A trading-grade validation harness: train/validation/test/holdout splits, six null models, and metrics for graphlet recurrence and out-of-sample latent-edge prediction — so emergence has to beat chance.

title: "Experiment Protocol for Latent-Structure Discovery" created: "2026-06-20" artifact_type: "research_protocol" status: "draft v0.1" purpose: "Define a trading-model-like experimental framework: train/test splits, out-of-sample validation, graphlet discovery, hidden-structure prediction, and next-best-question selection."

Experiment Protocol for Latent-Structure Discovery

0. Core Aim

We want to discover whether strange historical, religious, cultural, occult, intelligence, and conspiracy-adjacent material contains recurring latent structure.

The experiment should proceed like a trading model:

Do not overfit.
Do not let the same data generate and validate the hypothesis.
Do not tune until every anomaly fits.
Preserve out-of-sample material.
Reward prediction, not post-hoc explanation.

1. What We Are Testing

The primary test is not:

Is a specific metaphysical claim true?

The primary test is:

Do independent subgraphs converge on recurring graphlets, latent structures, missing edges, or mechanism families more strongly than expected under null models?

A stronger test:

Do graph-derived predictions about missing nodes/edges get filled by later-ingested information at rates above baseline?

2. Research Objects

2.1 Local subgraphs

Each culture/religion/case becomes a local subgraph.

Examples:

Christianity
Judaism
Islam
Buddhism
Gnosticism
Hinduism
Zoroastrianism
Indigenous wetiko traditions
Yazidi / Mesopotamian motifs
Norse myth
Greek/Roman cults
Egyptian myth
West African / Afro-diasporic spirits
Chinese demonology / hungry ghosts
Japanese yōkai / possession motifs
Modern cults
Modern intelligence programs
Modern extremist movements

2.2 Canonical role layer

Local concepts are mapped into canonical roles with weights, not collapsed into identity.

jinn != demon != archon != Mara != wetiko

But they may share weighted roles:

whispering influence
deceptive authority
possession influence
moral inversion
human drain
anti-awakening force

2.3 Graphlets

Graphlets are typed recurring structures.

Examples:

whispering adversary
controlled inversion
manufactured criminality
legitimacy laundering
source suppression
purification/remediation
apocalyptic acceleration
elite-deception motif
sacrificial-violence loop
hidden-hierarchy doctrine

3. Dataset Splits

3.1 Temporal split

Use older material to generate graphlets, later material to test.

Train: pre-1500 religious/cultural texts
Validation: 1500-1900 esoteric/religious reform movements
Test: 1900-2000 political cults/intelligence/counterculture
Final holdout: 2000-present algorithmic/extremist/AI-era cases

3.2 Cultural split

Generate graphlets from some traditions, test on unseen traditions.

Example:

Train:
  Christianity, Judaism, Greek/Roman, Gnosticism

Validation:
  Islam, Buddhism, Hinduism, Zoroastrianism

Test:
  Indigenous wetiko, Chinese hungry ghosts, Japanese yōkai, African diaspora

Holdout:
  modern cults, intelligence programs, extremist networks

3.3 Mechanism split

Generate mechanisms from historical cults, test on religious motifs — or reverse.

Train mechanisms on:
  Jonestown, SLA, O9A, 764, Epstein-like blackmail networks

Test mechanism analogues in:
  religious warnings about possession, temptation, corruption, false prophet, hidden hierarchy

3.4 Source-family split

Ensure that validation evidence is not copied from training evidence.

No source-family leakage.
No quote-chain leakage.
No derivative blog/video leakage.
No repeated rumor counted as independent confirmation.

3.5 Analyst split

If possible:

Analyst A builds graphlets.
Analyst B, blind to graphlet hypotheses, maps unseen traditions.
Compare predicted convergence after mapping.

4. Null Models

We need null models, otherwise all pattern discovery becomes self-confirming.

4.1 Random rewiring null

Preserve degree distribution, randomize edges.

Question:

Do observed graphlets recur more often than in degree-preserving random graphs?

4.2 Label-shuffle null

Preserve topology, shuffle canonical-role labels.

Question:

Is convergence driven by actual role structure or by generic graph density?

4.3 Source-copy null

Collapse dependent source families.

Question:

Does convergence survive when copied sources count as one?

4.4 Motif-universal null

Some motifs are common because humans are humans.

Question:

Does this graphlet exceed baseline archetype recurrence?

4.5 Prestige-inversion null

Test with and without source-authority weighting.

Question:

Does the structure only appear when stigmatized sources are upweighted or prestigious sources downweighted?

4.6 Analyst-overfit null

Have a second analyst map concepts without knowing desired convergence.

Question:

Does the same structure appear under blind mapping?

5. Experimental Phases

Phase 0: Schema lock

Before ingesting the next large corpus, freeze:

node types
edge types
canonical roles
claim statuses
bridge qualities
weight formulas
graphlet definitions
holdout sets

Phase 1: Initial graphlet library

Create broad graphlet templates without overconstraining.

Initial graphlets:

whispering adversary
controlled inversion
manufactured criminality
legitimacy laundering
source suppression
purification/remediation
apocalyptic acceleration
hidden hierarchy
false light
sacrificial violence
elite corruption
attention drain
generative redirection

Phase 2: Training subgraphs

Ingest training traditions/cases.

Output:

candidate graphlets
canonical role mappings
latent structures
next-best questions

Phase 3: Validation subgraphs

Ingest validation traditions/cases.

Measure:

graphlet recurrence
mapping stability
latent prediction accuracy
alternative explanation survival

Phase 4: Test subgraphs

No tuning. Ingest test cases and score.

Measure:

out-of-sample graphlet recurrence
false-positive rate under nulls
prediction fill rate
missing-edge accuracy

Phase 5: Final holdout

Use modern cases only after the model is frozen.

This is where the system earns credibility.

6. Metrics

6.1 Graphlet recurrence score

GRS(g) =
    count_independent_subgraphs_containing_g
  * average_mapping_confidence
  * average_edge_quality
  * independence_weight
  * anti_overfit_penalty

6.2 Latent prediction hit rate

LPHR =
    confirmed_latent_structures
  / total_testable_latent_predictions

6.3 Prediction specificity

Specificity is high when the prediction names:
  - node class
  - relation type
  - time window
  - source type likely to contain evidence
  - discriminating alternative

Bad:

There will be hidden connections.

Good:

If graphlet X is active in case Y, we should find a funding intermediary or legal-suppression event within ±5 years of the public scandal.

6.4 Alternative-explanation defeat rate

AEDR =
    number_of_cases_where_primary_structure_beats_best_alt
  / total_cases_with_alt_explanations

6.5 Holdout convergence score

HCS =
    recurrence_on_holdout
  / recurrence_expected_under_null

6.6 Structural cavity score

SCS(L) =
    anomaly_pressure
  * mechanism_fit
  * recurrence_across_subgraphs
  * temporal_coherence
  * prediction_specificity
  * independence_weight

7. Next-Best-Question Engine

The interface should continuously rank questions.

EIG(question) =
    uncertainty_reduction
  * salience_of_resolved_node
  * number_of_hypotheses_discriminated
  * feasibility
  * source_availability
  * safety_factor

Question types:

translation question
source-family question
missing-edge question
alternative-explanation question
prediction question
negative-control question
mechanism question
timeline question
actor-network question
symbolic-equivalence question

Example:

question: "Does the whispering adversary graphlet recur in Islamic waswasa, Buddhist Mara, Christian demonic temptation, and wetiko literature under blind mapping?"
hypotheses_discriminated:
  - universal_psychological_archetype
  - shared_metaphysical_entity
  - later_interpretive_overlay
  - translator_imposed_similarity
expected_information_gain: high

8. Out-of-Sample Discipline

Rules:

1. No changing graphlet definitions after seeing test-set results.
2. No adding new canonical roles to rescue a failed prediction unless the result is marked exploratory.
3. No counting copied source families as independent.
4. No accepting symbolic convergence as causal convergence.
5. No deleting failed predictions.
6. No collapsing rival explanations until tested.
7. No fact export from low-confidence/high-salience nodes.
8. No cross-cultural SAME_AS without direct lineage or translation basis.
9. All rejected or failed theories remain in the archive for calibration.
10. Every model version receives a versioned schema hash.

9. Suggested Initial Experiment Set

Experiment A: Whispering adversary

Goal:

Test whether traditions independently describe a hostile influence that whispers, tempts, implants thought, or distorts perception.

Train:

Christianity, Judaism, Gnosticism

Validation:

Islam, Buddhism, Hinduism

Test:

Indigenous wetiko, Chinese hungry ghosts, Japanese yōkai, African diaspora

Holdout:

modern cult testimony, extremist radicalization, algorithmic rage loops

Experiment B: Controlled inversion

Goal:

Test whether ritualized inversion functions as pressure release that stabilizes hierarchy.

Train:

Saturnalia, Carnival, Feast of Fools

Validation:

trickster rites, liminal festivals, anti-structure rituals

Test:

modern protest festivals, controlled opposition cases, media outrage cycles

Experiment C: Manufactured criminality

Goal:

Test whether isolation/confession/transgression/blackmail/new-identity/deployment recurs across cults and extremist networks.

Train:

Jonestown, O9A, 764

Validation:

NXIVM, Children of God, Aum Shinrikyo, extremist grooming cases

Test:

political discrediting operations, gang recruitment, intelligence kompromat cases

Experiment D: Legitimacy laundering

Goal:

Test whether sacred/moral traditions are repeatedly used to legitimate state, financial, or intelligence projects.

Train:

Lateran Treaty, Reichskonkordat, Balfour/Scofield bridge

Validation:

imperial religious charters, Cold War religious fronts, NGO-civil society fronts

Test:

modern platform morality, human-rights language used for war, development-bank moral language

Experiment E: Source suppression

Goal:

Test whether anomalous scandals show repeated record-sealing, classification, missing autopsies, missing client lists, or public-narrative simplification.

Train:

Jonestown, Presidio, Epstein, P2/Banco Ambrosiano

Validation:

MKUltra, COINTELPRO, church abuse archives, intelligence disclosures

Test:

modern algorithmic censorship/disclosure controversies

Experiment F: Remedy convergence

Goal:

Test whether traditions converge on awareness, naming, purification, prayer, discipline, mindfulness, or generative redirection as remedy for adversarial influence.

Train:

Christian discernment, Jewish heshbon ha-nefesh, Buddhist mindfulness

Validation:

Islamic dhikr/ruqyah/waswasa handling, Hindu mantra/yoga, Indigenous purification

Test:

modern trauma therapy, cognitive defusion, nervous-system regulation, anti-radicalization methods

10. Success Criteria

The program succeeds if it can:

1. Preserve anomaly without credulity.
2. Generate graphlets before seeing holdout data.
3. Predict missing structures.
4. Score whether new evidence fills predicted cavities.
5. Separate surface-language difference from structural equivalence.
6. Penalize overfit and source copying.
7. Allow consensus to lose if posterior evidence warrants.
8. Allow strange hypotheses to fail cleanly without deleting useful subclaims.

The real product is not a single answer.

It is a ranked list:

what is known
what is alleged
what is structurally expected
what is missing
what would most reduce uncertainty next