Neuruh Research Program — Protocol CZ-TEP-001

Case Zero, Built Runnable

The full Temporal Evidence Protocol v1.0 as working instruments: the eight-stage closed loop, the LCDE variable model, the evidence stack, cutoff adjudication, the blind atomization lab, schema eligibility, the sealed-prediction gates, the match-scoring laboratory that computes the protocol's own formulas, the ablation board, and the ruling engine that applies the protocol's frozen gates and returns Supported, Mixed, Unsupported, or Protocol Failure.

36 pages, 25 sections, Appendices A-H — read in full8 confirmatory hypotheses, 7 competing 8 controls B0-B77 support gates, 4 ruling classes 20-item preregistration freezeExecution standard: freeze first, predict blind, publish the misses

Fig 1The Closed-Loop Protocol

From preregistration to research ruling in eight stages. The protocol is designed to prevent hindsight, target-vocabulary leakage, and post hoc story construction. Six approval gates guard the loop: protocol freeze, corpus freeze, coder calibration, prediction seal, outcome seal, final ruling — each with a named owner and required evidence.

Idle. The decisive observation the loop hunts for: a blind evaluator, using only clean pre-AI relational evidence, predicts specific later Neuruh architecture features more accurately than evaluators using generic entrepreneurial material, shuffled archives, narrow source domains, or retrospective narratives.

§1The Claim and Its Enemies

Primary question: can a target-blind model of Jeramie's pre-AI relational schemas predict distinctive features of the later Neuruh AI architecture with materially greater specificity and accuracy than credible baseline and null models? The claim is not that broad experience is useful; that is already known. The distinctive claim is that deep, target-independent experience across structurally different domains creates a latent relational topology that predicts the later form of an integrated capability before the target domain becomes available.

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

Conceptual variable map only; not yet a validated law (the document's own label). Move the variables and watch the conceptual index; the denominator is the point: interference and fragility divide whatever the numerator builds.

Confirmatory (H1-H8) vs competing (H0, C1-C6)

Select a hypothesis.
Can support (§2.1)Temporal precedence of relational structures in the pre-AI archive; blinded, nontrivial prediction beyond lexical similarity; selective multi-domain contribution; a protocol that can reliably identify, score, and falsify candidate pathways; Neuruh as a possible externalized synthesis of prior schemas.
Cannot support (§2.2)A universal law of intelligence or consciousness; causal proof from a single retrospective case; superiority over AI specialists in every task; that every prior experience contributed; population-level estimates or commercial guarantees. This is a mechanism-detection and protocol-validation study. It can justify replication. It cannot carry a universal theory by itself.

§3The Evidence Stack (L0-L7)

Every research claim must retain a complete evidence path. The unit-of-analysis hierarchy runs from a preserved artifact to a research claim in eight levels, and the Appendix E worked example below walks one real thread up the whole stack. Click a level.

Select a level.

§5Roles, Blinding, and Independence

A single-person archive invites confirmation bias, so the protocol separates evidence preparation, schema construction, prediction, outcome coding, scoring, and audit: the recommended triple-blind structure. Click a role for its access, its forbidden knowledge, and its output.

Select a role.

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 are separately hashed before reveal; adjudicators see disagreements only after independent ratings lock; AI assistance only under logged, versioned prompts that never bridge blinded compartments.

§6Temporal Epochs and Cutoff Adjudication

The study depends on temporal precedence, and dates cannot be reconstructed from memory alone. Every date gets an evidence class and a confidence score. Adjudicate the sample artifacts: the ruling follows the protocol (conservatism wins; no convenient cutoff — the cutoff cannot be selected because it produces a better result).

§9Blind Atomization Lab

Atomization converts domain-heavy passages into auditable relational statements without importing the target outcome. It is the most important anti-hindsight stage. Pick a domain-heavy statement from the codebook (Appendix A.1) and see its acceptable neutral atom and its rejected target-loaded rewrite, side by side.

The atom record (§9.3)

Coder calibration gate (§9.4)

Protocol governance thresholds, not universal statistical laws. They must be frozen before primary coding and reported with confidence intervals and category prevalence. Reliability is calculated before adjudication.

§10-11Latent Schema Induction and the Schema Graph

A latent schema is not a clever summary. It is a recurring relational structure that appears independently across time or domains and preserves a coherent mechanism. Test the document's own candidate schema, S-014, against the seven eligibility criteria.

The graph the prediction team receives (§11)

Seven node classes (Schema, Constraint, Polarity, Value, Transformation, Domain, Evidence) and eight edge classes (supports/contradicts, generalizes_from, structurally_similar_to, activates_under, held_in_polarity_with, combines_with, bounded_by, precedes). The release packet carries neutral names, edges, domain-count and time-depth metadata, confidence and contradiction counts — and no Neuruh names, no post-cutoff terms, no product catalog, no narrative biography.

§12Sealed 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 the evidence that would disconfirm it. The document's example, P-007, is below, with the six quality gates as a live checklist.

Unsealed. Sealing hashes the prediction text and the gate state in this browser. Any later edit breaks the seal; that is the mechanism, not a metaphor.

§13Outcome Architecture Gold Standard

The outcome team independently codes what later existed. It does not ask whether a feature matches the theory. Inclusion rules: 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 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.

§16Match Scoring Laboratory

Click any cell to cycle its blind match score (0, 0.25, 0.50, 0.75, 1.00 — the §16.1 rubric: from no meaningful match, through vague thematic overlap and partial structural match, to a specific pre-registered prediction matching mechanism, architecture form, and boundary conditions). PP, FR, F1, and the primary descriptive score SAS compute exactly per §16.3.

Controls (§14): B0 generic founder, B1 narrow AI specialist, B2 random multidomain, B3 timestamp shuffle, B4 domain-label shuffle, B5 lexical baseline, B6 retrospective narrative (positive leakage control: should score high but fail integrity), B7 null graph. The primary endpoint (§16.4): the specificity-adjusted structural prediction score for the Case Zero graph, compared with the distribution of matched baselines and permutation nulls (at least 10,000 permutations where computation permits), subject to reliability, temporal-precedence, and leakage gates.

§15Domain Ablations and Counterfactual Tests

Ablation tests whether specific source domains contribute selectively, rather than merely increasing corpus size. The expected effect of removing each domain is preregistered before outcomes are revealed. Toggle a domain off to see its preregistered predicted weakening.

Counterfactual discipline (§15.1): 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; remove assistant-originated material to test user-originated structure; remove all target-adjacent vocabulary and rerun the pipeline.

§18Support, Revision, and Rejection — The Ruling Engine

Seven provisional v1.0 gates, frozen before primary data inspection. Set the study results and the engine applies the gates and the ruling classes exactly as written. A result can be theoretically interesting while failing the confirmatory gate.

No theory protectionIf the protocol produces an unfavorable result, the correct response is to revise the hypothesis or the measurement system — not to redefine emergence until the score turns green.

§19Bias, Leakage, and Failure Modes

Twelve named distortions, each with its control. Below them, the falsification conditions: the observations under which the hypothesis loses.

The eight falsification conditions (§19.1)
  1. The claimed relational structures do not appear in clean pre-AI records.
  2. They appear only after target vocabulary or AI collaboration begins.
  3. Independent coders cannot reproduce atoms, schemas, or match judgments.
  4. Generic or randomized controls predict the later architecture equally well.
  5. Removing purported source domains does not selectively weaken predictions.
  6. All impressive matches are generic, linguistic, or obvious from the prompt.
  7. Development is fully explained by ordinary exposure, iteration, and AI assistance.
  8. The hypothesis can reinterpret every outcome and therefore risks no failure.

§20-22Execution: Stages, Pilot, Full Study

Ten stages from freeze to ruling, a bounded minimum viable pilot before any full archive ingestion, and one extension the document calls the strongest: Case Zero II.

Minimum viable pilot (§21)100-300 pre-target artifacts (250k-1M words) across at least four domains; 1,000-3,000 coded atoms; 25-60 candidate schemas, 10-20 promoted; 20-40 sealed predictions capped in advance; 40-100 independently coded outcome features; controls: generic founder, randomized graph, lexical baseline, at least two domain ablations; two atom coders, two match raters, one auditor. Pilot gates: sealing works without leakage; reliability reaches the gate; packets produce both hits and misses; at least one control separates or reveals a design flaw; workload and cost are measured; every score reconstructs back to exact source evidence.
Case Zero II (§22.2)After the retrospective study, freeze the current latent schema graph and predict three to five Neuruh capabilities that do not yet exist. A future prospective hit would be stronger evidence than any retrospective match. The full roadmap after Case Zero (Appendix G): protocol validation, full retrospective case, prospective Case Zero, matched replication on other builders, experimental induction, computational simulation, theory comparison.

App FPreregistration Freeze — 20 Items

Nothing in this protocol means anything until these twenty items are frozen with an immutable timestamp. This checklist is the freeze instrument from Appendix F, in this browser only; checking boxes here does not preregister anything.

0 of 20 frozen.

Governance (§24): every post-freeze change receives a decision receipt. Deviations are classed D1 administrative (log), D2 methodological non-outcome-informed (log and test sensitivity), D3 outcome-aware methodological (primary confirmatory claim compromised; report as exploratory), D4 integrity breach (stop or restart the affected stage). Publication rule: the final report publishes the protocol, deviations, exclusions, failed predictions, control results, ablations, and uncertainty. Public claims must not exceed the ruling class.