Decision model

Decision Clone

I use the decision clone to make the private rehearsal harder than the real conversation, then record what it predicted before I know the outcome.

Decision Clone is where I rehearse difficult decisions and pressure-test the package. It records what the model predicted before the real meeting.

Why I built it

I was carrying decision packages into important conversations with arguments that felt complete until the first hard question exposed the assumption I had skipped.

Agreement was the dangerous failure mode for a decision clone because easy approval would make the rehearsal feel good while leaving the package untested; constant objection would be just as useless once the model hardened into a caricature.

Decision Clone prediction receipt using synthetic decision data — FIG. 01 - PREDICTION RECEIPT, LEDGER FORMAT, SYNTHETIC DECISION DATA

How it works

I built four rehearsal modes: interrogate, red-team, predict, and rehearse. The decision clone runs deliberately hawkish because I would rather absorb the harder version of a question in private and arrive with the answer.

Every prediction enters an append-only receipt ledger before the real interaction, and later evidence resolves verdict accuracy, question precision, condition accuracy, confidence, and surprises as separate dimensions. The scorecard cannot update the model on its own.

Design choices

Bias toward harder questions

I want the private rehearsal to expose the gap before the real room does, so I tune the clone to apply pressure rather than reward a polished package.

Record before resolution

I make prediction receipts immutable once written so the later outcome cannot quietly rewrite what the model originally forecast.

Keep dimensions separate

I keep verdicts, questions, conditions, confidence, and surprises separate so a correct question prediction cannot be laundered into a correct decision forecast.

Decision Clone calibration summary showing separate accuracy dimensions with synthetic data — FIG. 02 - CALIBRATION SUMMARY, SEPARATE ACCURACY DIMENSIONS, SYNTHETIC DATA

Block self-modification

I allow calibration evidence to support a proposed change, while every update still requires a separate review and an explicit promotion decision.

Using the system

The model can become a caricature of the person it represents. I use the rehearsal to test my reasoning and compare predictions with what happened afterward.