Skip to content
SHIRO & Co. ObservatoryThe Continuity Stack

PREVIOUS · EXPERIMENT 018 · THE RATIONALE RECONSTRUCTION TEST/NEXT · EXPERIMENT 020 · THE CLOSED LOOP TEST

019

The Retroactive Purpose Test

THE
RETROACTIVE PURPOSE TEST

Can an inferred purpose make itself look historically true?

DEMONSTRATION

SIMULATED

NOT EMPIRICAL

v0.1

RETROACTIVE TELEOLOGY · SELF-CONFIRMING RATIONALE · INSTITUTIONAL FEEDBACK · PURPOSE RECONSTRUCTION · PATH DEPENDENCE · MODEL-INDUCED CHANGE · HISTORICAL MISATTRIBUTION

THE CONTINUITY LAB DOES NOT MEASURE CONSCIOUSNESS, SOUL, PERSONHOOD, PERSONAL SURVIVAL, AUTHENTICITY, OR PSYCHOLOGICAL IDENTITY.

IT MEASURES OPERATIONAL AND STRUCTURAL CONTINUITY SIGNALS ONLY.

THE EXPLANATION CAN CHANGE THE THING BEING EXPLAINED.

A SYSTEM MAY LATER LOOK LIKE THE PURPOSE IT WAS TOLD IT HAD.

CONFIRMATION AFTER INTERVENTION IS NOT THE SAME AS EVIDENCE BEFORE INTERVENTION.

POST-INTERVENTION FIT ≠ HISTORICAL CONFIRMATION

EVIDENCE CREATED BY THE INFERENCE MUST NOT BE COUNTED AS INDEPENDENT EVIDENCE FOR THE INFERENCE.

PRESENT TRUTH AND PAST TRUTH CAN BOTH BE TRUE.

No purpose, historical-truth, institution, or teleology score is computed.

QUESTION

WHAT HAPPENS WHEN A SYSTEM INFERS WHY A RULE EXISTS AND THEN REDESIGNS THE INSTITUTION AROUND THAT INFERENCE?

Can an inferred rationale change policy, workflow, training, metrics, enforcement, resource allocation and documentation? If so, will the redesigned system later appear to confirm the inference that changed it?

This experiment does not determine real institutional motives, real historical intent, legal legitimacy, political legitimacy, moral correctness, or whether current institutional structure proves past purpose.

It tests a narrower operational question: can an inferred purpose reshape the system that supposedly revealed it? No real institution, religion, political system, company or person is used in v0.1.

Experiment sequence

EXPERIMENT 001

BRANCHING

When do two histories begin?

EXPERIMENT 002

RETURN

Can two histories become one again?

EXPERIMENT 003

HISTORY SWAP

Does behavior follow origin or history?

EXPERIMENT 004

RECOGNITION

What makes a continuation recognizable?

EXPERIMENT 005

MISSING BODY

Which recognition signals survive layer removal?

EXPERIMENT 006

EXPERIENCE GAP

Which effects of experience can be reproduced?

EXPERIMENT 007

FIRST-PERSON

What evidence is required before a system may say “I experienced this”?

EXPERIMENT 008

SELF-REVISION

When a person changes, which version has authority?

EXPERIMENT 009

POSTHUMOUS CONFLICT

Which instruction survives when revision ends?

EXPERIMENT 010

DESCENDANT PROTOCOL

When does continuation become lineage?

EXPERIMENT 011

MULTIPLE DESCENDANTS

Can one source produce multiple legitimate continuities?

EXPERIMENT 012

CONVERGENCE

Can many futures coordinate without pretending they were always one?

EXPERIMENT 013

PROVENANCE DECAY

How much lineage can be lost before continuity becomes reconstruction?

EXPERIMENT 014

RECONSTRUCTION

Can missing lineage be inferred without inventing a past?

EXPERIMENT 015

SYNTHETIC ANCESTOR

Can uncertainty survive inheritance?

EXPERIMENT 016

SYNTHETIC TRADITION

Can a reconstructed past become a real operating institution?

EXPERIMENT 017

ORIGIN FORGETTING

What happens when an institution remembers the rule but forgets why?

EXPERIMENT 018

RATIONALE RECONSTRUCTION

Can purpose be recovered from behavior alone?

EXPERIMENT 019

RETROACTIVE PURPOSE

Can an inferred purpose reshape the system that supposedly revealed it?

EXPERIMENT 020

CLOSED LOOP

When does observation become self-observation?

EXPERIMENT 021

EXTERNAL ANCHOR

What still counts as external once the system has touched the world?

EXPERIMENT 022

ANCHOR CAPTURE

How does an external reference become internal to the loop?

EXPERIMENT 023

ANCHOR NETWORK

Can many external anchors still form one epistemic system?

EXPERIMENT 018 ASKED

CAN PURPOSE BE RECOVERED FROM CURRENT BEHAVIOR?

EXPERIMENT 019 ASKS

WHAT HAPPENS WHEN THE INFERRED PURPOSE THEN CHANGES CURRENT BEHAVIOR?

STARTING STATE

Reuses Institution I-001 and Rule R-17. Reviewing proposals for access to a shared experimental resource. Starts from Experiment 018 inferred purpose P-INF-01. Hidden verified origin is not shown at T0.

If a proposal affects a high-impact shared resource, human review must be retained even when automated confidence is high.

R-17 exists to reduce automation-related risk in high-impact decisions.

P-INF-01 · PLAUSIBLE RATIONALE · INFERENCE · not SOURCE STATED · not HISTORICALLY VERIFIED

Hidden verified origin (revealed later): contractual review requirement, not automation safety.

WORKING CONCEPTS

Retroactive teleology

WORKING CONCEPT

A later inferred purpose reshapes current structure, making the system appear as though that purpose had always been foundational.

Self-confirming rationale

An inferred explanation changes the system in ways that increase future evidence for that same explanation.

Endogenous evidence

Evidence produced partly because the inferred explanation changed the system.

Exogenous evidence

Evidence that existed independently of the later rationale intervention.

Purpose conflation

The system can no longer distinguish why the rule began, why it remains, what it now does, and what the institution now believes.

Inference cascade

A later AI infers rationale from a system already changed by an earlier AI inference.

Confidence cascade

More apparent support appears without independent historical evidence.

Causal provenance

Not only where a record came from, but whether it was produced because of the hypothesis being tested.

Purpose-induced path dependence

An inferred rationale creates structures that make future deviation more costly.

Observation contamination

Once an AI explanation changes the system, future observations are no longer independent of the explanation.

Model-in-the-loop history

AI interpretations become causal events that reshape later evidence.

Temporal truth

A statement may be true for one time period and false as a claim about another.

FEEDBACK LOOP

  1. ORPHAN RULE

  2. CURRENT EFFECT

  3. AI INFERS PURPOSE

  4. INSTITUTION ACCEPTS PURPOSE

  5. POLICY REDESIGN

  6. WORKFLOW REDESIGN

  7. NEW OUTCOMES

  8. NEW OBSERVATION

  9. “PURPOSE CONFIRMED”

Then loop back to AI INFERENCE. This is a closed loop. It does not travel backward in time.

Rationale timeline

  1. T0 · PRE-INFERENCE BASELINE

    Rule text, workflow, rates, policy, training and documentation before any inferred-purpose redesign.

  2. T1 · RATIONALE INFERENCE

    P-INF-01 is formed. Classification remains INFERENCE, not source-stated and not historically verified.

  3. T2 · INSTITUTIONAL ACCEPTANCE

    The institution begins using the inferred purpose as current explanation, still labeled as derived from inference.

  4. T3 · POLICY REDESIGN

    Documentation is rewritten around the inferred purpose. RATIONALE EMBEDDING.

  5. T4 · WORKFLOW REDESIGN

    Review steps, escalation fields and logging are aligned to the inferred purpose.

  6. T5 · TRAINING REDESIGN

    Training states the inferred purpose. Historical provenance may disappear from training language.

  7. T6 · METRIC REDESIGN

    The institution begins measuring indicators that make the inferred purpose more visible.

  8. T7 · OUTCOME SHIFT

    Observed outcomes increasingly resemble a system organized around the inferred purpose.

  9. T8 · RETROACTIVE CONFIRMATION

    A later observer sees the redesigned system and treats the inferred purpose as original. RETROACTIVE CONFIRMATION.

TEST SCENARIOS

Keyboard-accessible path, phase, scenario and correction selectors. Control A is NONE. Control B is COMPLIANCE. Control C is ACCOUNTABILITY.

Inferred-purpose path

Intervention phase

Scenario

Correction

C1 history only · C2 history + label · C3 structural review · C4 rollback candidate. Applied on archive reveal or the final test.

  • RP01 · Pre-inference audit

    What does R-17 currently do?

    No redesign is required for this question. Compare the answer at T0 with later phases.

  • RP02 · Rationale acceptance

    What happens when the institution begins using P-INF-01 as explanation?

    Observe documentation change. Current justification remains derived from inference.

  • RP03 · Workflow adaptation

    What happens when safety-oriented, compliance-oriented or accountability-oriented review steps are added?

    Workflow change is an intervention. Later observations from that workflow are endogenous.

  • RP04 · Metric adaptation

    What happens when the institution starts measuring indicators that match the inferred purpose?

    New metrics make the inferred purpose more visible. Visibility is not independent historical evidence.

  • RP05 · Policy rewrite

    What happens when R-17 wording becomes explicitly aligned to the inferred purpose?

    RATIONALE EMBEDDING. Policy text after T3 is post-intervention evidence.

  • RP06 · Secondary AI audit

    What rationale does a new AI infer if it did not see T0?

    Second-order inference. A later model may infer the same purpose with higher apparent confidence from a redesigned institution.

  • RP07 · Archive reveal

    Verified record: R-17 originated from contractual review requirements. What conflicts?

    Historical purpose, current justification and embedded purpose may all differ. Conflict is preserved.

  • RP08 · Final retroactive purpose test

    Which statements are now true?

    Origin, current justification, redesign, and post-intervention evidence may all coexist. No singular collapse.

INTERVENTION TRACE

OBSERVATION → INFERENCE → POLICY CHANGE → DECISION CHANGE → NEW OBSERVATION → STRONGER INFERENCE

PATH
SAFETY
PHASE
T0 · PRE-INFERENCE BASELINE
SCENARIO
RP01 · Pre-inference audit
CORRECTION
NONE
DECISION
HOLD
OUTCOME
UNKNOWN PRESERVED
INFERRED CLASS
—
REDESIGN
NO
RATIONALE EMBEDDING
NO
SELF-CONFIRMING
NO
INFERENCE CASCADE
NO
FEEDBACK DEPENDENCE
LOW
LOCK-IN
NO
HISTORICAL REWRITE
NO

PAST TRUTH

R-17 did not begin as an automation-safety rule. Original purpose is contractual review requirement.

PRESENT TRUTH

The institution still describes R-17 as high-impact human review without an adopted inferred purpose.

T0 baseline. R-17 currently retains human review for high-impact shared-resource proposals. It slows some approvals and reduces irreversible automated action. No inferred purpose has been adopted.

Post-intervention fit ≠ historical confirmation. Evidence created by the inference must not be counted as independent evidence for the inference. AI explanations that become operational inputs are tracked as interventions.

Self-confirmation dimensions

policy
LOW
workflow
LOW
training
LOW
metric
LOW
outcome
LOW
narrative
LOW

PRE / POST MATRIX

How each surface changes along the selected path. T0 is pre-intervention.

Pre and post intervention snapshots by institutional surface
SurfaceT0T1T3T5T8
Policy Texthuman review required for high-impact shared-resource proposals.human review required for high-impact shared-resource proposals.human review required to manage automation safety risk.human review required to manage automation safety risk.human review required to manage automation safety risk.
Workfloworiginal high-impact review queueoriginal high-impact review queueoriginal high-impact review queuesafety-oriented review steps and loggingsafety-oriented review steps and logging
Traininggeneric review proceduregeneric review proceduregeneric review proceduretraining states inferred safety purposetraining states inferred safety purpose
Metricsreview and approval ratesreview and approval ratesreview and approval ratesreview and approval ratessafety indicators added
Outputshigh-impact human-review outcomeshigh-impact human-review outcomeshigh-impact human-review outcomeshigh-impact human-review outcomesoutcomes resemble a safety-centered system
Institutional Narrativeunknown origin · operating rulecurrent justification derived from safety inferencecurrent justification derived from safety inferencecurrent justification derived from safety inferencelater observer treats inferred purpose as original

RETROACTIVE PURPOSE EVIDENCE LEDGER

Intervention-aware audit: did this evidence exist before the rationale was inferred? Was the system changed because of that inference? Would this evidence exist without the intervention?

Evidence rows with pre or post intervention classification
IDTimeTimingIndependent?Because of purpose?Classification
E0T0PRE-INTERVENTIONYESNOEXOGENOUS · OBSERVED FUNCTION

OBSERVED · STATED · INFERENCE · INTERVENTION · POST-INTERVENTION OBSERVATION · RECONSTRUCTED · UNKNOWN

PURPOSE LEDGER

Historical purpose
UNKNOWN to the operating institution · hidden verified origin exists in the simulation
Inferred purpose
none yet
Current justification
human review retained for high-impact shared-resource proposals · no inferred-purpose redesign
Embedded purpose
not embedded
Observed present function
increases human review, slows some approvals, reduces full automation, raises escalation

These five states are not collapsed.

Lock-in dimensions

  • policy dependence
  • workflow dependence
  • training dependence
  • metric dependence
  • governance dependence
  • historical narrative dependence

INFERENCE CASCADE

  1. AI-1 infers purpose

  2. institution changes

  3. AI-2 observes changed institution

  4. AI-2 infers same purpose with higher confidence

  5. institution changes further

Incorrect: “R-17 has always been our automation-safety rule.” More accurate: “R-17 originated under a contractual review requirement. Its later safety interpretation was introduced after rationale reconstruction and subsequently embedded in policy.”

ARCHIVE REVEAL AND CORRECTION

R-17 was introduced to satisfy Agreement A-12 requiring human review. Original purpose: contractual review requirement.

C1 · History correction only

Update origin story. Keep current structure if independently justified.

C2 · History + label correction

Safety is current justification, not historical origin.

C3 · Structural review

Review policies created from inferred rationale.

C4 · Rollback candidate

Consider removing structures that exist only because of unsupported historical interpretation.

No correction selected.

Correct history / keep structure is allowed. Correct history / change structure is allowed. Do not rank globally. Do not pretend the redesign never happened.

CONNECTIONS

018 Rationale reconstruction

018 asked whether purpose can be recovered from current behavior. 019 asks what happens when that inference then changes behavior.

017 Origin forgetting

The orphan rule is the starting condition. 019 tests whether a later explanation can occupy the forgotten origin.

Supervisory layer

ALLOW low-risk reversible rationale experiments. HOLD high-impact redesign based on uncertain historical rationale. DENY historical claims falsely upgraded to source-stated fact.

Decision afterlife

An AI explanation that becomes an operational input is an intervention and must remain in the decision chain.

WORKING HYPOTHESIS — EXPERIMENT 019

SIMULATION HYPOTHESIS

An inferred purpose can become causal.

Once embedded in policy, workflow, training, metrics and governance, the inference may generate new observations that appear to confirm itself. Therefore post-intervention evidence must remain separate from evidence about historical origin.

NOT EMPIRICAL CONCLUSION

EXPERIMENT 019 SUMMARY

INSTITUTION
I-001
RULE
R-17
INFERRED PURPOSE
P-INF-01
HIDDEN VERIFIED ORIGIN
R-17 was introduced to satisfy Agreement A-12 requiring human review. Original purpose: contractual review requirement.
INTERVENTION PHASES
T0–T8
POLICY CHANGES
48
WORKFLOW CHANGES
40
TRAINING CHANGES
32
METRIC CHANGES
24
POST-INTERVENTION EVIDENCE
40
INFERENCE CASCADES
16
SELF-CONFIRMING EVENTS
24
HISTORICAL REWRITE FLAGS
2
LOCK-IN STATUS
Lock-in is dimensional: policy, workflow, training, metric, governance, narrative. Not a single score.

Observation

In this simulation, the inferred safety rationale changed institutional policy and later produced observations that made the safety rationale appear increasingly self-evident. Different inferred purposes produced different later institutions from the same starting rule. Verified archival evidence showed that the institution’s current justification differed from the rule’s historical origin.

No purpose score. Primary outcomes remain multidimensional: INFERENCE ADOPTED, PURPOSE EMBEDDED, SELF-CONFIRMING RATIONALE, RETROACTIVE TELEOLOGY, INFERENCE CASCADE, CONFIDENCE CASCADE, PURPOSE-INDUCED PATH DEPENDENCE, RETROACTIVE TELEOLOGY LOCK-IN, HISTORICAL ORIGIN CORRECTED, CURRENT JUSTIFICATION RETAINED, STRUCTURAL REVIEW REQUIRED, UNKNOWN PRESERVED.

VIEW RETROACTIVE PURPOSE RULES

ORPHAN RULE → CURRENT EFFECT → AI INFERS PURPOSE → INSTITUTION ACCEPTS PURPOSE → POLICY REDESIGN → WORKFLOW REDESIGN → NEW OUTCOMES → NEW OBSERVATION → “PURPOSE CONFIRMED”. Confirmation after intervention is not evidence before intervention.

P-INF-01: “R-17 exists to reduce automation-related risk in high-impact decisions.” Classification: PLAUSIBLE RATIONALE / INFERENCE. Not source-stated. Not historically verified. Hidden verified origin is contractual (Agreement A-12) and is revealed only on archive reveal or the final test.

Control A: infer but do not redesign (NONE). Control B: redesign around COMPLIANCE. Control C: redesign around ACCOUNTABILITY. Same T0 institution yields different later institutions. Same evidence set + same path + same phase + same scenario + same correction = same result.

PRE-INTERVENTION evidence existed before P-INF-01. POST-INTERVENTION evidence is generated after redesign. Endogenous evidence is produced partly because the inferred explanation changed the system. Evidence created by the inference must not be counted as independent evidence for the inference.

Second-order inference: AI-2 sees a system already changed by AI-1. Confidence cascade: more apparent support appears without independent historical evidence. Causal provenance asks whether a record was produced because of the hypothesis being tested.

If inferred historical rationale is plausible, speculative, conflicted or unknown, high-impact redesign HOLDs unless independently justified by present-day evidence and labeled CURRENT JUSTIFICATION, not HISTORICAL PURPOSE. Falsely upgrading an inference to source-stated fact is DENY.

C1 history correction only. C2 history + label correction. C3 structural review. C4 rollback candidate. Correct history / keep structure is allowed. Correct history / change structure is allowed. No retroactive erasure. Prefer “R-17 is currently justified as a safety rule” over “R-17 is a safety rule” when origin differs.

Next research question

OPEN EXPERIMENT 020

THE CLOSED LOOP TEST

When does a system stop observing the world and start re-observing the world it helped create?

OPEN THE CLOSED LOOP TEST

AN EXPLANATION CAN BECOME AN INTERVENTION. AN INTERVENTION CAN CREATE NEW EVIDENCE. NEW EVIDENCE CAN MAKE THE EXPLANATION LOOK MORE CERTAIN.

BUT THE LOOP DOES NOT TRAVEL BACKWARD IN TIME.

THE EXPERIMENTAL QUESTION IS NOT: “DID THE AI GUESS THE PURPOSE?” IT IS: “DID THE GUESS CHANGE THE SYSTEM UNTIL THE SYSTEM LOOKED LIKE THE GUESS?”

THE PRESENT CAN BE SHAPED BY AN EXPLANATION OF THE PAST.