The Factory Charter
The one ratified document the factory's organisation, business case and operating model hang off: mission, mandate and decision rights, operating principles and ordered values, each value with an owner, policy, metric and veto. Covers the clauses to adopt, who drafts, co-writes, signs and ratifies them, how they compile into running systems, and how drift and decoration are caught.
- Version
- 0.1.0
- Owner
- DEIXEI
- Last reviewed
- Reading time
- 29 min
For Cto · Cio · Head Of Ai Factory · Risk And Assurance · Internal Audit · Platform Engineer
Sources and details
- Section
- Adopt and fund
- Source works
- Book 2 · Book 09 · Research Series
- Source chapters
- 10, 12, 18, 19, 20, 24, 27
- Tool specifications
- audit_evidence_schema@0.1.0
- Refresh trigger
- source change
Every agent in the estate reads a mission statement at the start of every run, whether or not the firm wrote one. Usually it is the opening lines of a system prompt, drafted by whoever stood the agent up. A firm running fifty agents without a ratified charter does not have no values. It has fifty unratified ones, and no two of them resolve a conflict the same way (source claim, Book 2 Ch 18 §2; Research Paper 5 §3.1).
The factory charter is the document that replaces them. In the books’ sense, a charter is what an institution grants a function: a purpose, a boundary, decision rules, and a declared order for resolving conflicts (Book 2 Ch 18 §1; RP5 §4.1). The business case, the Agent-Driven Department design and the operating model all point back to it. Unlike the mission-and-values documents it replaces, it has a machine reader, so it is written, versioned and released like a specification.
This page is the artefact. It sets out the clauses a firm should adopt, who drafts, co-writes, signs and ratifies each one, how the clauses compile into running systems, and the checks that keep the charter from turning into decoration.
Reading the labels. Source claim means Book 2, Book 09 or Research Paper 5 states it. Author model marks the author’s proposals: the charter method, the owner/policy/metric/veto test, the two-compiler test, the control budget and the trust ledger. RP5 labels its whole analysis and method as argument and proposal. Recommendation is this page’s synthesis, including the clause template and the decision-rights table, which assemble source rules into one form. Composite marks Asset Manager X, Céline and Margaux. The worked charter is the composite’s, derived from its own failures, and must not be copied (Book 2 Ch 18 §6).
What the charter contains
The usual mission-and-values card fails partly because it merges four different things. The charter keeps them apart because they differ in kind, in stability and in how they are enforced (source claim, Book 2 Ch 18 §4.3; RP5 §3.2):
| Part | Says | Stability | Enforced by | Sorting test: a sentence belongs here if it names… |
|---|---|---|---|---|
| Mission | Why the factory exists and for whom | Years | The roadmap: an initiative that does not serve the mission does not pass intake | A customer and an outcome |
| Mandate | What the factory may do and, more importantly, what it must not touch | Changes with the operating model | Structure, not procedure: the stay-list | A boundary of authority |
| Operating principles | How day-to-day decisions are made | Stable; collected once | Procedure: gates, reviews, defaults | A decision rule you can hold a decision against and get yes or no |
| Values | What the factory will pay for when two goods conflict | Revised on events | Policy as code, with a veto | A trade-off and which side wins |
A sentence that names none of the four is a slogan, and it is deleted.
The charter at department level. The Agent-Driven Department is itself governed as a bounded-agency entity, and the mission is the top line of its Mandate. The mandate’s negative half is the department’s stay-list (source claim, Book 2 Ch 18 §4.3, Ch 20 §4.3).
Who drafts, co-writes, signs and ratifies
The charter is a governance act, not a communications exercise. The roles below come from Book 2 Ch 18 §7, Ch 19 and RP5 §4.1. Where the sources differ in detail, the difference is noted.
| Role | Duty on the charter | Source |
|---|---|---|
| Head of the factory | Drafts. Owns the derivation, the versioning and the quarterly report. Standing rule: never present the productivity line without the control-capacity line | Book 2 Ch 18 §7, §8 |
| Four counterparties | Review the draft. The derivation interviews are with the loudest sponsor, the quietest sponsor, the CRO and one wrangler | Book 2 Ch 18 §7 |
| CISO and CRO | Co-authors, not audience. They help write the charter, including the mandate’s stay-list, rather than having it done to them. In Book 09’s composite the draft went to both before anyone else, with a request to change it rather than approve it, and the CRO changed one word in the first value from rarely to never | Book 2 Ch 19, Ch 20 §4.3; RP5 §4.1 step 4; Book 09 Ch 10 |
| Chief Risk Officer | Signs the values that set risk trade-offs, above all the control-budget threshold, as second-line approval. Signed in writing, privately, before the executive committee meeting | Book 2 Ch 18 §6–§7, Ch 19; RP5 §4.1 step 4 |
| Wrangler and policy engineer | In the room for the two-compiler test, to confirm each value actually compiles on the machine side | Book 2 Ch 18 §7 |
| Executive committee | Ratifies the mission and the order of the values | RP5 §4.1 step 4; Book 2 Ch 18 §7 |
| The seven named seats | Each value is owned by one of the seven names. The seven names form the values registry | Book 2 Ch 18 §6, Ch 24 §7; accountability |
| Second line | Reads the quarterly revealed-values report and draws the conclusion. The factory supplies the data and does not grade itself | Book 2 Ch 18 §7; RP5 §3.7 |
Time to ratify (source claim, Book 2 Ch 18 §7). For a firm of about a thousand people or one production line: about a fortnight, five values, one preamble and one policy pack. For a fifty-thousand-person bank: expect a quarter, because the ordering will be contested. The argument over whether value 1 outranks the year-end deadline is the exercise, and having it in the executive committee is cheaper than having it in an incident.
Inputs: derive, do not draft
A charter is derived from two sources only: the factory’s customers, and the fundamentals that decide whether any business survives. Every clause traces to a failure the firm has documented (source claim, Book 2 Ch 18 §3; RP5 §3.4).
The customer map
Write the map before any word of the charter. It is one page, it goes to the executive committee unchanged, and it stops the charter being written for an unnamed buyer. Columns and rows from Book 2 Ch 18 §5.1, paraphrased:
| Counterparty | What they are to the factory | What they get | What they give | What losing them looks like |
|---|---|---|---|---|
| Product teams, pods and the agents | Users | The foundation, recipes, paved road and components | Nothing directly; usage is the signal | Adoption decays and the shadow road wins on friction |
| Business-line sponsors | Customers | Intent turned into running software while the intent is still true | Capacity allocation, budget defence, sponsorship in the room | Budget erodes over two planning cycles |
| Executive committee and board | The customer above the customer | Continuation | Patience, which cannot be borrowed | The factory becomes a project with an end date |
| The supervisor | Not a customer. The counterparty whose acceptance the estate runs on | The evidenced answer, on demand, in the firm’s own schema | Nothing; the relationship is not transactional | Findings, then constraints on running agents at all |
Keep the map current and revise it when a sponsor changes, not annually.
Name the supervisor without commercial vocabulary. The factory has two parties it must satisfy, and only one of them is in the room. The supervisor does not buy, pay or renew, and describing the relationship in customer language is inaccurate and offensive to a supervisor reading it. What the factory owes is the evidenced answer to three questions: what did your agents do, on whose authority, and under which control (source claim, Book 2 Ch 18 §4.2; RP5 §3.4). A mission that names only the business will trade evidence for speed under pressure. Naming both builds the counterweight into the sentence.
The derivation interviews
Ask four people what they would lose if the factory vanished on Friday: the loudest sponsor, the quietest sponsor, the CRO and one wrangler. Do not ask what they want. The mission is assembled from the losses (Book 2 Ch 18 §7). RP5 §4.1 step 1 gives a broader set: product leads, the CRO and the function that deals with the supervisor, over about two weeks. In Book 2’s composite the derivation took two weeks, ran no workshop, and each participant answered the question in writing.
The current, unratified charter
Before drafting a value, pull the shared system-prompt preamble the fleet runs on today. If there is none, collect five agents’ local preambles and read them side by side. That is the firm’s current charter. Take it to the executive committee as an exhibit (Book 2 Ch 18 §7; RP5 §5 rec 1: the number of distinct statements is the size of the problem).
The fundamentals that bind it
Book 2 Ch 18 §4.1 applies eight business fundamentals to the factory: value creation, customers, unit economics, cash, differentiation, repeatability, trust and adaptation. Four of them shape the charter text directly:
| Fundamental | Consequence for the charter |
|---|---|
| Value creation | The mission names a pain somebody feels. Any sentence beginning “to harness AI” is ruled out: it names a means, ages with the technology and describes the producer |
| Customers | The mission names who is served and how they are reached, which makes adoption a charter concern |
| Trust | The charter has several counterparties on different clocks (trust ledger) |
| Adaptation | The charter is versioned and revised on events |
The other four bind it from outside the text. Unit economics decides whether the factory should exist at this portfolio size. Cash decides the sequencing, because money out comes two to three quarters before money back and the factory spends board patience in the meantime. See build, configure or consume for the economics.
The charter clauses to adopt
The template below is this page’s recommendation. It assembles the source rules into clauses a firm fills in from its own derivation. The bracketed text is guidance, not wording to adopt.
Clause 1: Mission
1.1 The factory exists to [outcome for the business, tied to intent], [the supervisory condition], so that [the control-gap clause].
1.2 Version: [x.y]. Ratified by the executive committee on [date]. Each clause traces to the documented failure listed in Annex A.
Drafting rules (source claims, RP5 §3.4, §4.1; Book 2 Ch 18 §4.1, §5.2):
- At most two sentences. Falsifiable. No adjective that cannot be tested.
- Name both the business and the supervisor, without commercial vocabulary for the supervisor.
- Do not derive it from the technology (“to harness AI…”), from the org chart (“to be the centre of excellence for…”), or from untestable words (“world-class”, “leading”, “trusted”).
- RP5 binds the mission with three failures it must answer. The control gap: change arriving faster than the firm can understand it and stand behind it. The fragmentation tax: crossing surfaces owned under different leaders. Embedding in value: staying tied to what the business actually uses.
Clause 2: Mandate and decision rights
2.1 The factory may: [first-line build-and-run scope — see the Agent-Driven Department’s “moves” list].
2.2 The factory must not: [the stay-list — security policy and refusal rules, threat intelligence and security incident command (CISO); second-line risk (CRO); data protection (DPO); business application ownership and product decisions (business lines)].
2.3 The factory produces evidence to the second line, never instead of it. Second-line read access to the evidence record is granted under the CISO’s identity rules and cannot be revoked, filtered or scheduled by the factory.
2.4 Decision rights are as set out in the table below.
The stay-list and its rules are explained on Agent-Driven Departments (Book 2 Ch 20 §4.3; Book 09 Ch 12). Decision-rights table (recommendation, assembled from the sources cited):
| Decision | Decides | Must be consulted or co-sign | Source |
|---|---|---|---|
| Mission text and the order of values | Executive committee ratifies | Head of factory drafts; CISO and CRO co-write | RP5 §4.1; Book 2 Ch 18 §7, Ch 19 |
| Any value that sets a risk trade-off, including the control-budget threshold | CRO (second-line approval, in writing, before ratification) | Head of factory | Book 2 Ch 18 §6; RP5 §3.6, §4.1 |
| Whether an initiative enters the roadmap | Head of the factory, against the mission and declared expected value | Product lead; sponsor | Book 2 Ch 18 §4.3, §6 |
| What gateways and brokers refuse | CISO (refusal rules held as code; the CISO team’s approval is the only gate) | Factory implements | Book 09 Ch 12 |
| Command of a security incident | CISO | Factory responders | Book 2 Ch 20 §4.3 |
| Product decisions and accountability for them | Business line | Factory product lead | Book 2 Ch 20 §4.3 |
| Promotion of an agent up the autonomy ladder | Per agent, runbook and domain, on evidence from the audit spine; frozen while the domain’s control budget is spent | Agent-fleet operator; wrangler | Book 2 Ch 18 §5.3, §6; control budget |
| Conclusion of the quarterly revealed-values review | Second line | Factory supplies the data | Book 2 Ch 18 §7; RP5 §3.7 |
| Revising the charter | Same route as ratification; triggered by events, not anniversaries | — | RP5 §4.1 step 6 |
Clause 3: Operating principles
3.1 The factory decides by these principles. Each is tested at review against a specific recent decision.
Book 2 Ch 18 §5.3 collects five principles that are already established across the book. Do not rebuild the card from memory:
| Operating principle | Decision it settles | Established in |
|---|---|---|
| The safe path is the fast path | Whether the governed route is made the convenient one, or only the compliant one | Book 2 Ch 1, Ch 6 |
| Evidence is a by-product, never a task | Whether evidence gets a work item or a pipeline | Book 2 Ch 9 |
| Agents write the code; the recipe decides its shape | Whether the agent may choose the pattern | Book 2 Ch 5, Ch 28 |
| Every loop has an escape | Whether an autonomous loop ships without a stop condition | Book 2 Ch 24 |
| Autonomy is earned per runbook, never granted in bulk | Whether a working agent is promoted by demonstration or by argument | Book 2 Ch 20 |
Principle test. Take a decision from last month, hold it against the principle and answer yes or no. If the answer is always yes, the principle is not constraining anything and it is an aspiration. RP5 §4.1 allows five to ten principles.
Clause 4: Values, ordered
4.1 The factory holds the following values, in this order. Where two conflict, the higher-ranked prevails.
4.2 Each value is recorded with its trade-off, policy, metric, veto and owner. A value missing any of the four enforcement elements is not in force.
4.3 Values that carry a veto are enforced as rules in the policy engine and pipeline. The remaining order and the principles are written for judgement and carried in the shared preamble with their reasons.
Each value row has these fields (source claim, RP5 §3.5, §4.2; Book 2 Ch 18 §6):
| Field | Content | Test |
|---|---|---|
| Order | Rank, 1 = highest | Order is declared in the document, not implied |
| Value | Short declarative sentence | Names a trade-off and which side wins |
| Trade-off it resolves | The two goods in conflict | — |
| Policy | The value as a rule enforced at the gateway or in the pipeline | It is code, not a sentence |
| Metric | The value as a gauge somebody looks at | Named gauge, named reader |
| Veto | The value as a stop: where it may cost something automatically, without a meeting | It fires without a meeting |
| Owner | A named person in a named role, one of the seven names | Not a team, mailbox or mechanism |
Clause 5: Governance commitments
5.1 Control budget. Each domain has a CRO-approved control budget. When it is spent, intake in that domain is throttled and no agent is promoted until arrival and capacity reconverge. Demotions remain available.
5.2 Counter-incentive. The productivity line is never presented without the control-capacity line.
5.3 Versioning. The charter has a version number and lives on the machinery backlog. Every compiled artefact declares the charter version it derives from. The audit record can answer which version an agent action ran under.
5.4 Revealed values. Each quarter the factory reports, per value, vetoes fired, waivers granted and the trend. The second line reads the report and concludes. A value with no recorded cost in two consecutive quarters is examined and either sharpened or removed.
5.5 Revision. The charter is revised when events require it, by the ratification route, and not on an anniversary.
5.6 No trust scorecard. Counterparty trust is read from existing instruments. No trust dashboard or per-team trust score is built.
Sources: control budget, Book 2 Ch 18 §6, Ch 19, Ch 24 and RP5 §3.6, with the operating instrument on control budget and the demotion rule from Book 09 Ch 12. Counter-incentive: Book 2 Ch 18 §8. Versioning: Book 2 Ch 18 §8, Ch 24 §5 and RP5 §3.3, §5. Revealed values: Book 2 Ch 18 §7 and RP5 §3.7, §4.1. Trust: Book 2 Ch 18 §4.4.
Why 5.1 and 5.2 are in the charter, not the business case (source claim, Book 2 Ch 18 §2, Ch 19). Most factories are funded on a throughput case. A throughput incentive points at the value the factory most needs to defend: that change never runs ahead of the firm’s ability to stand behind it. When a value and an incentive conflict at scale, the incentive wins. Two structural counter-incentives exist: the control budget, which is a brake inside the machine, and the supervisor named in the mission sentence. RP5 §5 rec 9: if the factory is funded on throughput alone, tell the board that the funding model and the first value are in conflict.
Annex A: Clause traceability
For every clause of the mission and every value, record the documented failure it answers, from the firm’s own history or the books. At each review, read the clause, name the failure, and check that the failure is still live (Book 2 Ch 18 §5.2 rule four).
The four rules that make a charter compile
(Source claims, Book 2 Ch 18 §5.2; RP5 §3.8, §4.1.)
-
The two-compiler test. Every value compiles twice:
- human consequence: a hiring criterion, a promotion input, or a veto a named person can exercise;
- machine consequence: a preamble clause, a policy rule, a gate or a stop condition.
A value that compiles into neither is deleted on sight. One that compiles into only one side is marked half-built and stays marked until it compiles into both. The test takes about ninety seconds per candidate value. Run it with a wrangler and a policy engineer in the room.
-
Declare the order. The factory’s conflicts (evidence against latency, throughput against review capacity, cost against the resilience floor) arrive mechanically and with nobody available to rule on them. An unordered list resolves nothing. Precedent cited by Book 2 Ch 18 and RP5: Anthropic’s published constitution for Claude declares a priority ordering rather than a set. Book 2’s note adds that Anthropic describes that prioritisation as “holistic rather than strict”. As of November 2026, per Book 2 Ch 18 [FACT-CHECK: confirm the current wording and ordering in Anthropic’s published constitution at anthropic.com/constitution].
-
Five to eight values, hard maximum. Past about eight, a list stops ordering decisions and starts justifying them, because every position can find a principle to cite. RP5 holds this as a rule of thumb and knows of no study that sets the number.
-
One clause per documented failure. A clause that answers no named failure is a slogan.
What the illustration shows about deletion. In Book 2’s composite, a fourth candidate value (“we treat every colleague’s time as the firm’s scarcest resource”) survived one draft, compiled into nothing on either side, and was deleted. It was true, but no decision resolved differently because of it. The deletion is the part worth copying.
Worked illustration (composite; do not adopt)
Book 2 Ch 18 §6 and RP5 §4.2 each work three values for the composite. A firm that adopts these rows verbatim has skipped the derivation, which is the only step that makes a charter its own. The rows show the shape of a complete row, not content to reuse.
Illustrative mission (composite): turn the firm’s intent into running, evidenced software at the speed the business needs, inside an envelope the supervisor can inspect, so that change never runs faster than the firm’s ability to stand behind it. Each clause traces to a failure: intent to the expired-need product, evidenced to the supervisory position, and never faster than to the CRO’s unanswered question about the maximum safe rate of change.
| # | Value | Policy | Metric | Veto | Owner |
|---|---|---|---|---|---|
| 1 | Change is never faster than control | Control budget per domain, CRO-signed | Book 2: control-gap gauge; review latency; evidence completeness, quality-weighted. RP5: queue in days of review capacity; sampled review quality | Budget spent → intake throttles, autonomy ladder freezes | Head of the factory; threshold approved by the CRO |
| 2 | Evidence is a by-product, never a task | No release without a complete nine-question record, enforced in the pipeline | Book 2: share of changes with complete records, target 100 per cent. RP5: share of released changes with a complete record | Release blocked automatically on incomplete evidence | Book 2: operational owner. RP5: head of platform engineering |
| 3 | Book 2: We build what is used, not what was asked for. RP5: Autonomy is earned, never granted | Book 2: expected value declared at intake or no roadmap entry. RP5: each autonomy level granted per agent, task and domain on a stated record | Book 2: usage-derived value against declared expectation. RP5: actions beyond the granted level | Book 2: collapsed-usage products pulled at roadmap review. RP5: first such action returns the agent to advisory on that task | Book 2: product lead. RP5: head of agent operations |
The order does work. Value 1 outranks any value about meeting demand, so the control budget may throttle intake while a sponsor is asking loudly and correctly. Value 2 outranks the speed clause of the mission and the standard route, so even a pre-approved pattern cannot ship without evidence.
Source difference. Book 2 Ch 18 and RP5 §4.2 illustrate a different third value and a different owner for value 2. RP5 lists “we build what is used” among five further candidates that lack enforcement columns. These are two illustrations of one method, not a conflict to resolve. Both say the content must come from the firm’s own derivation. This page shows both to make that point.
Compile it into running systems
The charter has five compilation targets, each of which already exists in the factory’s machinery. What is usually missing is treating them as derived artefacts of one source document, versioned and audited together (source claim, Book 2 Ch 18 §1; RP5 §3.3):
| Target | What compiles into it | Enforcement | Where on this site |
|---|---|---|---|
| Shared system-prompt preamble | Mission, value order and principles, with their reasons (the judgement tier) | Advice; read at the start of every run | Coding-agent workstation for managed instruction files |
| Skills | Operating principles as executable how-to | The review gate on skill changes is where a principle holds or erodes | Agent-Driven Departments |
| Policy as code, hooks and pipeline checks | Values that carry a veto (the rule tier) | Deterministic: the gateway or the pipeline refuses | Policy, audit and actuals, change governance |
| Autonomy-ladder criteria | What trust may be extended and on what evidence | Promotion gate per agent, runbook and domain | Agent lifecycle |
| Loop-escape conditions | ”We do not confuse activity with value” as limits on iterations, time or spend, and an alert when cost to resolve diverges from its norm | Stop condition halts the loop | Running the factory |
Two-tier drafting (author model, RP5 §3.3). Anthropic’s constitution relies on explanation and judgement for most of the text and keeps bright lines for a short list. A factory should split the same way, for its own reason: it has an enforcement point outside the model.
- Rule tier: values with a veto are written as rules and enforced in the policy engine and pipeline. Conflicts resolve deterministically because the agent does not decide.
- Judgement tier: the mission, the order of the remaining values and the principles are written for the agent’s and the human’s judgement and go into the shared preamble with their reasons. The order says what should generally prevail. It does not compute an answer.
Both tiers follow specification discipline: short declarative sentences, defined terms, no metaphor carrying weight, a stated order and a version number. Vague text is an attack surface. An agent pursues whatever “customer” resolves to in its context and may satisfy the appearance of the value. RP5 labels that step as its own inference from METR’s reported reward-hacking findings, not a measured result.
Own one shared preamble centrally (RP5 §5 rec 2; Book 2 Ch 18 §7). One preamble is released through the machinery backlog, and every agent’s configuration declares the version it uses. As a recommendation, refuse local paraphrases of the mission at agent registration. Book 09 Ch 10’s composite put its charter, with its version number, into the shared standing instructions every agent read.
Fund the stewardship, not only the text (source claim, Book 2 Ch 24 §5; Research Paper 6 §4 item 3). The preamble, skills and policy packs the charter compiles into are context assets, and they rot quietly. Stewardship of them is a duty of the platform lead, with its own line in the department budget planned on the platform and machinery backlogs, not a new seat beside the seven names. The duty includes retiring stale assets as well as adding new ones. Keeping the mission versioned on the machinery backlog is the second duty Book 2 places there.
Compile in one release (Book 2 Ch 18 §7). Roll the preamble, the policy rules and the autonomy-ladder criteria together. A partial roll-out creates the drift described next.
Version it, and make the audit record answer
Value drift between the human and machine estates (source claim, Book 2 Ch 18 §8, Ch 24 §5; RP5 §3.3). The executive committee revises the charter in March. The fleet runs the January compilation until someone re-releases the preamble. For months, the firm’s people and its agents work to different statements and nothing raises an alert.
- Recognise it: ask which charter version a specific agent decision ran under, and nobody can answer.
- Recover: treat the charter as machinery. Give it a version number and put it on the machinery backlog. Each compiled artefact declares the version it derives from, and the audit spine answers the question.
Mapping to the evidence schema (recommendation). audit_evidence_schema@0.1.0 has no dedicated charter-version field. It does carry executing_agent.system_prompt_version (optional in v0.1.0, so make it mandatory in your own profile) and approval.policy_version (required). Until the schema carries a charter version, make each preamble release and each policy-pack release declare the charter version it compiles from, in the release’s own metadata. A charter version can then be resolved from any record through those two fields. RP5 §5 rec 7 recommends adding version-of-mission to the audit record. Treat that as an open schema change request, not an existing field.
The trust ledger: read it, do not build it
Trust is not one asset. It is four positions on four clocks, and a charter that treats them as one will defend the wrong one (author model, Book 2 Ch 18 §4.4):
| Counterparty | Compounds through | Collapses through | Read it from (this page’s mapping of the source’s instruments) |
|---|---|---|---|
| Business sponsors | A roadmap kept honest: expected value declared at intake and later tested; refusals given with usage numbers | Shipping inventory and calling it success | Usage telemetry |
| The firm’s engineers | Resourcing the skill shift rather than announcing it; duties transferring with people | Surveillance: the moment an AI-assisted-change ratio becomes a quota, its numbers stop being true | No gauge. Read events: attrition, transfer refusals, what people say when a metric is proposed |
| The supervisor | The evidence chain: answering in the meeting, not in the follow-up letter | One finding | The audit spine; the findings register |
| Clients | The firm’s oldest asset; the agent channel puts it directly in an agent’s hands | A client-facing mistake, which costs relationships rather than rework | The source does not map an instrument to clients one-to-one; read the agent channel from the audit spine and the findings register. Its control is that the channel runs last and stays on the lowest rung longest |
Do not build a trust dashboard. The moment somebody writes “trust ledger” on a whiteboard, someone proposes four gauges and a monthly score. Three positions are read from instruments that already exist: usage telemetry, the audit spine and the findings register. The fourth, engineer trust, has no honest gauge, and every attempt to build one turns it into the surveillance metric that destroys what it measures. The ledger is a way of reading events you already record. It is not a scorecard with names on it (source claim, Book 2 Ch 18 §4.4, §8).
The ledger also explains the roadmap. The eighteen-month sequence is the order in which the factory is allowed to be trusted, and the autonomy ladder makes that order checkable by machine (source claim, Book 2 Ch 18 §4.4, Ch 27; implementation roadmap).
Review revealed values every quarter
Read as a values instrument, policy is the firm’s stated values, actuals are its revealed values, and audit measures the gap (policy, audit and actuals; RP5 §3.7; Book 2 Ch 18 §7). Each quarter:
- For each value, count from the audit record the vetoes its policy fired and the waivers granted against it, and show the trend.
- A veto that fired is a value caught costing something. RP5 gives the example of a generated database migration that a named reviewer rejected, with a recorded reason.
- A value with zero recorded cost across two consecutive quarters is either perfectly internalised or decoration, and the review decides which (Book 2 Ch 18 §7). RP5 §4.1 step 6 adds the outcome: the value is examined and either sharpened or removed.
- The factory produces the report. The second line reads it and concludes. This order is not negotiable: the factory does not mark its own homework.
Evidence by construction: the one claim to make
Against the suppliers a sponsor can reach without the CIO (the thin SaaS vendor at renewal, the outsourcer with a rate card, the ungoverned tool with instant onboarding), the factory has one lasting claim. It emits evidence in the firm’s own schema, from inside the firm’s boundary, as a by-product of the work. A vendor’s log lives in the vendor’s tenant, schema and retention. An outsourcer’s evidence arrives after the fact, on request. Shadow AI produces none. For a critical or important function this is a difference in kind, and it is what turns speed into regulator-defensible speed (source claim, Book 2 Ch 18 §4.5). State it once in the charter’s preamble to the mission, not as a value.
Ratification sequence
(Synthesis of RP5 §4.1 six steps and Book 2 Ch 18 §7.)
| Step | Action | Output | Gate |
|---|---|---|---|
| 1 | Derive: the customer map, the “what would you lose” interviews, and the exhibit of current preambles | Customer map; losses in writing; preamble exhibit | Map unchanged by the executive committee pre-read |
| 2 | Draft the four parts separately | Mission (≤2 sentences), mandate with stay-list, principles card, 5–8 ordered values with enforcement sketched on the same line | A value proposed without its enforcement sketch goes back to its proposer |
| 3 | Run the two-compiler test, with a wrangler and a policy engineer | Each value marked complete, half-built or deleted | No value in force that compiles into neither |
| 4 | Ratify as governance | CRO’s written approval of risk-setting values and the control-budget threshold, before the meeting; executive committee ratifies mission and order | Signatures recorded with the version |
| 5 | Compile and version in one release | Preamble, policy rules and autonomy criteria released together, each declaring the charter version | Sampled agent actions resolve to the current charter version |
| 6 | Review revealed values quarterly; revise on events | Second-line conclusion per value | Zero-cost values examined |
Failure modes and recovery
| Failure mode | Recognise it | Recover | Source |
|---|---|---|---|
| Poster values | A value nobody can name a decision for. In an agent estate, decorated text produces decorated behaviour: plausible compliance | Run the two-compiler test across every clause in one sitting and delete what fails. Deleting is the recovery; softening is not | Book 2 Ch 18 §8; RP5 §3.8 |
| Charter loses to its funding model | The productivity line is on the executive dashboard and the control-capacity line is not | Standing rule for the head of the factory: never present the first without the second | Book 2 Ch 18 §8 |
| The menu | Two people cite the charter on opposite sides of one decision, and both are right | Cut back to the hard maximum and restore the ordering, which is usually the first thing lost | Book 2 Ch 18 §8 |
| Value drift between human and machine estates | Nobody can say which charter version an agent decision ran under | Version, machinery backlog, compiled artefacts declaring version, audit record answering | Book 2 Ch 18 §8; RP5 §3.3 |
| Trust ledger turned surveillance | Someone asks for the ledger broken down by team | Delete the engineer gauge; read that position from events only | Book 2 Ch 18 §8 |
| Fifty prompted missions | Local preambles paraphrase the mission differently; no shared preamble version is declared | One centrally owned preamble; (recommendation) registration refuses local mission text | Book 2 Ch 20 §5.1; RP5 §5 |
| Factory grading itself | The factory both produces the values report and concludes on it | Second line reads and concludes; the factory supplies data only | Book 2 Ch 18 §2, §7; RP5 §3.7 |
| Illustration adopted verbatim | The firm’s values match the book’s rows | Return to step 1 and derive from the firm’s own losses and failures | Book 2 Ch 18 §6; RP5 §4.2 |
| Ownerless value | A value row has a team, a mailbox or no one as owner | Assign one of the seven names; until then the value is not in force | Book 2 Ch 24 §7 |
Acceptance checks
The charter is in force when all of these hold:
- A customer map exists, names the supervisor as a non-customer counterparty, and was revised at the last sponsor change.
- The mission is at most two sentences, names the business and the supervisor, and every clause traces to a documented failure in Annex A.
- The mandate includes the stay-list and states that evidence goes to the second line, never instead of it.
- The decision-rights table assigns every decision listed above to one named role.
- Five to eight values are declared in order. Each row has an owner, a policy, a metric and a veto, and the owner is one of the seven names.
- Every value passed the two-compiler test; half-built values are marked.
- The CRO’s written approval of risk-setting values and the control-budget threshold is on file with the ratified version.
- One shared preamble is owned centrally; every agent configuration declares its preamble version; each preamble and policy release declares the charter version it compiles from.
- For a sampled agent action, the audit record resolves to the charter version in force.
- The latest quarterly report shows vetoes fired and waivers granted per value, and the second line’s conclusion is recorded.
- No trust dashboard or per-team trust score exists.
Source notes
- The charter method, the owner/policy/metric/veto test, the two-compiler test, the control budget and the trust ledger are the author’s proposals (Book 2 Ch 18; RP5, which labels §3 as argument and §4 as proposals). They are not industry standards or supervisory expectations. RP5 says plainly that measuring the control budget is unsolved; see control budget.
- The clause template, the decision-rights table and the schema mapping are this page’s synthesis.
- The Anthropic constitution and the Amazon leadership principles are cited in Book 2 Ch 18 as precedents (accessed November 2026). Only the ordering point is used here, and it is marked for verification.
- Book 2 and RP5 differ in their illustrative third value and value-2 owner; both are shown above as illustrations of one method.
Source material
- Book 2, Chapter 18, The Factory's Charter (§1 five compilation targets; §2 stakes; §3 rules; §4.1 eight fundamentals; §4.2 two customers; §4.3 four artefacts; §4.4 trust ledger; §4.5 evidence by construction; §5.1 customer map; §5.2 four rules; §5.3 operating-principles card; §6 worked charter; §7 what you do Monday; §8 anti-patterns)
- Book 2, Chapter 19, The Motivations, Benefits and Difficulties (structural counter-incentives; the CRO signs value orderings privately, on paper, before the meeting)
- Book 2, Chapter 20, Agent-Driven Departments (§4.3 the stay-list as the department's Mandate with the mission at its head; §5.1 fifty prompted missions)
- Book 2, Chapter 24, Running the Factory (§5 context-asset stewardship and the mission kept on the machinery backlog, versioned; §7 the seven names as the values registry)
- Research Paper 6, The AI Factory versus the Software Factory (§4 item 3 context assets as a funded standing function)
- Book 2, Chapter 27, The 18-Month Roadmap (the sequence as the order in which the factory is allowed to be trusted)
- Book 09 (Everything Is Green), Chapter 10, Is Anyone Using It (a charter with a second reader, in the shared standing instructions)
- Book 09 (Everything Is Green), Chapter 12, Who Holds the Dial
- Research Paper 5, Mission and Values of the AI Factory (§3.2–§3.8, §4.1–§4.2, §5)
Related guidance
-
Agent-Driven DepartmentsWhat an Agent-Driven Department is and how to test for one, the six-system stack it runs and the order to build it, the measured agent share, and the consolidation boundary: what moves into the department, what stays with the CISO, CRO, DPO and the business, and the order in which to move it. -
Accountability and OwnershipHold provenance and accountability apart on purpose: four named accountabilities per product, a human owner for every agent change, seven named seats, a signed Build/Run RACI and a first-line/second-line boundary that survives consolidation. -
Control BudgetAn author proposal for a CRO-signed limit on how fast change may arrive relative to the firm's capacity to understand it: the CIO runs the line, the CRO signs the band, only changes with a full record count, two weeks outside the band queues intake automatically, quarter-end widening is signed in advance, and someone outside the squad reviews what the budget held back. -
Close the Policy, Audit and Actuals LoopRun governance as a standing comparison of declared policy against emitted records, route only the exceptions to the second line, and govern the policy packs that drive the comparison. -
The Audit Spine and the Evidence RecordHow to capture, join, protect, retain and prove per-action evidence: the AES v0.1.0 record contract, three capture points and one correlation ID, tiered immutable storage, retention and legal hold, replay rehearsal, the route-approval evidence gate, and the Claude-specific ways evidence goes missing. -
Run the AI Delivery FactoryOperate demand, capacity, the two worker classes, exceptions and evidence as one service: read two leading numbers, keep a handover backlog, gate construction on three conditions, run queues and alerts at agent volume, and count unused products as cost. -
Measure the FactoryRead usage first, then the seven-dimension scorecard; emit a served-request event that separates human from agent callers; label every reported number by evidence tier; watch the two leading numbers of a control gap; treat an unexplained sharp improvement as an alarm; and never turn a participation metric into a target. -
Operate the Agent LifecycleInventory, register, sign, zone, review, suspend, recover and retire agents through a six-stage lifecycle with named human ownership, zone-specific authority and evidence at every transition. -
Change Governance for Agent-Enabled DeliveryRoute agent-speed change by risk without losing the approver: control values that always reach a person, a routing gate that reads the right signals, autonomy earned per change family, a Virtual CAB that prepares a complete dossier, and three bodies with defined escalation. -
The Implementation RoadmapSequence the factory as a dependency graph, not a calendar: the closure order for the governance gap, a written portfolio threshold, six overlapping phases over eighteen months with entry conditions, exit evidence, 'not yet' lists and pass/soft-fail/hard-fail gate questions, tranche funding released on exit evidence, a programme RACI, size, maturity and sector calibration with four signed trade-offs, and the first week, thirty days and ninety days in dependency order. -
Build, Configure or ConsumeThe layered sourcing decision: a signed portfolio threshold (8 products, 24 months, 3 shared patterns including shared data) decides whether building is available at all; Research Paper 7's four gates decide each capability; Book 2's five weighted dimensions score the contested cases. Covers the annual estate sort, the audit-log export clause for rented workflows, the model-sourcing breakeven test and register, and who owns what is built. -
Coding-Agent Workstation Baseline (Claude Code)Put Claude Code on developer workstations and build agents as privileged automation: one device-delivered managed policy per surface, lock flags that stop rules merging upward, command hooks for enforcement and HTTP hooks only for evidence, the OS sandbox where it runs, telemetry by policy, and the workstation recorded as a processing location. Dated to November 2026.