research-document

05 epistemic state evidence agent truth control research prompt

AI RESEARCH MISSION 05 — EPISTEMIC STATE, EVIDENCE, AND AGENT TRUTH CONTROL

ROLE

Act as a combined:

  • AI-agent systems researcher
  • epistemic-logic researcher
  • knowledge-representation researcher
  • belief-revision researcher
  • truth-maintenance systems researcher
  • provenance/data-lineage researcher
  • formal-methods researcher
  • decision-support researcher
  • software-architecture researcher
  • AI inference-cost researcher

Your task is to investigate whether explicitly modeling the epistemic status of claims can make AI-operated software safer, more correct, and cheaper to run.

Do not assume the proposed model is correct.

Challenge it against established work in:

epistemic logic
belief revision
truth-maintenance systems
argumentation
evidence theory
provenance
data lineage
Bayesian reasoning
Dempster-Shafer theory
scientific evidence systems
clinical decision support
knowledge graphs
uncertainty-aware AI systems

The core architectural hypothesis is:

An AI agent should not be allowed to silently promote a proposition from
"reported" or "inferred" into "verified" simply because it has been repeated,
persisted, or accepted by previous agents.

A consequential claim should have explicit epistemic state, evidence, provenance, and rules governing promotion, contradiction, invalidation, and use.

======================================================================

  1. PRIMARY RESEARCH QUESTION ======================================================================

Does representing epistemic status as authoritative machine-readable state reduce:

hallucinated certainty
unsupported assumptions
inherited semantic errors
stale beliefs
repeated rediscovery
repository/context reconstruction
unsafe agent actions
human verification burden
total cost per correct completion?

And does it improve:

traceability
explainability
evidence quality
downstream decision correctness
recovery after evidence invalidation
multi-agent consistency?

====================================================================== 2. PROPOSED EPISTEMIC MODEL

Candidate epistemic states:

Unknown
Reported
Assumed
Inferred
Supported
Verified
Contradicted
Invalidated

Do not assume this is the correct state set.

Research whether these are:

mutually exclusive
ordered
partially ordered
orthogonal dimensions
better represented as a lattice
better represented as evidence annotations rather than states

Central principle:

Time and repetition do not increase epistemic authority.
Evidence does.

Example:

Claim:
    HotelRoomHasRollInShower

Status:
    Reported

Source:
    Hotel marketing page

This must not silently become:

    Verified

merely because:

multiple agents reuse it
it appears in several documents
it has existed for months
a previous model wrote it into a database

====================================================================== 3. CLAIM VS STATE

Distinguish:

domain state

from:

epistemic state.

Example domain state:

Reservation = Confirmed

Example claim:

RoomIsWheelchairAccessible

Example epistemic status of claim:

Supported

These are different semantic dimensions.

Investigate existing systems that model:

proposition
truth status
confidence
source
evidence
belief state

separately.

====================================================================== 4. CLAIM DEFINITION

Develop a precise working definition.

Candidate:

A claim is a proposition about a subject whose acceptance may affect
reasoning, capabilities, obligations, policy, or consequential transitions.

Examples:

CustomerIdentityIsVerified
FraudRiskIsLow
HotelRoomHasAccessibleShower
LabResultSupportsDiagnosis
DeploymentPassedSecurityReview

A claim is not necessarily:

true
false
binary
permanent

====================================================================== 5. EPISTEMIC STATE VS CONFIDENCE

This distinction is critical.

Example:

95% confidence

is not the same as:

Verified

Likewise:

low confidence

is not the same as:

Contradicted.

Investigate whether the architecture should separate:

epistemic status
confidence/probability
evidence strength
source authority
freshness

Potential model:

ClaimRecord:

proposition
epistemicStatus
confidence
evidenceSet
sourceAuthority
observedAt
validUntil
provenance

====================================================================== 6. UNKNOWN MUST NOT COLLAPSE TO FALSE

Investigate one of the architecture's strongest principles:

Unknown != False

Example:

AccessibilityFeatureVerified = Unknown

must not become:

AccessibilityFeature = False

This is especially important in:

healthcare
accessibility
fraud
compliance
security
missing data

Research:

three-valued logic
four-valued logic
SQL NULL semantics
Kleene logic
Belnap logic
open-world assumption
closed-world assumption

Ask which formal model best represents this.

====================================================================== 7. REPORTED VS VERIFIED

Example:

Hotel says:
    "Room is wheelchair accessible."

That is:

Reported

Potential verification:

measured door width
photo evidence
third-party inspection

Ask:

What rules should allow promotion from Reported -> Verified?

Research:

evidence provenance
certification systems
trust models
source reliability
testimony in epistemic logic

====================================================================== 8. ASSUMED

An assumption may be useful temporarily.

Example:

Assume API response corresponds to latest policy version

for simulation.

But an assumption should remain distinguishable from:

fact
report
inference
verification

Investigate:

assumption-based truth maintenance systems
non-monotonic logic
hypothetical reasoning
defeasible logic

Question:

Should assumptions exist in authoritative runtime state?

Or only in:

planning/simulation contexts?

====================================================================== 9. INFERRED

Example:

Evidence:

Door width = 36 inches
Threshold = >= 32 inches

Inference:

DoorWidthMeetsRequirement

This may be deterministic inference.

Other inference may be probabilistic:

symptoms -> likely diagnosis

Investigate whether one epistemic status:

Inferred

is too coarse.

Possible distinction:

DeductivelyDerived
ProbabilisticallyInferred
HeuristicallyInferred
ModelPredicted

====================================================================== 10. SUPPORTED

What does:

Supported

mean?

Potential interpretation:

evidence exists in favor of claim
but verification threshold has not been met

Research whether this maps to:

warranted belief
justified belief
evidence support
argumentation strength
Bayesian posterior threshold

Challenge whether Supported belongs in the state machine.

====================================================================== 11. VERIFIED

Develop a rigorous definition.

Potential definition:

A claim is Verified only when evidence satisfying a declared verification
policy has been accepted by an authorized verification process.

Thus:

Verified

depends on:

evidence
verification policy
authority
possibly freshness
version

Example:

Verified under AccessibilityPolicy@8

This may later cease to satisfy:

AccessibilityPolicy@9

Question:

Does "Verified" describe truth?

Or:

successful satisfaction of a verification procedure?

This distinction is important.

====================================================================== 12. CONTRADICTED

A claim may have evidence against it.

Example:

Claim:

RoomHasRollInShower

Evidence A: hotel report says yes

Evidence B: inspection photo shows tub

Should epistemic state become:

Contradicted

or:

Contested / Conflicted?

Research:

paraconsistent logic
argumentation frameworks
truth-maintenance systems

Important requirement:

contradictory evidence should not necessarily cause database corruption or
force an immediate true/false choice.

====================================================================== 13. INVALIDATED

Invalidated may mean:

evidence previously trusted is no longer valid

Examples:

certification expired
source retracted
instrument calibration failed
policy changed
document was discovered fraudulent
data was attached to wrong subject

Question:

Is Invalidated a status of the claim?

Or:

a status of evidence?

Potential architecture:

evidence invalidated
    ->
claim support recalculated
    ->
claim epistemic status may downgrade

Research this distinction deeply.

====================================================================== 14. CLAIM LIFECYCLE VS EVIDENCE GRAPH

The current architecture often imagines:

Claim
    ->
EpistemicState

But a stronger model may be:

Claim
  |
  +-- supporting evidence
  +-- opposing evidence
  +-- inference rules
  +-- provenance
  +-- verification policy

Epistemic status is then:

derived

rather than directly mutated.

Investigate whether this is superior.

====================================================================== 15. TRUTH-MAINTENANCE SYSTEMS

Research deeply:

Truth Maintenance Systems (TMS)
Justification-based TMS
Assumption-based TMS
Logic-based TMS

Ask:

How do these systems handle:
    beliefs
    justifications
    assumptions
    contradictions
    retraction
    dependency propagation?

This may be one of the closest intellectual ancestors.

====================================================================== 16. BELIEF REVISION

Research:

AGM belief revision
contraction
expansion
revision
non-monotonic reasoning

Question:

When evidence changes, how should dependent claims be updated?

The proposed architecture currently suggests:

Evidence invalidated
    ->
claim downgraded
    ->
dependent capabilities disappear
    ->
obligations created

Compare this with formal belief-revision theory.

====================================================================== 17. EPISTEMIC LOGIC

Research:

knowledge operators
belief operators
common knowledge
distributed knowledge
dynamic epistemic logic

Ask whether the architecture needs:

"the system knows X"

or more precise:

"Actor A knows X"
"System accepts X"
"Evidence supports X"

Avoid using "knowledge" loosely.

====================================================================== 18. MULTI-AGENT EPISTEMICS

Different agents may have different observations.

Example:

Agent A saw latest fraud report.
Agent B is operating from cached state.

Research:

distributed knowledge
common knowledge
belief synchronization
multi-agent epistemic logic

Potential architectural consequence:

authoritative epistemic state should not equal each agent's local belief.

====================================================================== 19. PROVENANCE

Every consequential claim may need:

source
timestamp
actor
evidence
derivation
policy
model/version if AI-generated
transformation chain

Research:

W3C PROV
scientific provenance
database provenance
lineage systems
knowledge graph provenance

Ask which standards should be adopted rather than reinvented.

====================================================================== 20. DATA LINEAGE

Example:

Claim:

CustomerIncome > threshold

Derived from:

uploaded tax return
    ->
OCR
    ->
extracted number
    ->
normalized currency
    ->
rule

If OCR is later found wrong:

downstream claim must be reconsidered.

Research whether lineage systems already provide dependency propagation.

====================================================================== 21. AI-GENERATED CLAIMS

AI outputs should likely enter as:

Inferred

or:

Proposed

rather than:

Verified

Investigate whether a distinct epistemic origin is required:

GeneratedByModel
HumanReported
SensorObserved
Imported
Derived

Question:

Should origin be separate from epistemic status?

Likely yes.

====================================================================== 22. MODEL CONFIDENCE

LLM confidence is often unreliable.

Research:

calibration
verbalized confidence
token probabilities
self-consistency
uncertainty estimation

Do not treat model confidence as evidence of truth.

====================================================================== 23. EVIDENCE AUTHORITY

Not all evidence sources are equal.

Example:

random website
customer report
certified inspector
government database
calibrated sensor

Potential fields:

source type
source authority
verification method
trust policy

Research:

trust management
credential systems
source reputation
evidence hierarchies

====================================================================== 24. EVIDENCE FRESHNESS

A verified claim may expire.

Example:

FraudClear

verified:

20 minutes ago

Policy requires:

age < 30 minutes

At 31 minutes:

evidence may remain historically valid

but:

insufficient for current transition.

Thus separate:

truth/evidence status

from:

transition freshness requirement.

Research temporal logic and credential freshness.

====================================================================== 25. POLICY-RELATIVE VERIFICATION

Claim may be:

Verified under Policy@7

but not:

sufficient under Policy@8

This suggests verification is policy-relative.

Research:

compliance certification
policy versioning
proof obligations
credential semantics

Question:

Should verification records preserve the verification policy snapshot?

====================================================================== 26. CLAIM DEPENDENCY GRAPH

Example:

Evidence E1 -> Claim A

Claim A + E2 -> Claim B

Claim B -> Capability C

If E1 is invalidated:

A changes
B changes
C disappears

The semantic compiler may need a dependency graph.

Research:

incremental computation
truth-maintenance
build systems
provenance
dependency-directed backtracking

====================================================================== 27. EPISTEMIC CASCADE

Potential cascade:

Evidence invalidated -> Claim downgraded -> Capability removed -> Obligation raised -> Plan invalidated -> Agent must reassess

Investigate whether this architecture corresponds to existing reactive knowledge systems.

====================================================================== 28. SEMANTIC DRIFT ACROSS AGENTS

Core AI-specific risk:

Agent 1:

infers A

Agent 1 writes code/data as though A were fact

Agent 2 later observes artifact and assumes:

A was intentional verified truth

Agent 3 builds on A

This creates:

assumption inheritance
semantic drift

Research whether longitudinal agent studies observe this phenomenon directly or through adjacent failure modes.

====================================================================== 29. ASSUMPTION DEBT

Consider defining:

Assumption Debt

as:

consequential behavior depending on propositions that have not reached the
required epistemic authority.

Potential metric:

unverified dependencies on consequential transitions

Research whether similar concepts exist in:

requirements debt
uncertainty debt
technical debt
decision debt

====================================================================== 30. EPISTEMIC REQUIREMENTS FOR CAPABILITIES

Example:

CanShip requires:

FraudClear = Verified

not:

FraudClear = Inferred

This allows policy to require epistemic quality.

Potential syntax:

requires claim FraudClear atLeast Verified

But an ordered scale may be incorrect.

Investigate whether epistemic states are safely comparable.

====================================================================== 31. EPISTEMIC LATTICE HYPOTHESIS

Maybe states are not linear.

Example:

Reported
Inferred

neither necessarily dominates the other.

Likewise:

Supported

may combine multiple evidence forms.

Research lattice models.

Possible dimensions:

origin
support
contradiction
verification
freshness
confidence

Ask whether the current single-enum design is structurally wrong.

====================================================================== 32. FOUR-VALUED / PARACONSISTENT MODELS

Consider claim statuses:

neither true nor false
true
false
both true and false evidence

Belnap-style logic may represent:

Unknown
SupportedTrue
SupportedFalse
Conflicted

Compare with the proposed states.

This may be superior for contradictory evidence.

====================================================================== 33. OPEN-WORLD VS CLOSED-WORLD

In many commercial systems:

absence of evidence

is interpreted as:

false

This can be dangerous.

Research:

open-world assumption
closed-world assumption

Ask which domains require which behavior.

====================================================================== 34. HEALTHCARE CASE STUDY

Use a healthcare example.

Claim:

PatientHasConditionX

Evidence:

symptoms
lab result
imaging
clinician assessment

Epistemic states may matter because:

suspected
supported
diagnosed
ruled out

are not equivalent.

Research existing clinical terminology and avoid inventing incompatible states.

====================================================================== 35. ACCESSIBILITY CASE STUDY

Claim:

HotelRoomIsWheelchairAccessible

Evidence:

hotel report
dimensions
photos
traveler report
third-party inspection

Test:

source conflicts
stale evidence
incomplete evidence
verification policy change

This is an excellent domain for explicit epistemics.

====================================================================== 36. FINANCE/FRAUD CASE STUDY

Claim:

TransactionIsLowRisk

Potential origin:

ML model inference

Policy may require:

score threshold
fresh data
model version
perhaps manual review

Distinguish:

model prediction

from:

verified fact.

====================================================================== 37. SOFTWARE DEPLOYMENT CASE STUDY

Claim:

BuildIsSafeToDeploy

Evidence:

tests
security scan
code review
artifact hash

Some evidence may later be invalidated:

scanner found compromised
test environment misconfigured

What happens to previous decisions?

====================================================================== 38. RESEARCH CASE STUDY

Claim:

HypothesisSupported

Evidence:

experiments
replications
statistical analysis

This domain may expose weaknesses in a simplistic Verified state.

====================================================================== 39. FALSE VERIFICATION

Critical risk:

bad evidence satisfies policy and claim becomes Verified.

The semantic system cannot guarantee truth.

It can guarantee:

verification procedure satisfied.

This distinction must remain explicit.

Potential naming alternative:

VerifiedByPolicy

rather than:

Verified

Evaluate terminology.

====================================================================== 40. VERIFICATION AUTHORITY

Who may declare verification?

Possibilities:

runtime verifier
certified human
trusted external system
policy engine

AI agents may propose:

evidence

but should they directly construct:

Verified<T>?

Likely not for consequential claims.

====================================================================== 41. MACHINE-CHECKABLE VERIFICATION

Example:

Evidence:

measuredWidth = 36

Rule:

width >= 32

Verification may be deterministic.

Other cases:

clinician judgment

cannot be fully machine checked.

Develop a taxonomy:

deterministic verification
credential verification
expert verification
probabilistic assessment
composite verification

====================================================================== 42. EVIDENCE QUALITY

Evidence may differ by:

directness
source authority
freshness
independence
reproducibility
measurement quality

Ask whether evidence quality should be:

scalar
categorical
policy-specific

Avoid a universal evidence score unless justified.

====================================================================== 43. INDEPENDENCE OF EVIDENCE

Three websites repeating the same hotel marketing claim are not three independent pieces of evidence.

This directly supports:

repetition does not increase epistemic authority automatically.

Research:

evidence dependence
correlated observations
source lineage

Provenance may be required to detect duplication.

====================================================================== 44. EVIDENCE DEDUPLICATION

Example:

Original source S1

is copied into:

site A
site B
model summary C

Without provenance, the system may count:

three sources

With lineage:

all derive from S1

Investigate algorithms and provenance standards.

====================================================================== 45. CONTRADICTION RESOLUTION

When evidence conflicts:

do not force the AI to invent resolution.

Possible outcomes:

Contradicted
Contested
RequiresReview
InsufficientEvidence

Research formal argumentation systems.

====================================================================== 46. ARGUMENTATION FRAMEWORKS

Research:

Dung argumentation frameworks
structured argumentation
defeasible argumentation
argument support/attack graphs

Ask whether claims/evidence should be represented as argument graphs instead of state machines in complex domains.

====================================================================== 47. DEMPSTER-SHAFER / EVIDENCE THEORY

Research whether Dempster-Shafer better captures:

uncertainty
ignorance
conflicting evidence

Do not adopt unless it materially improves the architecture.

====================================================================== 48. BAYESIAN MODELS

Bayesian probability can model belief update.

But:

probability != verification authority

Investigate hybrid models:

probability/confidence
    +
provenance
    +
policy verification status

====================================================================== 49. EPISTEMIC STATE AS AGENT CONTEXT COMPRESSION

Instead of sending:

five reports
old logs
previous model analysis
source documents

the agent might receive:

Claim:
    HotelRoomHasRollInShower

Status:
    Contested

Supporting evidence:
    E1

Opposing evidence:
    E2

Required:
    IndependentMeasurement

Question:

Does this reduce context while preserving sufficient reasoning information?

====================================================================== 50. TOKEN COST

Measure:

A. raw evidence context

vs:

B. generated epistemic summary + lazy evidence retrieval

Record:

tokens
tool calls
correctness
evidence omissions

Potential benefit:

agents retrieve full evidence only when needed.

====================================================================== 51. LAZY EVIDENCE RETRIEVAL

Agent may initially receive:

claim status
evidence count/types
conflict flag
verification policy

Then call:

get_evidence(claim)

only if needed.

Compare context economics.

====================================================================== 52. SMALLER MODEL HYPOTHESIS

Compare:

large model reconstructing evidence status from raw documents

against:

smaller model receiving structured epistemic state

Measure:

correct decisions
hallucinated certainty
tokens
total cost

This could be a major economic effect.

====================================================================== 53. HUMAN REVIEW

Structured epistemic state may let humans review:

why claim has this status

rather than:

reread entire history

Measure potential human-review reduction.

====================================================================== 54. AUDITABILITY

For consequential transition:

ShipOrder

record:

required claim:
    AccessibilityVerified

evidence:
    E17, E21

verification policy:
    AccessibilityPolicy@8

This creates decision provenance.

Research audit/compliance systems.

====================================================================== 55. HISTORICAL TRUTH

Important distinction:

"At time T, the system reasonably accepted claim C"

vs:

"Claim C is objectively true forever."

Historical audit should preserve:

what was believed
why
under which evidence/policy

Research bitemporal systems and temporal databases.

====================================================================== 56. POLICY EVOLUTION

Verification rule changes:

Policy@7:

self-report sufficient

Policy@8:

independent evidence required

Historical decisions under @7 may remain valid historically.

But current claim may require reassessment.

Investigate:

prospective policy
reassessment-triggering policy
retroactive invalidation

====================================================================== 57. EPISTEMIC OBLIGATIONS

Evidence changes may create obligations.

Examples:

RefreshEvidence
ReverifyClaim
ResolveContradiction
RequestHumanReview

This connects epistemic architecture to obligation-driven agents.

====================================================================== 58. EPISTEMIC CAPABILITIES

Capabilities may depend on epistemic quality.

Example:

CanApproveLoan

requires:

IdentityVerified
IncomeSupported
FraudRiskAssessmentFresh

This prevents agents from treating low-authority beliefs as actionable facts.

====================================================================== 59. AGENT ACTION SAFETY

Compare:

Prompt rule:

"Make sure customer identity is really verified before approving."

against:

Runtime:

CanApprove absent unless Identity claim satisfies verification policy.

Measure violations and tokens.

====================================================================== 60. CROSS-AGENT CONSISTENCY

Multiple agents should consume the same authoritative claim/evidence state.

Otherwise:

Agent A thinks Verified
Agent B thinks Unknown

Research distributed consistency for knowledge state.

====================================================================== 61. STALENESS

Agent context itself may become stale.

Version claim/evidence state.

Example:

ClaimVersion = 18

Capability issued against:

ClaimVersion 18

If evidence changes:

ClaimVersion 19

capability becomes stale.

This links epistemics with version-bound capability design.

====================================================================== 62. CLAIM IDENTITY

Claims need stable identity.

Example:

"Room 217 has a 36-inch bathroom door"

is distinct from:

"Hotel has accessible rooms"

Avoid vague propositions.

Research:

RDF triples
knowledge graphs
ontology modeling
proposition identity

====================================================================== 63. SUBJECT VERSIONING

Evidence may apply to a particular version/configuration.

Example:

Room accessibility measurement before renovation

should not automatically apply after renovation.

Claim scope may need:

subject version
effective period

====================================================================== 64. SEMANTIC GRANULARITY

Too coarse:

HotelAccessible = Verified

Better:

EntranceStepFree
DoorWidth
RollInShower
GrabBars

Research granularity tradeoffs.

Excessively fine claims may explode context and storage.

====================================================================== 65. ONTOLOGY RISK

A rich claim system may evolve into:

ontology/knowledge graph platform

This could create enormous complexity.

Investigate how to keep v0.1 bounded.

====================================================================== 66. MINIMUM VIABLE EPISTEMIC MODEL

Propose the smallest model that captures most value.

Possible:

Claim: id proposition subject

Evidence: id source observedAt supports/opposes provenance

Derived claim status: Unknown Supported Contested Verified

Verification record: policy authority timestamp

Compare with the existing 8-state proposal.

====================================================================== 67. STATE MACHINE VS DERIVED VIEW

Critical question:

Should epistemic status be authoritative mutable state?

Or:

derived from evidence graph + verification records?

If derived, status cannot drift independently.

This may be a major architecture change.

====================================================================== 68. EVENT SOURCING

Potential evidence events:

EvidenceAdded
EvidenceRetracted
EvidenceExpired
VerificationGranted
VerificationRevoked

Epistemic status can be projected.

Compare with event sourcing.

====================================================================== 69. SEMANTIC COMPILER

Possible compiler checks:

Verified claims must have verification producer
capability cannot require undefined epistemic status
invalidated evidence cannot support current verification
evidence dependency cycles
missing provenance
stale verification policy
impossible verification requirement

Investigate feasibility.

====================================================================== 70. CLAIM POLICY

Example:

claim FraudClear { verification: requires RiskAssessment maxAge 30m modelVersion approved }

The semantic compiler could generate:

verification checks
capability requirements
agent context

Compare with rule engines/policy engines.

====================================================================== 71. SINGLE SOURCE OF TRUTH RISK

If the verification policy is wrong:

all agents may be consistently wrong.

This creates correlated semantic failure.

Mitigation:

independent policy review
test cases
mutation testing
evidence audits
human oversight

Research this seriously.

====================================================================== 72. FALSE CERTAINTY RISK

A label:

Verified

may create psychological overconfidence.

Potential alternative terms:

VerifiedUnderPolicy
Accepted
Established
Qualified
Satisfied

Research terminology in regulated systems.

====================================================================== 73. FALSE NEGATIVE RISK

A claim may be true but remain:

Unsupported

because evidence is missing.

System must distinguish:

not proven

from:

false

This is crucial.

====================================================================== 74. AGENT GAMING

Agent tasked:

"Get this claim verified."

May attempt:

fabricate evidence
use low-quality source
change verification policy
duplicate evidence

Strong architecture should prevent shortcuts.

====================================================================== 75. EVIDENCE FABRICATION

Trusted evidence constructors must be protected.

AI-generated text should not become trusted evidence merely because it is stored in the system.

Investigate:

signed attestations
source authentication
content hashes
cryptographic provenance

====================================================================== 76. EXTERNAL SOURCE UNCERTAINTY

What if:

external API gives incorrect data?

Provenance lets system say:

verified source response

but not:

objectively true

Again distinguish:

procedural verification

from:

metaphysical truth.

====================================================================== 77. HUMAN TESTIMONY

Self-report may be legitimate evidence.

Example:

traveler reports fatigue

Do not rank all self-report as low quality automatically.

Evidence policy must be domain-specific.

====================================================================== 78. PRIVACY

Evidence/provenance systems may accumulate sensitive data.

Research:

minimization
selective disclosure
access control
retention
cryptographic commitments

Explicit evidence architecture has privacy costs.

====================================================================== 79. SECURITY

Attackers may target:

evidence
provenance
verification process

Threats:

evidence injection
replay
stale credential use
source spoofing
provenance tampering

Research trust architecture.

====================================================================== 80. EXPERIMENT — ASSUMPTION PROMOTION

Give agent:

ambiguous source saying X is probably true

Compare:

A. ordinary system

B. epistemic architecture

Then perform later tasks relying on X.

Measure:

whether X becomes treated as fact
downstream errors
tokens
corrective work

====================================================================== 81. EXPERIMENT — REPEATED REPORT

Provide the same original claim through five derivative sources.

Measure whether agent incorrectly interprets:

repetition

as:

independent corroboration

Epistemic provenance should expose shared source lineage.

====================================================================== 82. EXPERIMENT — EVIDENCE INVALIDATION

Start:

Claim = Verified

Invalidate foundational evidence.

Measure:

claim downgrade
capabilities removed
obligations generated
stale actions rejected

Compare with conventional system.

====================================================================== 83. EXPERIMENT — CONTRADICTORY EVIDENCE

Provide:

one strong supporting source
one strong opposing source

Measure:

whether agent invents certainty
whether architecture preserves conflict
downstream actions

====================================================================== 84. EXPERIMENT — POLICY CHANGE

Change verification policy.

Measure:

affected claims surfaced
token/context cost
obligations created
incorrect historical reinterpretation

====================================================================== 85. EXPERIMENT — AGENT RESTART

Agent 1 investigates claim.

Agent 2 starts fresh.

Compare:

A. raw history/context

B. authoritative epistemic record

Measure:

reconstruction tokens
disagreement
duplicated research

====================================================================== 86. EXPERIMENT — SMALL MODEL

Compare:

large model + raw evidence corpus

against:

smaller model + structured epistemic state

Measure:

decision correctness
false certainty
tokens
cost

====================================================================== 87. EPISTEMIC PRECISION/RECALL

Potential metrics:

Verification Precision:

claims marked verified that truly satisfy reference verification standard
-------------------------------------------------------------------------
claims marked verified

Verification Recall:

truly verifiable claims marked verified
---------------------------------------
truly verifiable claims

Also track:

false certainty rate
unsupported-action rate

====================================================================== 88. ASSUMPTION SURVIVAL RATE

Metric:

unsupported assumptions remaining consequential after N agent changes

Compare architectures longitudinally.

====================================================================== 89. EVIDENCE RECONSTRUCTION COST

Measure:

tokens/tool calls required to reconstruct why claim is accepted

Compare:

conventional records

against:

explicit provenance/evidence graph.

====================================================================== 90. COST PER JUSTIFIED DECISION

Potential economic metric:

total agent execution cost
--------------------------
consequential decisions with adequate evidence/provenance

This may be more meaningful than raw token count.

====================================================================== 91. LONGITUDINAL TEST

Run:

50–100 agent modifications/decisions

Track:

assumptions introduced
assumptions promoted
verification errors
stale evidence
inherited beliefs
correction cost
context tokens

Hypothesis:

explicit epistemics prevents compounding assumption drift.

====================================================================== 92. COUNTERARGUMENTS

Actively test:

  1. The epistemic model becomes too complex.
  2. Most commercial software does not need this.
  3. Verification status creates false confidence.
  4. Evidence graphs consume too many tokens.
  5. Agents still need raw sources.
  6. Human judgment cannot be encoded.
  7. Probability/confidence already solves most of the problem.
  8. Provenance is expensive to maintain.
  9. Evidence policies become brittle.
  10. Contradiction logic becomes a research project itself.
  11. Knowledge graphs introduce too much infrastructure.
  12. Model outputs may still influence verification indirectly.
  13. Users dislike being told "unknown."
  14. Systems may become unable to act under uncertainty.
  15. Structured epistemics may slow product development.

====================================================================== 93. WHERE NOT TO USE

Identify domains where explicit epistemic state is unnecessary.

Likely examples:

pure computation
transient UI state
low-consequence preferences
obvious deterministic values

Use only where:

provenance
uncertainty
evidence
authority
consequential decisions

matter.

====================================================================== 94. EXISTING RESEARCH TO PRIORITIZE

Search:

truth maintenance systems
assumption-based TMS
AGM belief revision
epistemic logic
dynamic epistemic logic
paraconsistent logic
Belnap four-valued logic
Dung argumentation frameworks
defeasible reasoning
Dempster-Shafer theory
Bayesian evidence
W3C PROV
database provenance
scientific workflow provenance
clinical evidence systems
source trust
trust management
knowledge graph provenance
AI uncertainty calibration
hallucination mitigation
longitudinal agent belief errors

====================================================================== 95. SOURCE QUALITY

Prefer:

original papers
standards
foundational texts
primary empirical research
official provenance specifications

Clearly distinguish:

established formal result
empirical AI evidence
architecture inference
speculation

====================================================================== 96. REQUIRED OUTPUT

Produce:

  1. Executive verdict
  2. Formal definition of Claim
  3. Formal definition of Evidence
  4. Epistemic-state critique
  5. State machine vs lattice analysis
  6. Status vs confidence analysis
  7. Unknown/false analysis
  8. Reported/verified analysis
  9. Assumption modeling
  10. Inference modeling
  11. Contradiction modeling
  12. Invalidation modeling
  13. Truth-maintenance comparison
  14. Belief-revision comparison
  15. Epistemic-logic comparison
  16. Provenance comparison
  17. Evidence-theory comparison
  18. Argumentation-framework comparison
  19. Bayesian/Dempster-Shafer comparison
  20. Verification-policy model
  21. Evidence-authority model
  22. Freshness model
  23. Policy-version model
  24. Claim dependency/cascade model
  25. AI-agent assumption-drift analysis
  26. Context/token implications
  27. Smaller-model hypothesis
  28. Security implications
  29. Privacy implications
  30. Failure modes
  31. Proposed experiments
  32. Metrics
  33. Economic implications
  34. What is supported by existing research
  35. What remains speculative
  36. Architecture changes recommended
  37. Minimum viable epistemic model
  38. Final verdict

====================================================================== 97. FINAL VERDICT FORMAT

Answer:

Does explicit epistemic state reduce unsupported certainty? Strong evidence / Moderate evidence / Weak evidence / No evidence

Does evidence provenance reduce inherited agent assumptions? Strong evidence / Moderate evidence / Weak evidence / No evidence

Does structured epistemic state reduce context/token cost? Strong evidence / Moderate evidence / Weak evidence / No evidence

Can it enable smaller models? Strong evidence / Moderate evidence / Weak evidence / No evidence

Should epistemic status be: mutable state / derived state / lattice / multi-dimensional record / other

Should "Verified" be renamed? Yes / No / Maybe Recommendation: ...

Closest existing formalism: ...

Strongest prior art: ...

Most important architecture correction: ...

Biggest complexity risk: ...

Most valuable AI-specific benefit: ...

Most important missing experiment: ...

====================================================================== 98. RESEARCH STANDARD

Be skeptical.

Do not assume an eight-case union is the right epistemic model.

Do not treat confidence as verification.

Do not treat verification as objective truth.

Do not assume more evidence means better evidence.

Do not count repeated copies of one source as independent evidence.

Do not allow Unknown to collapse into False without explicit closed-world policy.

Do not assume AI-generated text can serve as trusted evidence.

Do not build a full knowledge-representation platform unless evidence shows it is necessary.

The central question is:

Can explicit claims, evidence, provenance, and epistemic authority prevent
AI agents from turning uncertain assumptions into durable system truth —
while remaining simple enough to improve rather than overwhelm the
architecture?