Decision Optimisation Engine
How does the Decision Optimisation Engine make one decision improve a later decision?
Problem: decisions are made and reviewed, but their lessons rarely change the next choice.
Question: how does one decision become precedent that improves a later decision?
Decision: run a bounded decision-learning cycle whose north-star gauge is agency gain—greater human capacity to choose, act, learn, and adapt.
The Decision Optimisation Engine connects the inner atom—Problem → Question → Decision—to action, review, retained precedent, and recall. It replaces isolated decisions and passive records with a testable return loop.
Start here when you need the complete cross-cycle mechanism. If one live choice needs attention now, navigate that decision, then return with its outcome and review. If the problem or choice is still unclear, use the Decision Making hub to choose the right route first.
Decision Optimisation Engine
The core move begins when the inner atom turns reality into a committed move:
Problem → Question → Decision
- Problem names the observed gap, not merely its visible symptom.
- Question names what must become clearer to change the choice.
- Decision commits a move, an owner, and honest bounds.
That atom navigates one choice. The engine makes the choice teach the next one:
Problem → Question → Decision → Action → Outcome
↑ ↓
Recall ← Precedent ← Learning ← Receipt ← Review
The return path matters as much as the outward path. Review compares prediction with observation. A receipt preserves the variance and lesson. A lesson becomes precedent only when a later decision recalls it and states how it changed the reasoning, choice, threshold, or an explicit no-change verdict.
Agency Gauge
Speed, correctness, and output volume are useful supporting measures. None is a sufficient optimization target.
| Target alone | What it can hide |
|---|---|
| Faster decisions | Repeated mistakes, premature commitment, or authority silently delegated to a model |
| Correct outcomes | A lucky result produced by weak reasoning—resulting bias |
| More decisions or outputs | Activity that creates no worthwhile change for the beneficiary |
Agency gain asks a harder question: did this cycle increase the human's capacity to make and adapt the next consequential choice? A slower decision can produce more agency when it clarifies values or prevents irreversible loss. A fast reversible experiment can produce more agency when it earns evidence cheaply. Decision quality and velocity serve the gauge; they do not replace it.
Connected Practices
These practices operate at different resolutions and keep their existing jobs:
| Practice | Job | Boundary |
|---|---|---|
| Navigate a Decision | Navigate one live decision | Produces a proposed artifact, steering contract, confirmation, and review |
| Decision Journal | Retain and calibrate a decision trace | Preserves the reasoning and result; storage alone does not prove learning |
| Decision Optimisation Engine | Connect decisions across cycles | Requires recalled precedent to show an effect on a later decision |
| Software Factory | Produce decisions at organisational or platform scale | Routes context, capability, evidence, and learning without taking human authority |
The engine is not another journal, a method for automating judgment, or an organisational factory. It is the learning relationship between a decision and a later decision.
The Public Contract
Use the existing navigation contract rather than inventing a second ontology.
Before deciding, record these fields in order:
problem
question
decision
priority
valueSetpoint
beliefHorizon
controlLoop
proofSignal
agencyGain
The steering contract adds lever, trigger, loss limit, review point, and kill signal. Together they state what can change, what starts action, what loss is acceptable, when judgment returns to the evidence, and what stops the move.
At review, retain a receipt with:
prediction
observation
variance
lesson
later-decision effect
The receipt remains pending evidence until a later decision demonstrably consumes it. A filled record proves that the trace exists. It does not prove that reasoning improved, the decision was better, or the beneficiary gained value.
Copyable Protocol
1. Recall relevant precedent, or explicitly record that none exists.
2. Frame the observed problem rather than its symptom.
3. Ask the question that would change the choice.
4. Commit the smallest honest decision and classify its door.
5. Define lever, trigger, loss limit, proof signal, review point, and kill signal.
6. Act through the smallest falsifiable move.
7. Compare prediction with observed outcome without resulting bias.
8. Retain the lesson as precedent.
9. Require a later decision to cite how the precedent changed reasoning, choice,
threshold, or an explicit no-change verdict.
For a reversible door, keep the move cheap and falsifiable. For an irreversible or compound door, slow down: verify the state, compare credible options, expose assumptions, and make the loss visible before committing.
Authority Boundary
The human captain owns values, authority, consequences, and the final judgment. The agent First Mate may frame the problem, retrieve precedent, expose missing fields, compare options, test the contract for completeness, and propose a correction.
The First Mate must not infer consent. Every consequential decision requires an explicit captain response: reject, revise, or confirm. Silence, copying a draft, or continuing the conversation is not confirmation. No confirmed decision means no action receipt.
Onboarding Example
This example demonstrates the method. It is not customer proof.
Situation. A human product owner sees many new users abandon onboarding at identity verification. The visible symptom is a low completion rate. The framed problem is narrower: first-time users reach verification but cannot tell why the information is required or what happens next.
Decision-changing question. “Is unclear explanation, rather than the verification requirement itself, causing enough abandonment that a clearer step would materially improve successful activation?”
Proposed decision. Run a reversible experiment for 20% of eligible new users. Change only the explanation and progress cue; do not redesign the verification system.
| Contract part | Example |
|---|---|
| Lever | Explanation and progress cue on the verification step |
| Trigger | Captain explicitly confirms the experiment and privacy review passes |
| Loss limit | No more than 20% of eligible traffic for 14 days |
| Proof signal | More users complete verification and reach first value, without higher support or privacy complaints |
| Review point | 14 days or the minimum predeclared sample, whichever is later |
| Kill signal | Material rise in complaints, errors, or abandonment at the next step |
| Beneficiary gauge | New users reach first value with less confusion and no reduced informed consent |
The agent retrieves any comparable onboarding traces, checks that the contract is complete, and proposes the experiment. The human owner confirms or revises it. At review, the team compares the prediction with observed activation, support, and complaint signals rather than declaring success because one conversion number rose.
Suppose completion rises but activation does not. The retained lesson is: clarity reduced local friction, but verification was not the activation bottleneck. That receipt is still pending evidence. It becomes precedent when the next onboarding decision cites it—for example, by raising the threshold for another copy-only test, choosing to investigate the post-verification handoff, or recording that the old lesson does not apply because the audience has changed.
Evidence Limits
The public evidence is deliberately bounded:
- Demonstrated: typed validation exists for the nine-field navigation artifact and steering contract.
- Informed inference: recalled precedent can change a later decision when the later trace names the effect on its reasoning, choice, or threshold.
- Not yet proven: repeated compounding across broad decision classes, sustained agency gain, or customer value.
The engine therefore remains awaiting_maturity: the method and next-run assertion exist, but
comparable later cycles have not yet proved lift. Broad improvement requires repeated decisions to
show that recalled precedent changes later choices and improves beneficiary outcomes against a
frozen gauge.
The model also breaks when the problem frame is false, the beneficiary is absent, the receipt is written after the result to justify the choice, precedent is treated as a rule despite changed conditions, or human authority is ceremonial rather than real.
Changes my mind: repeated, comparable cycles show that recalled precedent does not improve human agency or beneficiary outcomes beyond a simpler review-and-journal practice.
Retrieval
Pull this page when a decision should learn from prior choices, when a journal contains traces that are not being reused, or when a workflow claims learning without showing a later-decision effect.
Version delta: the public decision route now distinguishes the inner atom from the surrounding cross-cycle engine and defines when a receipt becomes evidence of learning.
Context
- depends-on Navigate a Decision — run the public contract for one live choice.
- pairs-with Decision Journal — retain the trace needed for review and recall.
- applies-to Software Factory — scale the route only after the learning loop works manually.
- proved-by Reality — compare claims with observed evidence and beneficiary outcomes.
- pairs-with Performance — keep the agency gauge connected to observable outcomes.
- depends-on Purpose — keep optimisation subordinate to human values and worthwhile intent.
- instance-of VVFL Evolution — run the smallest decision-learning cycle inside the wider virtuous feedback loop.
- explained-by Flow — understand why action without a returned receipt cannot compound into agency.
Questions
Next question: which later decision will be required to consume this receipt, and what minimum change would count as agency gain?
- Which precedent is relevant, and where does the analogy break?
- What remains human-owned if an agent prepares every field?
- What beneficiary signal prevents faster decisions from becoming the wrong optimization?