L0 · Governance Plane · Gate Load
Autonomy levels, ceilings, postures and gate load
79.6% of steps run without a human in the execution path. Every level, ceiling, posture and failure mode behind that number is written down, owned and reversible on one person's authority.
GATES / PERSON · WK
9.2
RATE
79.6%
SHADOW COVERAGE
56.3%
FAILURE WATCH
2 of 3 tripped
Autonomy rate
0.0%
Counted, not asserted
Human-in-path steps
0
A0 + A1 + A2, concentrated at gates
Declared ceilings
0
Reviewed quarterly
Named failure modes
0
Each with an indicator and a control
The meta-metric
Gates per person per week, falling while output rises
One number judges whether the program is working, and two failure modes end it: gate-load inflation and apprenticeship collapse. This ledger instruments all three.
Gates per person · week
0.0
Down 44.0% over 11 weeks
Decline from absorption
0.0%
The other 70.8% is a bigger bench
Shadow-gate coverage
0.0%
118 days dormant at G1
Years to first gate
0.0
Up from 3.1 · 7 quarters
Two of these four numbers are pointed the wrong way
Gate load per person fell 44.0% while artifacts released rose 93%. Only 29.2% of that decline came from the platform absorbing firings; the rest came from putting more people on the bench, which is real relief and a different claim. Meanwhile the analyst population that feeds that bench fell 35.9% — 4.6 leaving for every seat added — and at G1, the one gate that cannot be delegated to anyone, no junior has recorded a shadow decision in 118 days. Both of those are in the ledger below rather than underneath it.
16.4 → 9.2 per person per week, output up 93%
Kills the program fastest — visible executive pain inside one quarter.
Indicator · gates_per_person_per_week trending up
G1 dormant · longest 118 days at G1
Kills the program most completely — invisible for three years, then irreversible for ten.
Indicator · No junior has held a shadow gate in 90 days
4.6 juniors out per bench seat added · years-to-first-gate 3.1 → 4.4
Second indicator on the same failure mode.
Indicator · Junior headcount falling faster than the bench grows; years-to-first-gate rising
Gates per person per week · 11 weeks
16.4 → 9.2
Down 44.0% while artifacts released rose +93% — 61 to 118 a week.
The platform working as designed. Fewer decisions asked of humans, and more of them agreed.
Read the decline before you present it
Two entirely different things move this number. The platform can absorb gate firings, or the bench can get bigger and the same work can be spread thinner. Only the first is the platform working. Over this window the split is 29.2% absorption and 70.8% bench.
Absorption effect
-2.11
Fewer firings reaching humans at all: 148 a week down to 129.
Bench effect
-5.12
More named holders carrying it: 9 up to 14. Real relief, but not the same claim.
Load down, output up · 12 weekly readings
W-11
W-10
W-9
W-8
W-7
W-6
W-5
W-4
W-3
W-2
W-1
Now
Output bought per unit of human attention rose from 0.41 to 0.91 artifacts per gate held — a factor of 2.22×.
Gate load rising
Autonomy ceilings are too low, or materiality thresholds are miscalibrated. The program has industrialized approval rather than production.
Load falling, overrides rising
Autonomy has been pushed past what the humans trust. The gates are firing less and disagreeing more.
Load falling, overrides falling
CurrentThe platform working as designed. Fewer decisions asked of humans, and more of them agreed.
Who actually holds it · 14 named individuals
Fair share 9.2 firings / week
General Counsel
2.28× fair share · 16.3% of the week
21
+17%
Canon Steward
1.52× fair share · 10.9% of the week
14
-33%
Regional Comms Lead · EMEA
1.41× fair share · 10.1% of the week
13
-41%
Chief Compliance Officer
1.30× fair share · 9.3% of the week
12
-25%
Channel Owner · Owned & Social
1.19× fair share · 8.5% of the week
11
-42%
Ops Domain Lead
1.09× fair share · 7.8% of the week
10
-55%
Chief Communications Officer
0.98× fair share · 7.0% of the week
9
-36%
Regional Counsel · EMEA
0.87× fair share · 6.2% of the week
8
Joined
Evidence Owner · Finance
0.76× fair share · 5.4% of the week
7
-42%
Regional Counsel · APAC
0.65× fair share · 4.7% of the week
6
Joined
Market President's Office · US
0.65× fair share · 4.7% of the week
6
Joined
CFO's Chief of Staff
0.54× fair share · 3.9% of the week
5
Joined
Chief Sustainability Officer's Office
0.43× fair share · 3.1% of the week
4
Joined
Chief Executive Officer
0.33× fair share · 2.3% of the week
3
-25%
The top three holders carry 37.2% of the week between them, and 6 of 14 sit above the fair-share line. Across the 9 people who were already on the bench at the start of the window, load fell -32.4% — with exactly 1 exception. General Counsel is up +16.7%, sits on 3 of the five gates, and 2 of those cannot be delegated to anyone.
Definition
What “80% autonomous” means here
The phrase is defined against steps, not outcomes. The full definition is below.
"80% autonomous" means at least 80% of discrete process steps execute at autonomy level A3 or A4 — without a human in the execution path — with all remaining human effort concentrated at named, pre-declared judgment gates. It does not mean 80% of hours.
Counted as autonomous
Counted as human-in-path
A3 removes the human from the execution path even though a human still decides release. That distinction is the whole design.
Counted, not asserted. Every step in Appendix B carries a declared level.
Step distribution by autonomy level
98 steps across five workflows
L0 · The ladder
Five levels, mapped to the published task mix
A3 carries the program. The distinction between A2 and A3 is whether a human touches the artifact or only approves it.
A person makes the call. The system assembles the inputs, states a recommendation and a confidence, and stops.
Agent role
Task-mix archetype
A person produces the output. Agents supply research, precedent, evidence and voice reference in-line.
Agent role
Task-mix archetype
Agent drafts, human materially rewrites. Expected human edit distance is high and is measured.
Agent role
Task-mix archetype
Agent produces the finished output. A named human approves at a gate before release. The human does not touch the artifact.
This is the level that carries the program. Most of the 80% lives here.
Agent role
Task-mix archetype
Agent executes and releases within a declared envelope. Humans see results, not requests.
Agent role
Task-mix archetype
The arithmetic
Autonomy per workflow — including the one that fails the target
78 of the 98 specified steps run with no human in the execution path. The five workflow readings show where each sits.
| Workflow | Steps | A0 | A1 | A2 | A3 | A4 | Auto | Rate | Ceiling |
|---|---|---|---|---|---|---|---|---|---|
WF1 Value dispute with the franchise system | 22 | 4 | 0 | 1 | 6 | 11 | 17 | 77.3% | A3 |
WF2 Global position cascade | 17 | 2 | 0 | 0 | 2 | 13 | 15 | 88.2% | A3 |
WF3 The golden hour | 23 | 3 | 0 | 0 | 6 | 14 | 20 | 87.0% | A3 |
WF4 The conduct issue | 18 | 5 | 0 | 2 | 3 | 8 | 11 | 61.1% | A2 |
WF5 Multi-jurisdiction policy campaign | 18 | 3 | 0 | 0 | 5 | 10 | 15 | 83.3% | A3 |
| Aggregate | 98 | 17 | 0 | 3 | 22 | 56 | 78 | 79.6% |
Autonomy rate by workflow
Target line is 80%. WF4 is deliberately below it.
L0 · Policy & Risk Register
Declared ceilings ratchet down instantly and up only with evidence
Per workflow, per topic, per market, per voice. Published in the Policy & Risk Register. Visible to every gate-holder.
61.1% autonomous by design. Sixteen points below the next-lowest workflow and nineteen below the 80% target.
Ceilings ratchet DOWN instantly, on any named individual's authority. They ratchet UP only at the governance cycle, with evidence.
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.
This is what compounds. Without it, the platform is a very expensive drafting tool.
Legal must be able to ratchet down a ceiling at 23:00 on a Friday without a deployment.
L0 · Posture system
One switch changes every deadline and every ceiling in sixty seconds
Posture is declared by a named human, never inferred. It is the highest-leverage control on the platform.
Standing operation.
Gate deadline
Autonomy modifier
Trigger
A material situation is live but contained. Not a crisis.
Gate deadline
Autonomy modifier
Trigger
Declared by
A posture, not a department. The same agents run; the ceilings, cadences and gate-holders change.
Gate deadline
Autonomy modifier
Trigger
Declared by
No AI toolkit handles a crisis end to end. The system compresses preparation from hours to minutes; the judgment stays human and the leadership stays human.
Posture declaration
sign-in requiredTopic-scoped ceiling drops one level. Sensing cadence moves to continuous on the affected topic.
L1 · Chief of Staff
The orchestrator that never writes, never publishes, never decides
L1 orchestrator. One per function.
- The CoS does not write.
- The CoS does not publish.
- The CoS does not decide.
Rule R1 routes every handoff through this layer. Swarms that self-coordinate cannot be audited, debugged or governed.
Every inbound signal, request and event resolves to a situation class or is escalated as unclassified.
Select the domain agent and task agents. Record the routing rationale.
Read the Policy & Risk Register before every invocation. No exceptions, no caching beyond 60 seconds.
Order dependent steps; parallelize independent ones; hold steps whose inputs are stale.
A gate-holder receives a recommendation, a named alternative, the evidence, the precedent and a deadline. Never a raw draft.
Gates approaching deadline escalate automatically. Chasing is the CoS's job, not a human's.
When two agents produce contradictory outputs, the CoS surfaces the contradiction rather than picking a winner.
Propose posture changes with evidence; execute declared changes across every running workflow within 60 seconds.
Track token and tool spend per workflow against the spend cap. Degrade gracefully, never silently.
Confirm every Canon write completed before marking a workflow step done.
Build governance
THE GATE — which agents get built, and when they get killed
Build governance. Which agents get built, by whom, and when they get killed.
Owner
Artifact
Owner
Disqualifying question
Artifact
Owner
Rule
Artifact
Owner
Rule
Artifact
Below a declared spend and risk threshold, anyone builds anything without approval. Above it, THE GATE applies. The threshold is published.
Anything a single person can build and test in one to two days needs no sponsor, no gate, no paperwork. Governance that reaches down to the individual prompt kills the experimentation that produces the good candidates in the first place.
The sponsor is a market comms lead, a practice head, a regional GR director — not the AI team, not IT. The sponsor's commitment is that their people will use it. Confusing these two is how programs produce excellent tools with 6% adoption.
"Why can’t the general-purpose assistant already deployed do this?" is asked at stage 1 and has to be answered in writing. Most proposals fail it, and that is the point. Voice you judge by ear — that is a prompt. A materiality score a GC will be asked to defend — that is a product.
Every build carries a date and a threshold set before a line is written. Adoption below the threshold at the date means the agent is retired, its Canon writes are frozen, and its budget is released.
Governance artifacts
Three files that make the whole thing auditable
Agents read them at runtime. Changing a ceiling is a data edit with an audit record, not a deployment.
The machine-readable file holding every autonomy ceiling, every no-fly zone, every jurisdictional constraint, every claim-substantiation rule. Agents read it at runtime. Changing a ceiling is a data edit with an audit record, not a deployment.
Owner
Review
Every gate fired, who held it, what they saw, what they chose, how long they took, and what happened afterwards. The audit trail, the training corpus for the override loop, and the evidence base for raising or lowering ceilings.
Owner
Review
Every agent live, its sponsor, its ceiling, its sunset date, its adoption rate, its override rate, its cost. One page. If it does not fit on one page the roster is too big.
Owner
Review
Operating model
Five roles the function does not have today
2.7 net new FTE plus a distributed gate-holding assignment, including the Canon Steward.
Canon currency, completeness and drift. Owns the single number 'percentage of live positions past expiry'.
Owns
Decides
Sits
Scale
Agent register, evaluation sets, cost per workflow, incident response.
Owns
Decides
Sits
Scale
Simulator accuracy against observed outcomes. Owns the calibration loop and the confirmation-bias control.
Owns
Decides
Sits
Scale
Named per gate per workflow. Gate-holding is in the job description and the objectives, not a favor.
Owns
Decides
Sits
Scale
Reinvestment of recovered capacity. Comp-tied. Without this role, 67% of leaders report GenAI enables new work and only 8% actually reinvest.
Owns
Decides
Sits
Scale
A small senior judgment layer, a thin technical layer, and nothing in between. Cheapest. Also has no mechanism for producing its own next generation of gate-holders and no surge capacity when three situations run at once. Every hour recovered goes to the P&L.
The same senior layer, a deliberately preserved and re-purposed middle — people who hold shadow gates, curate Canon, run calibration, and do the strategic work the function has never had time for — and a thin production base. Requires committing roughly half the recovered hours back into the function before the savings are booked.
“You can't break the apprenticeship model.”
Declared failure modes
Eight ways this program dies, each with an indicator and a control
Gate-load inflation kills it fastest — it produces visible executive pain within one quarter and generates an internal constituency for shutting the thing down. Apprenticeship collapse kills it most completely, because the damage is invisible for three years and then irreversible for ten.
Indicator
Control
Indicator
Control
Indicator
Control
Indicator
Control
Indicator
Control
Indicator
Control
Indicator
Control
Indicator
Control