CHORUS
ELEVATED

L3 · The Gates

A gate is a decision, not a review. A gate-holder who returns 'looks good' has not held the gate.

Twenty of ninety-eight workflow steps route to a human; fourteen of those are gate steps. Everything here makes those fourteen decisions fast, evidenced and auditable.

Gates

5

Firings across WF1–WF5

16

Default on timeout

Escalate · never auto-approve

Live posture

ELEVATED

Live console

The queue

Each card carries one question, one recommendation, one named alternative, the evidence and a deadline scaled to posture.

see where gates fire

Gates pending

0

awaiting a named human

Past deadline

0

escalates one level automatically

Decisions recorded

0

written to the Precedent Library

Recommendation overrides

0

each one is training data

The queue is readable by anyone. Deciding is not. A gate-holder is a named human — sign in to hold one.

reading queue…

Posture declaration

sign-in required

Topic-scoped ceiling drops one level. Sensing cadence moves to continuous on the affected topic.

G2 substantiation · Evidence Ledger

A claim without a source, an owner and an expiry is not a claim. File the source document against the claim reference and the gate stops firing on it.

Meta-metric

gates_per_person_per_week

Target direction: down, while output volume trends up. If gate load per person rises with volume, the program has industrialized approval rather than production. This is the single number the ExCo should watch.

The distinction that carries the whole design

A review is not a gate

Review

A review asks a human to read something and feel responsible for all of it. It has no completion criterion, no stated options, and no way to be done well or badly. Reviews expand to fill available senior time.

Gate

A gate presents a specific question, a recommended answer, the evidence behind it, the consequences of each option, and a deadline. It is answerable in under five minutes by the right person and is auditable afterward.

In a conventional function, a market press statement touches seven to twelve people across three to five days and produces an unbounded number of review cycles. In CHORUS the same statement touches two gates — G2 TRUTH (silent, unless a claim is unsubstantiated) and G5 RELEASE — for a combined human time of roughly four minutes. The work does not get less rigorous. The rigor moves from serial human reading into the coherence checks the Chief of Staff runs before anyone is asked anything.

G1 → G5

The five gates, specified

Each gate names what it decides, who holds it, what the holder sees, the options, when it fires and what happens on timeout.

G1POSITIONCANON-LEVELfires 4×|writes to narrative_canon

Is this what we believe, and is it a change from what we said before?

What the company's stance actually is. A leadership judgment about values and risk appetite, not a communications choice.

What the holder sees

The proposed position as a message triangle, a diff against the current Canon entry, the list of downstream artifacts that will change, and the markets affected.

When it fires

Once per position, not per output. Adoption triggers the change cascade.

adoptadopt with editrejectescalate to ExCo

Held by

CEO · CCO · General Counsel

Delegable to: none

Deadline by posture

NORMAL

4h

ELEVATED

30m

CRISIS

8m

On timeout: Escalate one level. Never auto-approve.

Evidence pack

  • precedent_matches
  • stakeholder_exposure
  • simulated_reactions
  • recommended_position
  • one_named_alternative
G2TRUTHCLAIM-LEVELfires 2×|writes to evidence_ledger

Can we prove every factual assertion, in every market this will appear in?

Whether every factual claim in the artifact is supported, current, cleared and jurisdictionally safe.

What the holder sees

Only the unsubstantiated or expiring claims, extracted and listed, with their current evidence and expiry.

When it fires

Automatically when any claim in an output is missing, expired or out of jurisdictional scope. Passes silently when everything checks.

B2 pre-resolves every claim against the Evidence Ledger and presents only unresolved or expiring claims. A clean G2 should take under two minutes.

substantiateamendstrike

Held by

Evidence owner · Finance for financial claims · General Counsel for legal claims

Delegable to: Canon Steward, for claims sourced entirely from filed or audited material

Deadline by posture

NORMAL

4h

ELEVATED

30m

CRISIS

8m

On timeout: Block release.

G3RISKESCALATINGfires 4×|writes to policy_risk_register

What is the worst plausible reading of this, and can we live with it?

Regulatory, litigation, employment-law and jurisdictional exposure. Whether this can be said here, now, by this person.

What the holder sees

The output, the adversarial read produced by D4, the simulated reaction from the three most hostile personas, litigation and regulatory exposure flags, and the closest three precedents with their outcomes.

When it fires

On materiality score, regulated topic, or crisis posture.

proceedproceed with named mitigationsholdkill

Held by

General Counsel · Chief Compliance Officer · Regional counsel

Delegable to: none

Deadline by posture

NORMAL

4h

ELEVATED

30m

CRISIS

8m

On timeout: Block release.

Hard blocks

  • active_litigation
  • regulatory_filing_window
  • employment_tribunal_matter
  • embargo_window
  • material_nonpublic_information
G4VOICEPRINCIPAL-LEVELfires 1×|writes to voice_corpus

Does this sound like the person whose name is on it?

Whether this sounds like the named principal and is one they will say without rewriting.

What the holder sees

The draft, a voice-fidelity read against the Voice Corpus, and the three passages flagged as least characteristic.

When it fires

On any first-person attributed content. Edits are captured and feed the Voice Corpus, so the gate gets cheaper every time it fires.

approveeditreject

Held by

The principal, or their named chief of staff

Delegable to: Principal's office, by standing instruction recorded in Voice Corpus

Deadline by posture

NORMAL

4h

ELEVATED

30m

CRISIS

8m

On timeout: Drop to A2 and route to a human writer.

G5RELEASEALWAYSfires 5×|writes to evidence_ledger,precedent_library

Right thing, right audience, right moment?

Whether this goes out, to whom, on which channel, at what time, and with what monitoring attached.

What the holder sees

The final artifact, the distribution list, the timing against the corporate calendar, the embargo state, and any conflicting release in the same window.

When it fires

On every external release without exception. Below the materiality floor it is a one-click confirmation; above it, a full stop.

releaseholdreschedule

Held by

Channel owner · Regional comms lead · Ops domain lead

Delegable to: A4_envelope, where the artifact class, audience and channel are pre-approved

Deadline by posture

NORMAL

4h

ELEVATED

30m

CRISIS

8m

On timeout: Hold. A missed release is recoverable; a wrong release is not.

Load

Where the gates actually fire

Fifteen firings across fourteen steps in five workflows. G5 RELEASE fires in all five. G4 VOICE fires once, in the conduct workflow.

Gate firings across WF1–WF5

rendering…

Counted per firing, not per workflow: one step in WF3 combines G2 and G5 under a single eight-minute deadline.

GateFiresWorkflowsHolder
G1POSITION3WF1, WF4, WF5CCO, or CEO+CCO+GC where the position is a values statement
G2TRUTH2WF2 (exception-only), WF3 (combined)Claim owner; incident commander + GC in crisis
G3RISK4WF1, WF3, WF4, WF5GC with CCO or incident commander
G4VOICE1WF4The principal whose name is on the words
G5RELEASE5All fiveVP Comms, market leads, CCO, or GR leads by jurisdiction

Four operating principles

How the gate system stays cheap

Silent passes are the goal

G2 fires only when something fails. A function running well sees G2 almost never, which is exactly the intended state: the Evidence Ledger is doing the work a legal reviewer used to do by reading.

Every override is training data

When a human decides against the recommendation, the reason is captured in structured form. The gate that fires wrongly twice should not fire a third time.

Deadlines are real

Every gate carries a deadline scaled to posture: 4 hours normal, 30 minutes elevated, 8 minutes in crisis. On expiry the decision escalates one level automatically. A gate without a deadline is a queue, and a queue is how a 4-hour statement becomes a 3-day statement.

Gate load is a first-class metric

The platform reports gates-per-person-per-week and median time-to-decision. Rising gate load means autonomy ceilings are set too low or materiality thresholds are miscalibrated — a tuning problem with an owner, not an inevitability.

“Our function would have end-to-end automated processes in place for almost every workflow and conceivable output and integrate human perspective and decision making at precisely the necessary moments along the way.”

Joe Evangelisti · JPMorgan Chase

gate = decisiondeadline scaled to postureoverride = training data