Authority follows the cost of a mistake.

Move quickly while mistakes are cheap.Add controls as value and consequence rise. A capable agent is not an unlimited authority.

Process should scale with what can be lost.

A throwaway experiment should not carry commercial process. Customer money, private data, valuable work and physical safety change the required evidence.

Autonomy should scale with reversibility. Governance should scale with value and consequence.Core doctrine
Choose controls for the actual consequence. These are classes, not stages a project has to climb, and classification is not an automatic risk score.
Value classes and controlsChoose controls for the actual consequence. These are classes, not stages a project has to climb, and classification is not an automatic risk score.View the system model (D2)
Read the diagram in text
  • Class 00 · ThrowawayExplore freely; discard cheaply.
  • Class 01 · ExperimentPreserve learning; add useful checks.
  • Class 02 · ProjectProtect source, repeatability and records.
  • Class 03 · ProductProve users, access and production quality.
  • Class 04 · CommercialControl money, promises and rollback.
  • Class 05 · TreasuredNarrow authority; prove recovery and safety.

Agents do the routine work. MP is not the default planner.

Most of this factory is agents reasoning. MP is the human principal and the final authority, but not the default planner, memory store, specification writer or decision engine. An agent that brings MP a choice between two defensible options has handed back work it should have finished.

What genuinely needs MP is a capability an agent does not have: taste, physical hardware, money, credentials, external accounts, legal judgement, an irreversible action, or a gate this factory has deliberately reserved. Everything else is the agent's work, carried to completion and shown rather than proposed.

The default loop

  • Reconstruct current truth from the repository and the running product.
  • Treat the issue as a claim about a desired outcome, not as proof that the proposed fix or the existing test is correct.
  • For user-facing work, inspect or play the product before changing it.
  • Find the smallest material discrepancy blocking the outcome, and take it to completion.
  • Validate at the cheapest meaningful level, then at the integrated level its risk earns.
  • If a gate conflicts with current product intent, investigate the gate before changing the product to satisfy it.
  • Re-inspect the owning outcome after integration. If it is still incomplete, leave the parent open and name the next real discrepancy.
Done means the observable outcome exists at the authority level the project claims. Making the change is not the same thing as finishing the work.Core doctrine

A prompt can end with a real object.

The factory can compose software with a payment provider, a printer, a manufacturer, a contractor or a person doing work that software cannot. MAX makes the boundary obvious: a working PWA is not a real walker. The hardware, mounting, microphone, speaker, lights and the person's ability to use FIND still need physical proof.

The same route could take approved artwork to a poster, shirt or book through a fulfilment provider such as Printful. Agents can prepare the files, product details and order workflow. A provider makes and ships the object. The factory records what happened and what still needs a person to check.

An eligible event could ask an AI system to draft artwork, copy, a diagram or a product variant. That output stays a candidate. Check the brief, rights, visual quality, print dimensions and cost, then ask Dean to approve, return or reject it before it becomes a listing.

Before work crosses that boundary, confirm that it is legal, the assets may be used, the provider is authorised, the budget is available, the result is safe and the outcome can be verified. If any of those answers is missing, stop at a proposal.

Events can start work. Authority still has limits.

A milestone, failed check, shipment event or scheduled review can ask an agent to prepare the next change. Production access starts read-only.

Events, scheduled checks and safe production observations can create bounded work. Agents check scope, rights, cost and recovery. Ordinary work proceeds within existing authority; reserved account, money, rights and irreversible decisions go to MP. Existing product and release gates still apply.
Authority has a boundaryExisting authority permits routine work. A reserved decision applies only to its approved scope; a revision repeats the affected checks.View the system model (D2)
Read the diagram in text
  1. Event or scheduled reviewA release, failed check, order, shipment or recurring audit may start a bounded task.
  2. Read-only productionAgents gather permitted evidence. They do not turn an observation into a production write.
  3. ProposalThe output is a draft, issue or changeset with evidence, rationale, risk and rollback.
  4. Policy and qualificationCheck scope, asset rights, spend, required gates and recovery. Work already within authority proceeds and is qualified; an event does not grant new permission.
  5. Reserved decisions go to MPAccount access, money, rights, irreversible actions and explicit owning gates need their actual decision. Approve the exact scope, return it for revision and new checks, or reject it and stop. Agent approval does not replace that decision.

Scheduled improvement works the same way. A check can spot an old dependency, stale screenshot, broken journey or drifting document. It creates a proposal or work item. It does not silently edit or deploy production.

Project value sets the amount of process. A throwaway experiment stays cheap to change. Customer money, promises, private data, hard-to-replace work or physical safety add controls when a project earns them. A repository count or an AI-generated market estimate does not make something a company. Record actual customers, revenue, use, replacement cost, creative value and operating risk separately.

A payment event must lead to one controlled fulfilment.

A browser redirect cannot settle an order. Payment state comes from the payment provider, then a durable workflow records the fulfilment that follows.

Dean approves the catalogue item, artwork rights, price and spend limits. The customer pays through a payment provider. A validated, deduplicated event updates the order ledger and may create one Printful order within those limits. Shipment events update the same record.
One paid order, one fulfilment pathReference architecture, not a live shop. Retries and duplicate events must not create a second charge, order, shipment or refund.View the system model (D2)
Read the diagram in text
  1. Approve the catalogueDean approves artwork, rights, price, provider and per-order spend limits before a product is offered.
  2. Confirm payment from its sourceVerify the webhook signature, payment status, amount and order. A return to a success URL is not proof of payment.
  3. Deduplicate before actingRecord stable event and order IDs. Use idempotent provider requests so a retry cannot make a second order.
  4. Fulfil within policySend Printful an approved order only when the item and spend fit the published limits. Hold exceptions for Dean.
  5. Follow the shipmentProvider events update the order record and customer notification. Physical product quality still needs a real check.

A commercial pipeline can automate orders inside limits Dean approved in advance. A new product, unusual amount, refund, supplier purchase or other exception waits for a decision. The approval is durable and records whether Dean approved it, returned it with notes or rejected it.

Validate events, make handlers idempotent and keep an order ledger. A retry must not charge twice, buy two shirts, send duplicate refunds or tell a customer that an unshipped order shipped. Payment details stay with the payment provider; Woodhouse's order record needs the provider's event and transaction references, not card numbers.

Technical references: Stripe payment events · Stripe idempotency · Printful orders and webhooks · Cloudflare human approval workflows · Cloudflare scheduled agent tasks.

Scope: Woodhouse publishes these rules. This release does not process payments, call Printful, schedule agent work or write to production.

Observe more than you mutate.

Agents can collect production evidence and prepare a change. Before a consequential action, the change needs a reviewable diff, a reason, qualification and a rollback path. The approval is tied to that exact candidate and cannot be skipped by asking again in chat.

01

Observe

Read permitted state and source material.

02

Recommend

Explain what changed and what remains uncertain.

03

Prepare

Produce a reviewable change with evidence and rollback.

04

Execute by policy

Bounded work may proceed inside approved limits. New financial, destructive or physical actions wait for a human decision.

Failures teach the factory.

Contain the failure. Record it as an incident or tracked defect, attach the evidence, and investigate which control missed it. Link the work item to the changeset, qualification and production follow-up. Then decide whether the factory needs a regression check, release gate, policy, architecture or documentation change.

A recurring failure should become harder to repeat because the factory learned from it. Public claims remain smaller than the evidence supporting them.

Give agents room to work where mistakes are cheap. Add review where a mistake can cost users, money, data or safety.Core doctrine

A gate is evidence for a requirement, not the requirement.

Checks accumulate. Some outlive the behaviour they were written to protect, and a factory that treats an outdated gate as a specification will warp a correct product to satisfy a test that no longer describes anything real.

When a gate and current product intent disagree, investigate the gate first. Ask whether the requirement still exists. If it does not, delete or rewrite the gate and record what it used to protect. If it does, the product was wrong. What is not acceptable is deleting the requirement quietly to make the disagreement disappear.

The same bar applies to process work itself. New process, tooling or architecture needs a named observed failure behind it. A control added because a check could not actually run against production, or because one worktree could certify another, has earned its place. A control added because it seemed prudent is a cost with no evidence attached.

Spend against evidence, not against possibility.

Money is the one input in this factory that cannot be un-spent, and the temptation it creates is specific: a capability that could be paid for feels available, and availability is easy to mistake for readiness. The rule is a sequence, not a mood.

  1. Preserve the option. Keeping a path open is cheap and keeps the choice real.
  2. Validate cheaply. Find the smallest experiment that could be wrong.
  3. Increase spending only as evidence increases.

Applied honestly, this rules out a whole category of expensive mistake. Capital is not wasted by waiting for a technology that makes a project possible; it is wasted by spending early to find out whether a market exists. A factory that can already build the system has demonstrated one thing: it can build systems. That says nothing about whether strangers will pay for them, and the two claims are routinely confused because both feel like progress.

Technical maturity and commercial validation are separate claims with separate evidence. A deployed product, a passing suite and a growing catalogue are implementation, release and production verification. None of them is a customer. Where the only honest measurement is an empty table — the tables that would hold a rating, a report or a dispute, with no rows — that is a measurement worth publishing rather than a gap to route around.

This also constrains the negative direction. Creative work is allowed to be worth doing without a business attached, and a project with no commercial ambition is not a project awaiting justification. Forcing a revenue thesis onto work that has none produces a worse answer than leaving it alone, and manufacturing the appearance of demand is worse than recording its absence.

Preserve optionality, validate cheaply, and increase spending only as evidence increases.Core doctrine