ONE DECISION, THREE SELVES

STATUS: RESEARCH / DESIGN ONLY.
Use only synthetic private data until a separate privacy/authority lane is ratified.

Shared decision

Synthetic:

Stay in current New York role vs accept a San Francisco role and relocate.

Do not represent any fixture value as a real current city/employer fact.

Core question

Was taking the job a bad decision — or do I know and value different things now?

Three temporal selves

S0:

• 48 hours before decision

• facts T0

• beliefs/forecasts T0

• values T0

• no outcome knowledge

S1:

• two weeks after accepting

• decision known

• early narrative / rationale drift

• no major outcomes yet

S2:

• 18 months later

• later public policy edition

• later private outcomes

• changed values

• current regret/preferences

Required immutable objects

• original fact snapshot;

• original value state;

• original forecasts;

• decision rationale;

• decision commitment;

• later policy edition;

• later outcomes;

• preference revisions;

• post-decision narrative;

• current narrative.

Do not overwrite predecessors.

Required replay grid

Support four contexts:

1. facts_then × values_then

2. facts_then × values_now

3. facts_now × values_then

4. facts_now × values_now

Label:

• historical replay;

• counterfactual value-drift projection;

• counterfactual hindsight projection;

• current context.

Historical replay MUST NOT be confused with counterfactual recomputation.

Required forum

Question: Was taking the job a bad decision?

Agents:

• Maya-S0

• Maya-S1

• Maya-S2

Factor disagreement into:

• fact/evidence change;

• public source-edition change;

• forecast error;

• value drift;

• outcome;

• narrative drift.

Do not emit a naked contradiction verdict before comparing basis/time/value dimensions.

Required COLLIDE examples

• S0 “MOVE best” vs S2 “STAY best” should inspect basis_edition, valid_at, and preference context.

• current memory “I knew office policy would get worse” vs recorded T0 forecast of low probability should preserve discrepancy without rewriting either.

Required reverse traces

Public/private combined owner view:

OfficePolicy v1 → T0 fact set → forecast → Decision D1

and:

OfficePolicy v2 → later outcome → belief/value revision → current preference

Required negatives

• bad outcome does not prove bad ex-ante decision;

• good outcome does not prove sound ex-ante reasoning;

• current values do not overwrite old values;

• later facts do not enter T0 replay;

• post-decision narrative does not replace original rationale;

• “would I do it again?” is underspecified without fact/value basis;

• counterfactual result is not historical memory;

• private values never become public by default.

Acceptance

PC-04 succeeds if the demo can answer:

What did I know then?
What did I value then?
Why did I choose?
What changed in the world?
What changed in me?
Which forecast was wrong?
What do I prefer now?
Would I choose differently because of new facts,
different values,
or both?

End:

Replay the decision as I was then.