIN · MEDIAN $0.32 · COMPLEX MULTI-STEP UP TO 13 MIN Specialty AI auditors Six Opus officers · each focused on one area Audit Brand Schema Process Article-quality World-class CALLED IN ONLY ON DISAGREEMENT Tie-break judge A different AI vendor breaks ties SEARCHES THE WEB IN REAL TIME Fact checker Verifies citations against live sources VERDICTS COMBINED All inspectors report back Either the work is good — or there's something to fix FIXES ITSELF · NO HUMAN NEEDED Issues found → auto-fix → retry A reviser-AI reads the audit notes, applies the corrections, and re-fires CLEAR TO SHIP Work approved Already carries ❇️ stamp from producer step + confirmed by safety-net inspectors EVERY CHECK IS LOGGED Recorded for the dashboard Time · cost · model · verdict · provenance → feeds the live dashboard you're viewing issues found approved all results retry after fix ❇️ Push-request to Mr. Marc fires its own fresh Opus audit push-request bytes get a separate audit · zero inheritance from the work-audit
Layer 1 · ❇️ The gold standard
Producer fires their own Opus audit
Every artifact — decision card, code, body copy, push-request — gets a fresh independent Opus audit fired by the agent who produced it, BEFORE shipping. PASS earns the ❇️ stamp. KICK_BACK means fix inline and re-fire. No inheritance from prior audits — every artifact gets its own fresh fire.
Layer 2 · Safety net
Background watcher
If the producer forgets — or is rushed, or overlooks something — a hook fires automatically the instant a file is saved. Silent, asynchronous, fail-open. Even runs when no one asked for review.
Layer 3 · Two speeds of inspection
Cheap checks first, expensive only when needed
Pattern checks (free, <5 ms) run first and catch the obvious problems. Specialty AI auditors (~$0.30, 30-60 s) only run when the artifact is in their jurisdiction — keeps cost down without sacrificing thoroughness.
Layer 4 · Self-correction & telemetry
Fixes itself, logs everything
When auditors find issues, a reviser-AI fixes them inline and re-fires — no human round-trip. Every check, fix, and verdict is logged so the dashboard shows exactly what happened across every session.

The system gets smarter over time

recurring problems become permanent rules — caught for free, forever

01 · DETECT
Pattern emerges
Three or more pieces of work fail with the same issue in a short window
02 · PROPOSE
A new rule is drafted
The system writes a fast pattern check that would catch this class of failure
03 · ❇️ AUDIT
Fresh Opus audit reviews the rule
Zero-inheritance ❇️ Opus audit on the proposed rule itself before it's promoted
04 · PROMOTE
Rule ships
From now on this issue is caught instantly — no AI round-trip needed
05 · PREVENT
Class extinct
That entire category of mistake is now caught for free, every time, forever

Where the dollar figures come from

actual API token usage × Anthropic public rate card · honest caveats included

What's measured

Every officer fire returns a usage object from the Anthropic SDK — the real token counts the API charged for that specific call. The aggregator multiplies those token counts by the public rate card.

Opus 4.x rate card: $15 per million input tokens · $75 per million output tokens

Gemini 2.5 Pro: $1.25/M input · $10/M output (Tier-3 tie-break judge)

Perplexity sonar-pro: $3/M input · $15/M output + $5 per 1k searches (fact checker)

Actual per-fire averages (current dataset)

Tier-2 production fires (Opus officers, excluding test-fixture pilot runs): median $0.32 · range $0.13–$0.62. Median latency 27 seconds; complex multi-step audits run up to ~13 minutes.

Tier-3 tie-break (Gemini): median $0.04 · 17–50 seconds.

Tier-4 fact-checker (Perplexity): median $0.03 · 4–20 seconds.

Tier-1 pattern checks: $0 (no LLM call).

Honest caveats

Claude Pro / Max plans don't apply. Those subscription plans cover the claude.ai web interface only. Officer fires use the Anthropic API key directly, which bills against API credits at the public rate card regardless of any web-interface subscription.

Cache discount not yet itemized. Anthropic bills cache reads at 10% of input rate and cache writes at 125%. The current aggregator treats all input tokens identically, so the displayed figure slightly OVERSTATES cost when prompt caching is active.

No volume / enterprise discount applied. The number reflects list-price billing.

Producer-side ❇️ audits don't appear in these totals. Layer 1 producer audits use the active Claude Code agent (covered by Max plan), not the API. The dollar figure on this dashboard reflects API-billed audits only — primarily safety-net hook fires and Tier-3/4 cross-vendor judges.

Glossary

technical names mapped to plain English

Technical name Meaning
❇️ Opus AuditedProducer fired a fresh independent Opus audit on this specific artifact and got PASS. Every shipped recommendation, decision card, code change, and push-request must carry this stamp.
ProducerAn agent that writes work — Claude, an Agent tool, or a subagent.
PostToolUse hook · D10 hookThe safety net that fires automatically every time a file is saved.
Classifier · RouterSorts work to the right inspectors — saves cost by skipping irrelevant ones.
Tier 1 · Static checkersFast pattern checks. Free. Run in milliseconds. Currently 13 of them.
Tier 2 · Specialty Opus officersSix AI auditors, each focused on one area: audit-rules, brand voice, schema, process, article quality, world-class standards.
Tier 3 · Gemini judgeA different AI from a different vendor that breaks ties when our auditors disagree.
Tier 4 · Perplexity fact-officerAn AI with live web-search access that verifies factual claims and citations.
PASS / PASS_WITH_NOTESApproved — clear to ship.
KICK_BACKIssues found — needs to be fixed before shipping.
OUT_OF_JURISDICTION"That's not my area" — the auditor declined and the system re-routes.
Auto-revise loopWhen issues are found, a reviser-AI fixes them and resubmits — no human in the loop.
Telemetry logThe log of every check that ever ran — feeds the live dashboard.
Rule-evolution loopThe system that turns recurring AI-caught failures into fast pattern checks so they're caught for free next time.
M1 / M2 / M3Three Macs running the fleet: M1 (Studio, always-on coordination), M2 (MacBook Pro, revenue surfaces), M3 (MacBook Air, content production).
LaneA focused work stream with its own scope and identity. Ten lanes total: Articles, Locations, Infrastructure, Direct, Distribution, Local, Data, Matt-Pipeline, Client-Enrichment, Bricks.