REF.04 own project · psychology.games · 2026
A taxonomy whose rules were put through a blind reproducibility test
A classification of psychological games, phenomena and methods in which every category boundary is a yes/no rule — and in which those rules went through a blind test, repair and re-measurement before earning a new version number.
- R1–R87 branches · 26 leavesCategory ≠ mechanism
- TESTBlind test (3 × 25 items)Hidden names · 2 model families
- v4→v5Agreement with intent20/25 → 23/25 (24/25 stable)
- Branches / leaves
- 7 / 26
- Agreement with intent
- 20/25 → 23/25
- Stable across 3 runs (v5)
- 24/25
- Boundary-probe readings
- 18/18
-
Problem
The domain mixes cognitive tasks, Berne's interpersonal games, manipulation tactics, video games and tests. Earlier attempts grouped them by branch of psychology, and the branches overlapped — the same object could honestly be placed in two homes.
-
Approach
The governing principle: a category describes what an object is; a mechanism describes why it works. Each item is first normalised to its subject of classification, then passes through eight ordered rules into one of 26 leaves. The rules were tested blind: 25 items with names hidden, three independent runs per model, worked examples removed, across two model families. Divergences were diagnosed as either a rule defect or noise, and the rules were repaired and measured again.
-
Outcome
Agreement with the intended answer rose from 20/25 to 23/25 against an acceptance threshold of 22/25. The test exposed a systematic error that no single run reveals: both model families agreed, stably, on the same wrong leaf. The rule change shipped as a governed v4 → v5 release with a change log and PL/EN kept in lock-step.
The most valuable result was not the high agreement but the place where two model families agreed with each other — and were both wrong. Run-to-run stability looked like proof of correctness; it was really proof that the rule led everyone to the same wrong place. You only find an error like that by comparing results with an intent recorded in advance, not just comparing runs with each other.
A second finding: one item can only be settled by the table of worked examples that a blind test deliberately removes. That means the rule is incomplete at that point — the example is standing in for it.
What this engagement does not cover
Every run was a model run, not a human one — the human editor pass is still pending. The rules were themselves drafted with a model, so these are models reading rules written by models.