AMNESIAC NATURAL VARIATION P0 PRE IMPLEMENTATION PROMPT

STATUS: DESIGN ONLY. DO NOT IMPLEMENT A UNIVERSAL PREFERENCE SCHEMA UNTIL RATIFIED.

Goal

Prove that Amnesiac can preserve changing human relations without treating temporal difference as correction/error.

Use three synthetic fixtures:

1. NV-01 One Song, Ten Hearings

2. NV-02 One Movie, Four Ages

3. NV-03 One Flavor, Growing Up

Required invariants

• latest state never overwrites history;

• different states at different times are not automatically contradiction;

• context and encounter count are explicit;

• object version/subtype is explicit;

• scalar ratings do not erase dimensions;

• stale old preferences remain historically true;

• revisitability is distinct from recommendation;

• one encounter never becomes timeless identity by default.

Required data distinction

Separate:

RelationObservation
RelationTrajectory
DerivedCurrentRelation
DerivedRevisitability

These names are conceptual only unless separately ratified.

Required change classification tests

Show examples of:

• REVISION;

• DRIFT;

• CONTEXT;

• CYCLE.

Do not require every trajectory to have exactly one type.

NV-01 acceptance

Must preserve: dislike → familiarity → love → saturation → avoidance → dormancy → nostalgia.

At encounter #8: current negative response must not delete prior love.

After dormancy: return must not be treated as inconsistency.

NV-02 acceptance

Must preserve dimensions separately:

• surprise;

• craft admiration;

• protagonist sympathy;

• emotional resonance.

Latest overall rating must not replace earlier interpretations.

NV-03 acceptance

Must distinguish:

• yellow mustard;

• Dijon;

• whole-grain.

Must not infer person-only drift when object subtype changed.

Childhood dislike must be historical truth but low-confidence current predictor after contradictory later evidence.

Privacy

All fixtures synthetic.

No public release of user preference histories by default.

Negative tests

• first skip → permanent dislike: FAIL

• latest rating overwrites prior: FAIL

• different context → contradiction: FAIL

• old negative observation treated as false: FAIL

• old negative observation treated as forever-current: FAIL

• object changed but system labels person drift: FAIL

• revisitability treated as forced recommendation: FAIL

End state

The system can answer:

What do I seem to prefer now?
How stable is that?
How did I get here?
What contexts change the response?
What did I once love/dislike?
Is an old preference worth revisiting?

without manufacturing one timeless profile.