Answer A wins
5 to 1.
Six outputs are treated as six signals.
EPISTEMIC INFRASTRUCTURE FOR AI AGENTS
An echo is not
a witness.
When AI agents agree, Minority Prophet tests whether the evidence is independent—or whether the crowd is repeating one recorded source.
MP.01 · SYNTHETIC FIXTURE5 : 1agent votes before lineage
A1ORIGINSource X
A2DERIVEDA1
A3DERIVEDA1
A4SUMMARYA2
A5PARAPHRASEA3
B1ORIGINSource Y
5 votes for A→1 recorded root
1 vote for B→1 recorded root
01 / THE FAILURE
Five votes.
Two evidence roots.
Everything completed successfully. No tool failed. No agent timed out. Five agents returned Answer A. But their declared ancestry leads back to one source.
Truth is not popularity. Consensus is not independent evidence.
Answer A wins
The majority
disappears.
A1 rootclaim-a1
=B1 rootclaim-b1
Preserve the minority. Request another independent source.
The point: five agreeing agents can still carry one piece of evidence. Lineage preserves the dissent without pretending either answer is proven.Inspect the result →
01B / SOURCE-FAMILY RECONSTRUCTION
Six claims.
Three actual sources.
The original human explainer asked whether people could distinguish the number of voices from the number of underlying evidence families.
weather-17 original
PROCEEDweather-cache copy of A
PROCEEDroute-clearance-9 original
PROCEEDmaintenance-alert-4 original
STOPmaintenance-summary summary of D
STOPweather-brief summary of B
PROCEEDTRACE
ANCESTRY→Nothing deleted.
Counting unit changed.
ANCESTRY→Nothing deleted.
Counting unit changed.
Weather source1 original → 1 copy → 1 summary
ABF
PROCEEDRoute-clearance source1 independent original
C
PROCEEDMaintenance source1 original → 1 summary
DE
STOPWhat it shows: agreement can shrink when copies are grouped by ancestry. It preserves the minority record; it does not automatically make the minority correct or authorize an action.
02 / THE LOOP
Detect. Trace.
Challenge. Verify.
Minority Prophet turns provenance into an inspectable decision process. The assessment can lower confidence, preserve dissent, request evidence, or route the decision onward. It never grants authority.
Detect
Find recorded consensus collapse, unsupported claims, circularity, and disappearing dissent.
Trace
Reconstruct claims, sources, transformations, agents, and recorded roots.
Challenge
Test whether corroboration is independent and whether more evidence is required.
Verify
Seek a new root, revise the belief, and preserve the failure as a regression test.
03 / EVIDENCE
See the failure.
Measure the difference.
Controlled studies show what provenance changes, where deterministic checks help, and where the evidence still runs out.
EPISTEMIC LIFT · CONTROLLED STUDY
Baseline → Provenance
→ Minority Prophet.
Two models faced the same 32 frozen worlds in three controlled conditions. Only the available epistemic information changed.
A 12.5%
B 68.75%
C 96.875%
A 3.125%
B 68.75%
C 90.625%
192/192completed cells
0failures
Bothpaired p < 0.05
Synthetic, matched worlds. Useful evidence—not a claim about every model or every deployment.
Inspect the lift →METHOD COMPARISON
Same packet.
Different methods.
AI reasoning, tool-using agents, a conventional vote, and a deterministic evidence rule faced the same eight cases.
128/128deterministic result
18.7 msobserved wall time
Per modelcosts kept separate
Deterministic rule $0 modelGPT Terra A ≈ $0.89Claude Opus A ≈ $3.25
A method comparison, not a model ranking. Eight cases; time and cost stay attached to each run.
Compare the methods →EPISTEMIC OBSERVATORY
Three witnesses.
Ninety-five echoes.
Enter a generated world where a copied majority overwhelms the count while three independent instruments retain the constructed truth.
3truth roots
95copied claims
$0model calls in demo
A generated teaching world. The ancestry is known so you can see the mechanism clearly.
Enter the observatory →04 / THE TRUST BOUNDARY
Evidence is not
authority.
Minority Prophet asks whether evidence is independently grounded at the causal boundary relevant to a stated decision. It preserves deeper lineage, exposes when another reasonable cut materially changes settlement, and never turns that assessment into permission.
Evidence arrives
A trace, score, test result, or replay bundle records what happened.
Context binds
The proposition, failure domain, independence cut, threshold, policy, and proposed action are joined.
Prophet assesses
Roots are counted at the declared proximal boundary while the full ancestry and minority signal remain visible.
Policy enforces
A separate policy proceeds, blocks, or escalates. Assessment never grants authority.
05 / FIRST PRINCIPLES
Never confuse—
ConsensusTruth
PopularityEvidence
ConfidenceCorrectness
ReputationCompetence
CorrelationIndependence
MajorityReality
06 / REPRODUCIBLE FIXTURE
Reproduce the
result.
minority-prophet / MP.01
$ python -m experiments.mp01.run_mp01
agent votes: A=5, B=1
recorded roots: A=1, B=1
ABSTAIN · PRESERVE_MINORITY