Workbench file / 003-A
Live instrument / Deterministic
Decision Compiler
What should AI do here—and what would it need to prove?
This public instrument turns operating constraints into an explicit posture, first test, safeguards, and a reason to stop. It never predicts success.
The Decision Compiler / live instrument
Compile an AI product, not another demo.
Set the operating constraints. The rule system turns them into a product posture, evidence plan, guardrails, and an explicit reason to stop.
System trace
Every product constraint changes something downstream.
- 01DecisionStatus: readyRecommend
- 02WorkflowStatus: readyDaily / Team
- 03ModelStatus: readyReviewable to verify
- 04GuardrailsStatus: watchMaterial error cost
- 05FeedbackStatus: readyDays
Change one constraint to see which product decisions move with it.
The rule system is deterministic: the same constraints always produce the same operating plan.
Compiled posture
Recommend
Let AI propose a next move, but make evidence and user judgment part of the product.
Why this posture
- Errors create meaningful cost, so user judgment must remain inside the operating loop.
- Skilled review is possible, so recommendations can be tested without hiding uncertainty.
- Delayed feedback requires disciplined reconciliation rather than proxy-only success.
- First experiment
- Shadow 50–100 representative decisions, then expose ranked recommendations with evidence to a small expert cohort. Compare decision quality, elapsed time, and correction work against the current baseline.
- Evaluation stack
- Representative task set anchored to the current human-workflow baseline
- Expert rubric, evidence grounding, and reviewer-disagreement analysis
- Delayed outcome reconciliation segmented by workflow and user behavior
- Weekly slice review across user, task, and failure type
- Guardrails
- Show supporting evidence and preserve explicit user approval
- Make reject, edit, and escalation paths first-class product actions
- Adoption requirement
- Agree on a shared review rubric, workflow owner, escalation path, and team-level adoption measure before rollout.
- Kill condition
- Stop if recommendations fail to improve decision quality or time-to-decision after accounting for added review work.
Margin notes / 03
How to read the file.
- 01Human decision before AI capability
The intended human decision and behavior change come before AI architecture or model selection.
- 02Evidence before AI autonomy
AI autonomy is earned only when errors are recoverable, outputs are verifiable, and feedback is fast.
- 03AI operations are part of the system
Ownership, escalation, adoption, monitoring, and rollback are treated as first-class AI surfaces.
This is a deterministic AI decision aid, not a readiness score or substitute for discovery. The same inputs always produce the same recommendation, and every rule is inspectable in the source.