Status: CANDIDATE · AUTHORITY NONE · PRE-IMPLEMENTATION
Date: 2026-08-09
Purpose: make ordinary human change first-class rather than treating it as correction noise

1. CORE AXIOM

Variation is not noise around the person. Variation is part of the person.

Amnesiac already requires historical state preservation.

This doctrine adds three stronger rules:

CURRENT
≠
TRUE SELF

LATEST
≠
AUTOMATIC AUTHORITY

CHANGE
≠
ERROR

A person may:

• learn;

• drift;

• cycle;

• react differently by context;

• return to an old preference;

• outgrow something;

• rediscover it;

• separate craft admiration from enjoyment;

• change values without discovering an error.

Those are normal states.

2. REVISION IS ONLY ONE KIND OF CHANGE

Do not collapse all temporal change into supersedes.

Candidate change families:

REVISION
  changed because evidence / reasoning changed

DRIFT
  response changed gradually without one discrete correction

CONTEXT
  different response in different situations

CYCLE
  response returns/reverses over repeated exposure or long gaps

These are candidate vocabulary, not a frozen enum.

One relation may exhibit several families.

3. THE RELATION IS THE PERSISTENT OBJECT

Static personalization usually models:

Person
  likes
Object

Amnesiac should instead preserve:

Person
        │
        ▼
Relationship-to-Object
        │
        ├── Encounter T1
        ├── Encounter T2
        ├── Encounter T3
        └── Encounter Tn

The durable object is the relationship trajectory.

Individual reactions are observations on that trajectory.

4. MINIMUM ENCOUNTER MODEL

Conceptually:

RelationObservation
  person_state_ref
  object_state_ref
  observed_at
  encounter_index
  context
  response_dimensions
  confidence
  private_basis[]

No one field must reduce the response to like = true/false.

5. BOTH ENDPOINTS MAY CHANGE

The strongest form is:

PERSON STATE AT T
        ×
OBJECT STATE AT T
        ×
CONTEXT AT T
        ↓
RELATION OBSERVATION AT T

Examples:

• person changes with age;

• movie is seen in a different cut;

• restaurant changes chef;

• song is heard in a different recording;

• friendship changes circumstances;

• product changes generation;

• preference changes through familiarity.

The system must not assume only the observer changes.

6. EXPOSURE IS A FIRST-CLASS DIMENSION

Time alone is insufficient.

For many aesthetic/preferences domains:

encounter #1
≠
encounter #10

Possible trajectory:

novelty
→ familiarity
→ appreciation
→ saturation
→ avoidance
→ dormancy
→ nostalgia

Therefore encounter_index and recent-exposure history may be materially relevant.

7. MULTI-DIMENSIONAL RESPONSE

Do not flatten:

overall_rating = 6/10

when the human state is:

surprise              high → low
craft admiration      stable/high
emotional resonance   medium → high
novelty               high → zero
social enjoyment      context-dependent
nostalgia             absent → high

A scalar may be stored if the person supplied one.

It must not erase the vector.

8. PREFERENCE STALENESS

A historical preference observation can become stale without becoming false.

Example:

"I hated mustard at age six."

Historical truth: still valid.

Current recommendation authority: weak unless reinforced.

Candidate concept:

preference_staleness

must be derived from context, domain, age/elapsed time, and later evidence — not a universal decay formula.

9. REVISITABILITY

A prior negative response should not always suppress future exposure.

Candidate derived relation:

revisitability

Meaning:

Is this old response worth testing again?

Possible basis:

• very old observation;

• only one encounter;

• known slow-burn trajectory;

• large context change;

• object changed;

• person moved into adjacent tastes;

• long dormancy after saturation.

This is not “recommend anyway.” It is a reasoned reopening condition.

10. SETTLED PREFERENCE IS DERIVED, NOT ASSUMED

One skip, one bad meal, one unfinished book, or one awkward experience should rarely become identity.

Possible progression:

single_observation
→ repeated_pattern
→ context-stable pattern
→ settled preference

No universal encounter threshold is ratified here.

11. PERSONAL FORMAL CONCEPTS

Across many trajectories, the private graph can derive concepts such as:

slow burns

novelty spikes

things I love socially but not alone

things whose craft I admire but do not enjoy

things I outgrew

things I grew into

things I return to after long dormancy

things I only enjoy in a specific context

These are not profile labels to permanently stamp onto the person.

They are concepts supported by encounter histories.

12. LATEST STATE DOES NOT DELETE PRIOR STATE

If:

T1  love
T2  saturation
T3  dislike
T4  nostalgia

then current state may be:

nostalgic affection

but history remains:

love → saturation → dislike → nostalgia

The system should be able to answer:

• Did I ever genuinely love this?

• When did I tire of it?

• Did I rediscover it?

• Was the return caused by context, dormancy, or reinterpretation?

13. RELATIONSHIP TO BELIEF REVISION

Epistemic correction:

"I thought X was true."
→ evidence
→ "I now think X was false."

Preference drift:

"I loved X."
→ time/context/exposure
→ "I now enjoy X less."

The first may supersede an erroneous factual belief.

The second preserves both states as valid at their times.

Do not use one lifecycle model for both.

14. RELATIONSHIP TO COUNTERPEDIA

Counterpedia answers:

What is the public object?
What version is it?
What claims govern it?

Amnesiac answers:

What has my relationship to this object been?

Together:

PUBLIC OBJECT HISTORY
        ×
PRIVATE RELATION HISTORY
        ↓
inspectable encounter trajectory

15. RELATIONSHIP TO COLLIDE

A current preference and historical preference may differ without being contradiction.

COLLIDE must inspect:

• valid_at;

• context;

• exposure count;

• object edition;

• response dimension.

Example:

T1: "I dislike this song on first listen."
T10: "I love this song."

relationship:
  temporal variation
not:
  contradiction requiring correction

16. REGRESSION INVARIANTS

NV-R1

One reaction does not become timeless identity.

NV-R2

Latest reaction does not erase historical reactions.

NV-R3

Different contexts may legitimately yield different responses.

NV-R4

Repeated exposure may alter preference without new factual evidence.

NV-R5

Object version changes must not be mistaken for person drift.

NV-R6

A stale preference remains historically true but may be weak for current prediction.

NV-R7

A scalar rating never erases dimensional response.

NV-R8

Returning to an old preference is not inconsistency.

NV-R9

Revisitability is a reopening condition, not a forced recommendation.

NV-R10

Preference concepts remain evidence-derived and revisable.

17. IMPLEMENTATION HOLD

Do not yet ratify:

• a universal RelationObservation public schema;

• a global taste taxonomy;

• numerical preference-decay rules;

• one universal settled_preference threshold;

• automated psychological explanations;

• romantic/relationship scoring;

• public exposure of private preference histories.

First prove the doctrine with ordinary, low-risk content.

18. FIRST THREE PROVING SPECIMENS

NV-01
ONE SONG, TEN HEARINGS
→ exposure / saturation / nostalgia

NV-02
ONE MOVIE, FOUR AGES
→ reinterpretation / dimensional response / life-stage change

NV-03
ONE FLAVOR, GROWING UP
→ developmental acquisition / preference staleness / revisitability

These three deliberately avoid making intimate relationships the flagship.

Once the substrate is credible, the same doctrine naturally applies to:

• friendships;

• romantic attachment;

• places;

• careers;

• habits;

• identities;

• values.

19. DOCTRINE SENTENCE

Amnesiac should learn not only what a person likes, believes, and values, but how those relations change — because variation itself is part of the historical record.