Proof
Remove one belief. Replay the same event. The response disappears.
Astra reaches the same scheduled catch in two recorded versions, and her previous experiences remain in both. In one version she holds a belief about herself. In the other that belief has been cleared. The response appears only where the belief is present.
01 One run, two versions
Astra reacted in one version and carried on in the other.
Astra escaped three times and came to believe she was the fast one. Being caught twice changed her mind. Ember escaped two times without forming that belief.
The Believed stage at the second catch, where the two versions part.
The run as it happened
The belief is dropped: two contradictions in a row and Astra no longer holds it.
Inspect the recorded values
- trait swift
- conviction 0
- doubt 0
- held before yes
- held no
- times shaken 1
The run with one belief cleared
Nothing to contradict: Astra holds no belief about being the fast one, so the loss lands as an ordinary event.
Inspect the recorded values
- trait swift
- conviction 3
- doubt 0
- held before no
- held no
- times shaken 0
The two versions ran as separate processes, so their timestamps differ. The order of the events and their content do not.
What this supports
A persistent internal belief can be part of the causal path to a character response, in this recorded mechanism.
What this does not yet support
A general effect size, performance at population scale, or belief-driven choice of what the character does next.
02 What changed
Change the belief. Change the response.
Astra’s memories stayed. Her belief was removed before the catches. The game calculated the response again.
Both recordings use 2 characters, the same starting conditions and the same code path within a 90 second observation window.
The original logs and comparison file let you inspect the intervention and the resulting state changes.
What differs, in the run's own file
- Changed
- one self-belief row on Astra set to held=false, conviction unchanged
- Unchanged
- the same starting conditions, same world events, same schedule, same bodies, same code path
- Everything downstream
- re-derived by the running system in the cleared run's own process; no value in either file was edited by hand
03 The register
Six implemented. Seven being tested.
Twelve more are designed, not built. We judge every claim by one rule: withhold the claimed cause, hold the world constant, and see whether the behaviour changes.
Implemented , 6 claims
Working and tested in the setting named by the claim: saved game state or internal developer tools.
- Implemented
Event and autobiographical record
Each character keeps a bounded, curated record of what it has lived through, and that record changes how it develops.
- Implemented
Persistent individual identity
Each character keeps a stable identity, and that identity survives saving and loading.
- Implemented
Shared pair history
Two individuals share one bounded record of the history between them, written during ordinary play.
- Implemented
Grounded authored expression
What a character says is assembled from what it has actually lived, and it stays silent rather than claim something it cannot support.
- Implemented
Recording inspection, internal tool
Internal tooling opens saved recordings and shows each character’s recorded evidence, beliefs and choices at a selected moment.
- Implemented
Trace regression checks, internal tool
Internal tooling checks saved recordings against named rules, identifies failing records and reports when the evidence is insufficient to complete a check.
In production validation , 7 claims
Works in a bounded development setting and is undergoing broader validation.
- In production validation
Evidence held per character
Evidence is held per individual, and it gates what that individual is allowed to claim.
- In production validation
Belief-gated response
A character that has come to believe something about itself reacts when the world contradicts it, remembers the day, and lets the belief go; a character that never formed the belief does neither.
- In production validation
Appraisal and expectation revision
A character can hold an expectation, discover it was wrong, and be changed by that.
- In production validation
Belief-gated action
In a controlled scene, a character acts on a held belief about a friend, then revises it after observing a change.
- In production validation
CRI Engine prototype
An extracted decision kernel runs outside Luminids; the broader reusable CRI runtime remains in development.
- In production validation
CRI Studio prototype
An internal Studio prototype inspects and replays recorded character state, with a connected game showcase for following beliefs, choices and outcomes.
- In production validation
CRI Bench prototype
An internal Bench prototype changes supported inputs to recorded decisions, recomputes the result and checks recordings for regressions.
Proposed , 12 claims
The capabilities and tools we’re working towards.
Twelve more capabilities are designed and not yet built.
04 The tests
Test the character through play, change and saving.
The published suite snapshot is dated 11 October 2026: 32,281 tests executed, 99.82% passing, 59 known failures.
Some of the tests, as the suite names them
- a character remembers who raised it
- a name you gave is visible
- the hatch is witnessed
- care is witnessed
- every quirk is earnable
- a save never contains a statue
- the bench for two is raised by the colony
The published suite snapshot
- 32,281
- test cases run
- 59
- failures, included on purpose
- 6,300
- cases in the luminid character suite
- 4,304
- suite files
A 99.82% pass rate across the cases that ran, over about four and a half hours, three files at a time. Of the 4,304 suite files, 41 carry a failure and 55 ran out of time; 32 more passed only when run again.