L3 · Execution
Watch the work move and the outputs appear.
Pick a line and press play. Each step names the agent that holds it, its autonomy level, the gate that can stop it and the artifact it produces. Gates halt the run until a human releases them.
Steps simulated
98
Handoffs
91
Artifact classes
12
Gate stops
15
Choose the situation
The situation
A corporate position is adopted centrally and has to land in every operating market, in language, in the cascade window — and be measured where it did not land.
Trigger
Position ratified at G1
Clock
Hours to days — measured by cascade completion, not by send
Ceiling rationale
A3. Scanning, synthesis, position drafting and submission preparation run unattended. A filing that binds the company is signed by a person.
Run state
Steps done
0/17
Artifacts
0
Gates cleared
0
Human touches
0
Canon stores written
0
Overrides
0
Autonomous so far
—
Canon change committed at G1; change cascade opens as a first-class job with a named owner
Identify every affected artifact — live drafts, scheduled content, standing Q&A, market statements, briefing packs, prepared holding language
Identify every affected audience cohort across all three segments and their channel map
Retrieve per-market constraints — regulatory, works council, language, prior local commitments, live local issues
Generate audience-specific artifacts: corporate staff, market leadership, franchise owners, general managers, crew — each in the register that cohort actually reads
Localize across ~40 languages with a divergence report per market stating exactly what changed from source and why
Check every localized variant against the Canon for contradiction — emphasis permitted, contradiction blocked and named
Verify claims per market against Evidence Ledger jurisdictional scope; flag anything unsubstantiated in that market
G2 TRUTH — fires only on failure. Claim owner substantiates, amends or strikes. Silent when the ledger is clean.
Build the sequence — who hears what, when, through which channel, from which messenger; franchise partners ahead of external
Deconflict against the corporate calendar in every market — local holidays, local earnings, local launches, local sensitivities
G5 RELEASE — market leads confirm in parallel, not in series. Below the materiality floor this is one click.
Execute distribution across every channel and market simultaneously
Measure reach and comprehension per cohort — did franchise owners in Brazil actually receive and understand this?
Detect non-landing — cohorts where reach or comprehension fell below threshold — and auto-open a remediation job
Monitor for post-cascade divergence — markets saying something inconsistent in the following two weeks
Write back — cascade performance by cohort and channel, so the next cascade routes better
Output shelf
0Everything the run has actually produced. Click an item to read the output contract the producing agent is held to.
Nothing produced yet. Run the situation.
Canon touched
No write-backs yet.
Architectural rule R2: no output leaves the system without a Canon write-back. That is the difference between a tool that produces content and an organization that gets smarter.
Output taxonomy
What a step actually produces
Every specified step leaves an output. Grouping them into classes shows what was produced, by whom, under which level, and what it changed.
The work reached the audience it was built for.
Publishable language in the house voice.
The organization is permanently smarter for it.
A judgment with its drivers attached.
What the company says, and why it is defensible.
The same position, landed in-market.
A position of record with an external body.
A claim is substantiated and traceable.
What it changed, counted rather than claimed.
A named human accepted the consequence.
Something in the world was noticed and named.
A human is prepared to act or answer.
Why the bench behaves this way
Three rules govern every run
The simulation obeys the same three architectural constraints the platform is specified under.
R1
No agent-to-agent coordination. Every handoff goes through the Chief of Staff.
Swarms that self-coordinate cannot be audited, debugged or governed. The cost is a small latency penalty; the benefit is that every decision has a traceable path.
R2
No output leaves the system without a Canon write-back.
This is what compounds. Without it, the platform is a very expensive drafting tool.
R3
The autonomy ceiling is data, not code.
Legal must be able to ratchet down a ceiling at 23:00 on a Friday without a deployment.