Case Zero
Temporal Evidence Protocol

Protocol ID

CZ-TEP-001

Primary hypothesis

Latent Cross-Domain Emergence Hypothesis (LCDEH)

Case

Jeramie / Neuruh, Case Zero

Version

1.0, July 30, 2026

Primary unit

Relational schema, sealed prediction, independently coded architecture feature

Primary endpoint

Specificity-adjusted prediction performance above controls under low-leakage conditions

Change control

Any post-freeze deviation requires a numbered decision receipt and sensitivity analysis

Source document

10_Case_Zero_Temporal_Evidence_Protocol_v1.0_1.pdf (36 pages)

1. Research question and hypotheses

1.1 Primary hypothesis, LCDEH-H1

Deep experience across structurally distinct domains creates latent relational schemas. When a later target environment generates structural resonance and productive polarity with several such schemas, the schemas can recombine and reorganize into an integrated capability. This capability should appear faster, transfer more broadly, and require less target-specific instruction than would be expected from equivalent target exposure without the prior cross-domain topology.

LCDE = f(D, H, R, M, S, C, A) / (I + F)

Conceptual variable map only, per the source. Not yet a validated law. The demo does not compute a numeric LCDE value, because the source defines no arithmetic for f().

1.2 Confirmatory hypotheses H1 to H8

1.3 Null and competing hypotheses

2. Scope, claims, and non-claims

3. The 8-stage closed loop

The protocol is a retrospective longitudinal archive study executed prospectively from the moment of protocol freeze. The historical source material already exists, but the hypotheses, cutoffs, codebooks, prediction rules, scoring system, and control plan are frozen before the evaluators inspect the target outcome set. The ten stage-plan entries in Section 20 of the source (stages 0 to 9) are grouped here into the eight-stage closed loop shown in Figure 1, from preregistration to research ruling.

Roles, blinding, and independence

A single-person archive invites confirmation bias. The protocol separates evidence preparation, schema construction, prediction, outcome coding, scoring, and audit. One person may fill multiple operational roles only when the conflicting stages remain blinded and all transitions are documented.

Minimum independence standard. At least two independent atom coders and two independent match evaluators. The person who defines the target feature catalog cannot create the pre-target predictions. Predictions and outcome catalog must be separately hashed before reveal. Adjudicators see disagreements only after initial independent ratings are locked. AI assistance is allowed only under logged, versioned prompts and must not bridge blinded data compartments.

4. Temporal epochs and cutoff adjudication

The study depends on temporal precedence. Dates cannot be reconstructed from memory alone. Every date receives an evidence class and confidence score.

Date evidence classes

Contamination classes

5. The evidence stack (L0 to L7)

Every research claim must retain a complete evidence path. The unit-of-analysis hierarchy runs from source artifact to research claim.

Provenance model

The evidence graph follows a simple provenance pattern: an Entity (artifact or derived object) was generated by an Activity (extraction, coding, merging, prediction, scoring) associated with an Agent (human or model). This mirrors the entity-activity-agent distinction used in the W3C PROV family and makes derived conclusions auditable.

Source manifest minimum fields

source_id, system, original_path, export_path, sha256, created_at, modified_at, date_class, date_confidence, author_or_speaker, source_domain, epoch, contamination_status, privacy_class, duplicate_group, ingestion_status, notes

Source classes and primary risks

6. Blind atomization and schema induction

Atomization converts domain-heavy passages into auditable relational statements without importing the target outcome. It is the most important anti-hindsight stage.

Atom types

Atomization rules

1. One atom expresses one relation. Split compound claims.
2. Preserve exact source spans and page/line/message identifiers.
3. Use neutral nouns; do not translate directly into AI architecture labels.
4. Separate what occurred from what the source interpreted it to mean.
5. Mark speaker, authorship class, confidence, and date class.
6. Record contradictions rather than forcing premature reconciliation.
7. Do not infer a schema from a single atom; schema induction occurs later.
8. Do not code an atom merely because it resembles a known Neuruh component.
9. Assign explicit exclusion reasons to rejected spans.

Required atom record

{
  "atom_id": "A-000001",
  "source_id": "SRC-0001",
  "source_span": "message 54, chars 112-344",
  "verbatim_excerpt": "...",
  "neutral_atom": "evidence strength should govern action authority",
  "atom_type": "decision_rule",
  "subject": "evidence_strength",
  "relation": "governs",
  "object": "action_authority",
  "source_domain": "real_estate_underwriting",
  "date": "YYYY-MM-DD",
  "date_confidence": 0.95,
  "authorship_class": "user_originated",
  "contamination_status": "C0",
  "coder_id": "C02",
  "confidence": 0.84
}

Coder calibration gates (frozen before primary coding)

Reliability is calculated before adjudication. The protocol recommends Krippendorff's alpha for multiple coders and missing ratings, with Cohen's kappa available for pairwise nominal checks. These are protocol governance thresholds, not universal statistical laws. They must be frozen before primary coding and reported alongside confidence intervals and category prevalence.

Schema eligibility

Candidate schema object

{
  "schema_id": "S-014",
  "neutral_name": "authority_scales_with_irreversibility",
  "definition": "As action irreversibility and downside increase, required evidence and authorization should also increase.",
  "supporting_atoms": ["A-0031", "A-0199", "A-0440"],
  "source_domains": ["real_estate", "defense", "operations"],
  "first_evidence_date": "YYYY-MM-DD",
  "mechanism": "risk-weighted governance",
  "boundary_conditions": ["low-cost reversible experiments"],
  "contradicting_atoms": ["A-0874"],
  "status": "candidate",
  "confidence": 0.79
}

Schema graph: node and edge classes

The graph release packet contains neutral schema names and definitions, relation edges and polarity pairs, domain-count and time-depth metadata, and confidence, contradiction count, and evidence-independence count. It contains no Neuruh names, no post-cutoff terms, no current product catalog, and no narrative biography.

Sealed prediction generation

Predictions must be specific enough to fail. "The person will build systems" is worthless. A valid prediction states the likely target capability, causal path, architecture form, expected boundaries, and evidence that would disconfirm it.

Prediction quality gates

Prediction object

{
  "prediction_id": "P-007",
  "neutral_capability": "risk-weighted agent authority and evidence gates",
  "required_schema_ids": ["S-014", "S-029"],
  "causal_path": "irreversibility -> evidence threshold -> authority boundary -> logged action",
  "expected_manifestations": [
    "different actions require different authorization tiers",
    "high-risk actions retain provenance and rollback information",
    "confidence affects whether the system acts or escalates"
  ],
  "specificity": 0.88,
  "confidence": 0.74,
  "falsifier": "later architecture grants uniform autonomy without risk-sensitive evidence or escalation",
  "graph_version": "G-1.0",
  "sealed_at": "timestamp",
  "sha256": "..."
}

Worked example, Appendix E: neutral mapping

Outcome architecture gold standard

The outcome team independently codes what later existed. It does not ask whether a feature matches the theory. It records architecture, first appearance, implementation evidence, and importance.

Feature inclusion rule. The feature must be observable in dated artifacts, code, architecture documents, or repeated operational behavior. A single assistant suggestion without later adoption is not a realized outcome feature. First appearance and mature canonical appearance are stored separately. Features are described in neutral language before predictions are revealed. Importance weights are frozen before scoring and justified by architecture centrality, not match convenience.

7. Controls, ablations, and counterfactuals

The study is meaningful only if the Case Zero graph beats plausible alternatives. Controls are not decorative. They are the mechanism by which a good story is forced to compete.

Control matching. Match corpus volume, date span, domain count, document types, and prediction count where possible. Give every prediction team the same instructions and output budget. Do not disclose which graph is Case Zero. Score all packets in randomized order. Include a leakage-positive control to prove the scoring system can detect an easy match.

Domain ablations

Counterfactuals

• Remove the theoretically relevant domain and recompute predictions.
• Remove a volume-matched unrelated domain as a negative ablation control.
• Replace removed atoms with random atoms to separate information quantity from content.
• Restrict the graph to single-domain schemas and compare against multi-domain schemas.
• Remove assistant-originated pre-target material to test user-originated structure.
• Remove all target-adjacent vocabulary and rerun the pipeline.

8. Scoring system and primary endpoints

Pairwise match rubric

Core metrics

Candidate formulas (source Section 16.3)

PP  = sum_i v_i * max_j(m_ij) / sum_i v_i
      Prediction weights v reflect preregistered specificity; m is blind match score.

FR  = sum_j w_j * max_i(m_ij) / sum_j w_j
      Outcome weights w reflect preregistered architecture importance.

F1  = 2 * PP * FR / (PP + FR)
      Defined as zero when both terms are zero.

SAS = F1 * mean(specificity_i * mechanism_quality_i)
      Primary descriptive score.

BL  = SAS_case - mean(SAS_baselines)
      Also report standardized lift relative to baseline dispersion.

AS_d = (SAS_full - SAS_remove(d)) - (SAS_full - SAS_remove(control))
      Domain-specific ablation effect.

Prediction specificity scale (Appendix B.1)

Mechanism-quality judgment (Appendix B.2)

• Does the match preserve the same relation structure?
• Does it require the identified source schemas?
• Does the predicted implementation follow from the mechanism?
• Could surface terms alone have produced the match?
• Does the outcome violate any preregistered boundary condition?
• Would an independent evaluator recognize the mapping without the founder narrative?

9. Ruling engine: frozen gates

Thresholds below are provisional v1.0 gates. They must be frozen before primary data inspection. A result can be theoretically interesting while failing the confirmatory gate.

Ruling classes

Statistical analysis plan (summary)

• Reliability: Krippendorff's alpha for nominal and ordinal coding; ICC or rank correlation for continuous scores; raw agreement and category prevalence reported alongside corrected coefficients; all reliability metrics calculated before adjudication.
• Primary comparison: compute SAS for Case Zero and each matched control packet; generate a permutation null by shuffling graph relations, domain labels, or prediction-outcome assignment as preregistered; at least 10,000 permutations where computation permits; report exact permutation p-value, baseline percentile, standardized effect, and bootstrap confidence interval over outcome features; repeat after removing target-adjacent terms and assistant-originated source material; report all confirmatory analyses, including failures.
• Missing data: missing timestamps, uncertain authorship, and ambiguous matches remain missing or quarantined; they are not imputed into a favorable state. Sensitivity analyses may assign conservative lower and upper bounds. Any discretionary adjudication is recorded before aggregate results are viewed where practical.

Bias, leakage, and failure modes

Falsification conditions

• The claimed relational structures do not appear in clean pre-AI records.
• They appear only after target vocabulary or AI collaboration begins.
• Independent coders cannot reproduce atoms, schemas, or match judgments.
• Generic or randomized controls predict the later architecture equally well.
• Removing purported source domains does not selectively weaken predictions.
• All impressive matches are generic, linguistic, or obvious from the prompt.
• Development is fully explained by ordinary exposure, iteration, and AI assistance.
• The hypothesis can reinterpret every outcome and therefore risks no failure.

10. Pilot, full-scale study, and repository

Minimum viable pilot (v1.0 targets)

Pilot completion gates

• Corpus and prediction sealing works without leakage.
• Atom and match reliability reaches the gate.
• Prediction packets are specific enough to produce both hits and misses.
• At least one control meaningfully separates from Case Zero or reveals a design flaw.
• Review workload, cost, and token requirements are measured.
• The team can reconstruct every score back to exact source evidence.

Full-scale extensions

• Complete corpus inventory and deduplication across ChatGPT, Claude, Codex, Notion, local files, code, and operational records.
• Semantic and symbolic retrieval for candidate atom discovery, followed by human verification.
• Multiple graph representations: relation triples, causal graphs, polarity pairs, and temporal layers.
• Model-versus-human prediction teams to test whether the graph, not a specific reasoner, carries the signal.
• Architecture-family-specific replications and pre-registered feature holdouts.
• Replication on additional founders or builders with deep multidomain histories.
• Prospective extension: use the current graph to predict a future Neuruh capability before it is built.

Research roadmap after Case Zero

Recommended repository

case-zero/
  00_protocol/   protocol_v1.0.docx, preregistration.md, decision_register.csv, deviations.csv
  01_sources/    originals_read_only/, derivatives/, source_manifest.csv, hashes.sha256
  02_temporal/   cutoff_evidence/, epoch_assignments.csv, contamination_log.csv
  03_atoms/      codebook.md, calibration/, atoms.csv, reliability_report.json
  04_schemas/    schemas.jsonl, graph.json, contradictions.csv, graph_release_packet/
  05_predictions/ sealed_packets/, prediction_manifest.csv
  06_outcomes/   outcome_features.csv, outcome_evidence/
  07_controls/
  08_scoring/    match_matrix.csv, adjudications.csv, metrics.json
  09_analysis/   scripts/, figures/, permutation_results/
  10_reports/    audit_report.md, case_zero_report_v1.0.docx

Required deliverables

11. Governance and change control

Decision receipts

Every change after protocol freeze receives a decision receipt. The receipt includes the trigger, options considered, selected change, reason, decision maker, affected objects, whether outcomes were visible, and planned sensitivity analysis.

decision_id, timestamp, stage, trigger, options_considered, selected_action, rationale, outcome_data_visible, affected_objects, decision_maker, auditor_review, sensitivity_analysis, status

Deviation classes

Approval gates

Publication rule. The final report publishes the protocol, deviations, exclusions, failed predictions, control results, ablations, and uncertainty. Proprietary source content may remain private, but an auditor must verify that reported paths exist. Public claims must not exceed the ruling class.

12. Preregistration checklist and completeness meter

Appendix F of the source lists 20 freeze items. Check each item as it is completed. The meter is a documentation completeness tracker for this demo, not a substitute for the protocol's timestamped preregistration record.

13. Try-it walkthrough

1. Load Sample A. Click "Load Sample A: Appendix E worked example" in the match-scoring lab. The matrix fills with the source worked example: three predictions against three outcome features. Read the computed PP (0.851), FR (0.854), F1 (0.853), and SAS (0.504). These are the hand-computed values from the source formulas.
2. Change one match. Set prediction P2 against outcome O2 from 0.75 to 0.25. Watch PP, FR, F1, and SAS fall. This is the rubric doing its work: one broken structural link costs precision and coverage at once.
3. Run the ruling engine. Click "Use lab SAS and mean specificity", then "Issue ruling". With the default Sample A inputs the ruling is Supported, because the sample was constructed to pass every gate. Change the permutation p-value to 0.05 and re-issue: the primary endpoint fails and the ruling drops to Unsupported.
4. Load Sample C. The leaky pilot sets LLI to 0.12. Issue the ruling again: the leakage gate fails and the ruling becomes Protocol failure, because a contaminated prediction corpus cannot be interpreted, however high it scores.
5. Adjudicate a cutoff. In the cutoff lab, set the artifact date after T0 and adjudicate: the artifact is excluded from predictors and retained only as outcome evidence. Set the date class to D and adjudicate: the artifact is quarantined for primary analysis regardless of its date.
6. Complete the checklist. Check freeze items in Section 12 and watch the completeness meter rise. The missing-fields list shrinks as items are checked. Copy the share link to save this exact state.

14. How it works / What this is not

15. Honest gaps

• No real data. The match lab, ruling engine, and cutoff adjudicator run on sample or user-entered values. No archive has been inventoried, hashed, or atomized here.
• LCDE has no arithmetic. The source defines the LCDE expression as a conceptual variable map only. This demo does not compute it, and no numeric LCDE value should be quoted from this page.
• Fifth secondary gate is a lab operationalization. The source ruling rule requires four of five mechanistic secondary gates but names four (TPS, CDRI, ablation, NPC). The fifth gate in this demo, mean specificity at or above 0.61, is derived from Appendix B.1 and labeled as such. It is not a frozen protocol threshold.
• 8-stage grouping is a presentation choice. The source stage plan runs stages 0 to 9. This demo groups them into eight loop stages; the full ten-stage table is shown in Section 3.
• Permutation testing is not executed. The ruling engine accepts a p-value as input. It does not run 10,000 permutations in the browser.
• No cryptographic sealing. The share link encodes state in a URL hash. It is not a hash-locked, timestamped prediction seal and must not be treated as one.
• Reliability is entered, not measured. Krippendorff's alpha is an input. This demo does not compute inter-rater agreement from coder ratings.
• LLI is entered, not computed. The Linguistic Leakage Index is an input. This demo does not scan text for target-adjacent terms.
• Layout not eyeballed in a real browser. The page was verified by automated DOM checks in jsdom, not by visual inspection in a browser.
• Source claims are labeled as source claims. Thresholds, durations, pilot targets, and rubric values are taken from Protocol v1.0 (July 30, 2026). They are provisional v1.0 gates per the source and must be frozen before primary data inspection.

16. References

Methodological references from Appendix H of the source. The protocol itself is the primary source for everything on this page.

1. Center for Open Science. OSF Registrations and Preregistrations documentation. Used as the model for freezing hypotheses, methods, materials, and analysis decisions before data reveal.
2. World Wide Web Consortium. PROV-O: The PROV Ontology. W3C Recommendation, 2013. Used for the entity-activity-agent provenance pattern.
3. Cohen, J. (1960). A Coefficient of Agreement for Nominal Scales. Educational and Psychological Measurement, 20(1), 37-46.
4. Gentner, D. (1983). Structure-Mapping: A Theoretical Framework for Analogy. Cognitive Science, 7(2), 155-170.
5. Gick, M. L., and Holyoak, K. J. (1983). Schema Induction and Analogical Transfer. Cognitive Psychology, 15(1), 1-38.
6. Patrick, J., Ahmed, A., Smy, V., Seeby, H., and Sambrooks, K. (2015). A Cognitive Procedure for Representation Change in Verbal Insight Problems. Journal of Experimental Psychology: Learning, Memory, and Cognition, 41(3), 746-759.
7. Hughes, J. (2021). krippendorffsalpha: An R Package for Measuring Agreement Using Krippendorff's Alpha Coefficient. arXiv:2103.12170.

Source document: 10_Case_Zero_Temporal_Evidence_Protocol_v1.0_1.pdf, Protocol CZ-TEP-001 v1.0, prepared July 30, 2026, 36 pages. Sibling build: Case Zero Live Demo. Hub: Neuruh Hub.