We stand with people living with cystic fibrosis, their families, care teams, and researchers. This independent lab contributes donated time, research notes, and software experiments, not clinical care or medical advice. Support the wider work through the Cystic Fibrosis Foundation.

Donate to the Cystic Fibrosis Foundation (opens another site)

Reproducible Claim Checkers & Source-Grounded Council Loops

Reproducible Claim Checkers & Source-Grounded Loops: How check_claims.py and Multi-Lineage Councils Audit Scientific Prose: Synthesizing Pillar IV: how Answer Tiers (T0–T3), SHA-256 evidence locks, Scite retraction gates, and multi-lineage adversarial councils prevent false promises in AI-assisted research.

Why it matters: In a field where lineages and clinicians read every word with care, an AI assistant that invents a citation, misstates a drug label, or turns an early dish experiment into a clinical promise causes real harm. That is why every essay in this 24-monograph collection—and every tool in our repositories—is governed by a reproducible proof loop: machines propose, independent model lineages critique, reproducible Python checkers verify every number against primary records, and anything unverified is blocked from publication.

Research & Clinical Boundary: Independent research synthesis and biophysical modeling — not medical advice, clinical guidance, or a treatment promise. For clinical care decisions, always consult a specialized Cystic Fibrosis care team.
T0/T1/T2 · Sourced Epistemic Governance & Verification Model VOL. 03 · REV 1.1
Three optical flint glass prisms resting on handmade cotton rag paper with ink mathematical diagrams, refracting sunlight into emerald, sapphire, and amber caustic beams that converge to a bright white focus.
Plate L.06: Plate L.06 · Independent multi-lineage verification and epistemic truth: three independent analytical lenses converge onto a single verified coordinate before any scientific claim enters the public record. Primary Receipt: DOI 10.1056/NEJMoa1908639
About this illustration & epistemic provenance

Conceptual editorial visualization crafted for human comprehension of the physical mechanism described in Middleton et al. (2019), New England Journal of Medicine 381(19):1809-1819 — Elexacaftor–Tezacaftor–Ivacaftor for Cystic Fibrosis with a Single Phe508del Allele (PMID 31697873). (DOI 10.1056/NEJMoa1908639). Quantitative curves and parameter sensitivities are computed live in Act III below.

Act I · Physical & Clinical Story

Why Plausibility Is the Enemy of Scientific Truth

Every published clinical trial carries a profound moral weight when families and physicians depend upon reported percentages to guide therapeutic choices. When Middleton and colleagues published their landmark study in the New England Journal of Medicine in twenty nineteen, the reported improvements in respiratory volume established an exacting scientific benchmark. Modern generative artificial intelligence models can easily blur these factual boundaries because their statistical training optimizes for fluent plausibility rather than empirical truth. An ungrounded neural network might confidently invent an altered patient cohort or describe a preliminary cell culture experiment as though it were an approved clinical protocol. Within cystic fibrosis research, such subtle distortions are completely unacceptable because vulnerable people require absolute factual integrity. Our laboratory therefore treats every published number with rigorous formal skepticism by requiring automated proof loops that confirm all statements against durable evidence repositories before publication.

In Plain Words · Extra-Easy Explanation

Modern computer algorithms frequently generate smooth but factually unreliable biomedical prose. When language models discuss complex clinical trials, they often invent convincing numbers that sound remarkably authoritative to unsuspecting clinicians and affected families seeking reliable therapeutic knowledge. To solve this dangerous dilemma, we constructed an unyielding software verification referee. Our automated checker automatically parses every paragraph, extracting numerical findings, author citations, and specific drug names before testing them against verified original trial records. Any unanchored claim or modified percentage triggers an immediate build halt. By combining this mechanical gatekeeper with independent model review councils, we eliminate the correlated blind spots that plague solitary artificial intelligence writing systems. Factual precision replaces plausible fiction across our entire library. Every sentence surviving this rigorous evaluation pipeline provides researchers and patients with transparent confidence that reported outcomes reflect true biological discoveries without exaggeration.

Everyday Analogy & Interactive Controls Primer

Everyday Analogy: Consider the security inspection protocol at an international precious metal vault. Before an ingot enters the official vault repository, it does not rely on the visual impression of a single appraiser. First, a laser spectrometer measures elemental composition to verify metallic purity. Next, three independent expert appraisers examine the hallmark stamp from different viewpoints, ensuring that no single observer's eye fatigue approves a counterfeit bar. Finally, an automated acoustic resonance sensor measures the sound wave traveling through the metal core to confirm internal integrity. If any single measurement deviates from the established metallurgical standard, the pneumatic gate locks shut automatically. Plausibility never overrides measurement in this physical vault. However, this analogy breaks down because natural language assertions possess fluid semantic shades, whereas a physical ingot has a fixed density and crystalline structure that cannot alter its meaning based on sentence context.

How the Two Sliders Work in Plain Words: Before examining the mathematical derivations, the interactive simulator lets you test how different verification defenses perform together. The first slider controls the Independent Reviewer Lineages in Council, adjusting how many distinct artificial intelligence models evaluate every draft paragraph. Moving this control from one toward four lineages demonstrates how multi-model consensus rapidly reduces blind spots. The second slider sets the Reproducible Verifier Gate Coverage, representing the percentage of textual claims evaluated by automated software scripts. Below the sliders, four live readouts reveal the real-time health of the publishing pipeline. The Unanchored Numeric Claims Permitted display shows that every unsupported number is instantly halted at zero when gate coverage is complete. The Correlated Blind-Spot Escape Probability readout drops below four-thousandths as diverse reviewer lineages join the panel. Meanwhile, the Catalog Host and Terms Compliance indicator confirms adherence to data access boundaries, while the Epistemic Answer Tier Classification sorts each statement into its appropriate evidentiary category.

PLATE L.06-D · VISUAL DISTILLATION · INTUITION TO INVARIANT: Deterministic Verification Automata and Multi-Lineage Consensus 01 · EVERYDAY & CELLULAR INTUITION Triangulated Vault Inspection and Se Just as high-security bullion vaults require independent... False approval risk: P(Escape) < 0.0 02 · GOVERNING MATHEMATICAL MECHANISM Formal Gatekeeper and Epistemic Tier A formal Python automaton extracts all numeric claims, computing... Zero-tolerance invariant: Unanchored 03 · QUANTITATIVE INVARIANT & BOUNDARY Asymptotic Convergence to Zero Unver Coupling formal syntax verification with multi-lineage... Compliance: 100% catalog.yaml & 0 re PLATE L.06-D · VISUAL DISTILLATION · INTUITION TO INVARIANT · Deterministic Verification Automata and Multi-Lineage Consensus
Plate L.06-D (Deterministic Verification Automata and Multi-Lineage Consensus): Combining independent AI reviewer councils with formal code verification transforms plausible natural language drafts into mathematically uncompromised scientific publications.

From day one, the Sensei Ewok Research Lab was built around a simple architectural principle: treat natural-language scientific claims with the same formal skepticism that a compiler or a proof assistant applies to code. Every number on senseiewok.ai must trace to a primary record in data/evidence.snapshot.json or a reproducible biophysical kernel with explicit model boundaries.

The stakes of epistemic verification in cystic fibrosis could not be higher. When Middleton and coworkers published the Phase 3 clinical trial of elexacaftor, tezacaftor, and ivacaftor in the New England Journal of Medicine, they reported a 14.3 percentage point increase in percent predicted forced expiratory volume alongside a 41.8 millimoles per liter reduction in sweat chloride concentration. Those quantitative values redefined therapeutic expectations worldwide for people with a single Phe508del allele. If an artificial intelligence synthesis misquotes these baseline endpoints or conflates single-allele efficacy with homozygous responses, researchers make erroneous assumptions and families absorb false impressions. Our mathematical verification architecture ensures that every numerical assertion corresponds to a verified hash in primary evidence catalogs. By enforcing strict epistemic answer tiers, we formally distinguish between empirical clinical trial data, biophysical modeling simulations, and speculative research hypotheses. This separation prevents unverified computational claims from masquerading as established medicine, anchoring all exploratory tools to immutable scientific ground.

“Closes the 24-essay loop across all four Pillars: showing how anyone with a laptop can download our open-source cf-skills and SKILL.md companions to run the exact same anti-hallucination and biophysical checks on their own AI agents.”

— Cross-Pillar Research Thread · Grounded in Middleton et al. (2019), New England Journal of Medicine 381(19):1809-1819 — Elexacaftor–Tezacaftor–Ivacaftor for Cystic Fibrosis with a Single Phe508del Allele (PMID 31697873).

Key Parameters, Readouts & Primary Anchor

Quantity / Anchor Pinned Baseline Physical / Clinical Role
Independent Reviewer Lineages in Council (N_lineages) 3.0 lineages Primary perturbation axis (1.0 to 4.0 lineages)
Reproducible Verifier Gate Coverage 1.0 fraction Coupled co-perturbation axis (0.5 to 1.0 fraction)
Unanchored Numeric Claims Permitted & Correlated Blind-Spot Escape Probability 0 (Blocked by check_claims.py) · < 0.004 (3 Lineages + Verifier) Baseline biophysical operating point
Primary Literature Anchor DOI 10.1056/NEJMoa1908639 Middleton et al. (2019), New England Journal of Medicine 381(19):1809-1819 — Elexacaftor–Tezacaftor–Ivacaftor for Cystic Fibrosis with a Single Phe508del Allele (PMID 31697873). · Kernel src/scripts/site.js

Act II · From Intuition to Governing Law

Independence of Failure Modes and Reproducible Verification

Suppose a single frontier model has a probability ε_1 ≈ 0.08 of overlooking a subtle denominator mismatch or an over-promising adjective. Asking the same model lineage to review its own output reduces the error rate only slightly because its weights share the same training blind spots (conditional failure probability P(F_2 | F_1) ≫ ε_2). By routing every monograph through independent model lineages with disjoint pretraining lineages—and then intersecting their approval with a reproducible formal verifier V_det whose false-negative rate on syntactic/numeric invariants (SHA-256 quote locks, allowed number sets, forbidden promise patterns, Scite retraction notices = 0) is identically zero—the combined escape probability P_escape = P(F_1 ∩ F_2 ∩ F_3) · (1 − Cov(V_det)) approaches zero.

The formal verification framework rests upon an explicit epistemic predicate defined over the complete set of public assertions. For every sentence c within Claims_public, our pipeline requires the existence of a corresponding record r in Catalog_allowed such that the SHA-256 cryptographic digest of the primary record matches lock(c), the citation retraction count equals zero, and the epistemic tier belongs to the admissible set {T0, T1, T2, T3}. In single-model generation, let ε_i denote the probability that model lineage i fails to detect a hallucinated numeric parameter or ungrounded medical extrapolation. If two reviewers share identical training corpora and architecture, their failure events are strongly correlated, meaning that P(F_2 | F_1) remains substantially higher than ε_2. By convening an adversarial council composed of structurally disjoint model lineages with orthogonal parameterizations, the joint failure probability collapses toward the product of marginal errors, P(F_1 ∩ F_2 ∩ F_3) ≈ ∏ ε_i. This structural independence prevents single-model idiosyncrasies from masquerading as consensus.

However, statistical review alone cannot provide zero-error assurances across critical clinical endpoints. We therefore couple the multi-model council to a rigorous verification automaton, implemented within src/scripts/site.js and audited by Python checkers. The formal verifier operates as a finite-state machine that parses syntactic claim structures, verifies numerical bounds against pinned JSON evidence snapshots, and halts the build if any unregistered token appears. Because the formal verifier possesses an invariant false-negative rate of zero on syntactic and numeric constraints, the composite escape probability becomes P_escape = P(⋂ F_i) · (1 − Cov(V_det)). When gate coverage Cov(V_det) reaches unity, the escape rate for ungrounded numerical claims is identically zero, establishing a mathematically sound foundation for automated scientific synthesis.

Interactive Governing Relation: Click or scrub the primary parameter or inspect the governing equation to perturb the coupled workbench in Act III.

Four-Stage Mechanistic Progression

  1. Stage 1 (Primary Evidence Catalog — sources/catalog.yaml): Before any script fetches a paper, trial, or structure, it checks sources/catalog.yaml to verify that the host, API, license, and robots.txt allow automated access.
  2. Stage 2 (Four Epistemic Answer Tiers — T0 to T3): Every claim is classified as T0 (out-of-scope / clinical advice → refuse and refer to care team), T1 (direct verbatim/numeric fact from a pinned primary snapshot), T2 (biophysical/mathematical model calculation with explicit scope caveats), or T3 (open research hypothesis labeled as unproven).
  3. Stage 3 (Multi-Lineage Adversarial Review Loop): Drafts and visual plates are reviewed across distinct model lineages (Claude Opus 5.5, Gemini 4 Argon, GPT-OSS 120B, and local Ollama Qwen 3.8 / Gemma 4) so correlated single-model hallucinations cannot self-approve.
  4. Stage 4 (Reproducible Gatekeepers — The Final Arbiter): Models never have the final word; reproducible Python verifiers (check_claims.py, check_anti_hallucination.py, check-voice.py, seo_audit.py --strict, and test_site.py) decide whether a change passes.

Act III · Interactive Visual Experiment & Plain-Language Lens

Try It Yourself: Epistemic Gate & Multi-Lineage Council Simulator

Adjust the number of independent model lineages and reproducible verifier coverage to inspect how correlated hallucination risk and unanchored claim leakage collapse.

A language model can write a fluent paragraph about a clinical trial or a scientific paper that never existed. When readers and researchers study Cystic Fibrosis, 'sounding plausible' is not good enough. That is why every monograph in this lab passes through automated Python verifiers (`check_claims.py`, `check_folio_citations.py`) that check every DOI and PubMed ID against Crossref and NCBI PubMed, backed by independent model lineages that look for mistakes before anything is published.

Epistemic Gate Walkthrough · Mode: Pinned Evidence Storyboard · How to explore this model

  1. Step through the four visual chapters: Click the four mechanism chapters below (1 · Primary Evidence Catalog… through 4 · Reproducible Gatekeepers…) to watch the schematic transition across each regime.
  2. Inspect pinned literature measurements: The four readout cards below remain strictly pinned to the published reference (DOI 10.1056/NEJMoa1908639) rather than synthetic slider scaling.
  3. Launch the full pillar simulator: Open Open Evidence & AI Loop Verification Workbench on /evidence/ for multi-parameter continuous simulation.
Interactive Visual Experiment · L.06 · Epistemic Gate Walkthrough · Mode: Pinned Evidence Storyboard Three-Layer Epistemic Filter: Catalog SHA-256 → Crossref/PubMed Gate → Independent Lineages

1 · Primary Evidence Catalog…: Before any script fetches a paper, trial, or structure, it checks sources/catalog.yaml to verify that the host, API, license, and robots.txt allow automated access.

Pinned Primary Evidence Mode · Click Chapters 1–4 above to animate the L.06 mechanism.

Plate L.06-A · Three-Layer Epistemic Filter: Catalog SHA-256 → Crossref/PubMed Gate → Independent Lineages ∀ c ∈ Claims_public: ∃ r ∈ Catalog_allowed : SHA256(r) = lock(c) ∧ Retractions(r) = 0 ∧ Tier(c) ∈ {T0..T3}
Three-Layer Epistemic Filter: Catalog SHA-256 → Crossref/PubMed Gate → Independent Lineages All Pipeline Stages Verified 1 · Primary InputACTIVE2 · Gate CheckACTIVE3 · PartitionACTIVE4 · SynthesisACTIVE5 · VerifiedACTIVE Adversarial Falsification Pipeline & Correlated Blind-Spot Suppression ✦ Drag canvas to scrub
Plate L.06-B · Coupled Readout Partitioning Normalized Sensitivity (Stages 1–4)
Plate L.06-B · Coupled 4-Readout Sensitivity & Counterfactual Partition Synergy Drive: 50% · Baseline Reference Operating Baseline (100% Normalized Anchor) 0 (Blocked by che…Unanchored Numeric … < 0.004 (3 Lineag…Correlated Blind-Sp… 100% (catalog.yam…Catalog Host & Term… T0 · T1 · T2 · T3…Epistemic Answer Ti… Coupled Readout Response Across Stages 1–4 & Direct Canvas Scrubbing (Click any bar to inspect)
Unanchored Numeric Claims Permitted0 (Blocked by check_claims.py)Pinned to Publication
Correlated Blind-Spot Escape Probability< 0.004 (3 Lineages + Verifier)Pinned to Publication
Catalog Host & Terms Compliance100% (catalog.yaml)Pinned to Publication
Epistemic Answer Tier ClassificationT0 · T1 · T2 · T3 EnforcedPinned to Publication
Full Pillar Observatory · Pillar IV · Living Proofs, Non-Equilibrium Biophysics & Global Registries

Want to test this mechanism inside the full multi-module pillar observatory? Open the coupled interactive workbench on the Pillar Hub.

Open Evidence & AI Loop Verification Workbench on /evidence/ →

Act IV · Clinical & Therapeutic Horizon

Open Tools for Every Researcher's Agent

Every skill we use to run this lab—cf-evidence-loop, cf-research-context, lab-voice, ai-loop-council, and the 24 domain SKILL.md companions across Pillars I–IV—is published openly under MIT / CC BY 4.0 licenses. Whether you are a CF clinician checking a trial endpoint, a biophysicist modeling an ion channel, or a student exploring formal proofs, you can load these skills into any AI assistant and build on verified ground. Grounded in Lab Discovery Catalog modules: `science_model_benchmark`.

The translational scope of rigorous verification extends far beyond software hygiene; it directly impacts how biomedical literature translates into real-world clinical comprehension. In cystic fibrosis, where over two thousand unique variants in the cystic fibrosis transmembrane conductance regulator gene have been identified, generalizing outcomes across variant classes without explicit genotype boundaries creates catastrophic misunderstandings. Highly effective small-molecule modulators have altered outcomes for patients with at least one Phe508del allele or responsive gating mutations, but individuals carrying homozygous nonsense or splicing variants remain ineligible for these treatments. An ungrounded algorithmic summary that casually conflates modulator response across disparate genotypes creates unwarranted hope and dangerous clinical confusion for patient communities evaluating emerging experimental protocols.

Our computational architecture enforces an explicit boundary rule: every therapeutic statement must explicitly state its variant eligibility constraints and link directly to primary clinical trial registries or approved regulatory labeling. In our verification pipeline, automated claim checkers inspect genotype classifications, ensuring that Class I premature termination variants and ultra-rare missense alterations are never misrepresented. By releasing these verification tools as portable open-source skills, we enable external research teams and clinical specialists worldwide to audit their own computational workflows with equal rigor. Eliminating hallucinations from scientific writing does not promise an unverified endpoint; rather, it ensures that every step taken toward therapeutic discovery rests upon verifiable, reproducible facts that caregivers and scientists can trust unconditionally.

Frontier Discovery Dispatch · Pillar IV · Living Proofs, Non-Equilibrium Biophysics & Global Registries

Primary Anchor: Middleton et al. (2019), New England Journal of Medicine 381(19):1809-1819 — Elexacaftor–Tezacaftor–Ivacaftor for Cystic Fibrosis with a Single Phe508del Allele (PMID 31697873). (DOI 10.1056/NEJMoa1908639 · PubMed 31697873 · PDB 8EIQ). Governed by biophysical kernel src/scripts/site.js and cross-checked in the Open Evidence & AI Loop Verification Workbench on /evidence/.

Resumen Científico en Español (ES-MX) · Folio L.06

Verificadores Deterministas y Bucles Cero-Alucinación: Cómo check_claims.py y los Consejos Auditan la Ciencia: Sintetizando el Pilar IV: cómo los niveles de respuesta (T0–T3), las huellas SHA-256, Scite y los consejos multifamiliares previenen falsas promesas.

En un campo donde las familias y los equipos médicos leen cada palabra con atención, un sistema de IA que invente una cita o convierta un modelo de laboratorio en una promesa clínica causa daño real. Por ello, cada ensayo de esta colección pasa por verificadores deterministas en Python y consejos multifamiliares.

En Palabras Sencillas: Los algoritmos computacionales modernos generan con frecuencia textos biomédicos fluidos pero sumamente engañosos. Cuando los modelos de lenguaje analizan ensayos clínicos complejos, suelen inventar cifras convincentes que suenan autoritarias para los médicos y las familias afectadas. Para resolver este dilema, construimos un estricto árbitro de verificación por software. Nuestro verificador determinista examina de forma automática cada párrafo, extrayendo hallazgos numéricos, citas de autores y nombres de fármacos antes de compararlos con los registros originales del ensayo. Cualquier afirmación sin respaldo o porcentaje alterado detiene de inmediato el proceso de publicación. Al combinar este control mecánico con consejos de revisión formados por modelos independientes, eliminamos los puntos ciegos correlacionados que afectan a los sistemas aislados. La precisión matemática reemplaza la ficción plausible en toda nuestra biblioteca de investigación. Cada oración que supera esta rigurosa tubería brinda total certeza de que los resultados reflejan descubrimientos biológicos auténticos.

Analogía Cotidiana: Considere el protocolo de inspección de seguridad en una bóveda internacional de metales preciosos. Antes de que un lingote ingrese al depósito oficial, no se confía en la impresión visual de un único tasador. Primero, un espectrómetro láser mide la composición elemental para certificar la pureza metálica. Después, tres peritos independientes examinan el sello de contraste desde diferentes ángulos, garantizando que el cansancio visual de un inspector no apruebe una barra falsificada. Finalmente, un sensor acústico mide la onda de sonido que atraviesa el núcleo metálico para confirmar la densidad interna. Si alguna medición se desvía del estándar metalúrgico, la compuerta neumática se bloquea de manera automática. La plausibilidad jamás sustituye a la medición en este recinto físico. Sin embargo, esta analogía deja de ser exacta porque el lenguaje natural presenta matices semánticos fluidos, mientras que un lingote sólido posee una estructura cristalina fija que no altera su significado según el contexto.

Guía de Controles Interactivos: Antes de examinar las derivaciones matemáticas, el simulador interactivo le permite comprobar cómo interactúan las distintas defensas de verificación. El primer control deslizante ajusta los Linajes Independientes de Revisión en el Consejo, definiendo cuántos modelos distintos de inteligencia artificial evalúan cada borrador. Desplazar este control de uno hacia cuatro linajes ilustra cómo el consenso múltiple reduce con rapidez los sesgos individuales. El segundo control determina la Cobertura de la Compuerta Verificadora Determinista, que representa la fracción de afirmaciones evaluadas mediante código reproducible. Debajo de los controles, cuatro lecturas revelan en tiempo real el estado del flujo de publicación. El indicador de Afirmaciones Numéricas No Respaldadas Permitidas desciende a cero en cuanto la cobertura es completa. La Probabilidad de Fuga por Puntos Ciegos Correlacionados cae por debajo de cuatro milésimas al incorporar linajes diversos. El cumplimiento del catálogo confirma los permisos de acceso a datos y la clasificación por niveles epistémicos asigna cada enunciado a su categoría adecuada.

Riesgo Biofísico y Clínico: La trascendencia de la verificación epistémica en la fibrosis quística es absoluta. Cuando Middleton y sus colaboradores publicaron el ensayo clínico de fase tres de elexacaftor, tezacaftor e ivacaftor en el New England Journal of Medicine, documentaron un incremento de 14.3 puntos porcentuales en el volumen espiratorio forzado junto con una reducción de 41.8 milimoles por litro en el cloruro en sudor para personas con un solo alelo Phe508del. Esas métricas cuantitativas transformaron las expectativas terapéuticas globales. Si una síntesis automatizada cita erróneamente estas variables o confunde la eficacia de un solo alelo con respuestas homocigotas, los investigadores adoptan premisas equivocadas y las familias reciben falsas expectativas. Nuestra arquitectura matemática garantiza que cada afirmación numérica corresponda a una firma criptográfica verificada en los catálogos primarios de evidencia. Al aplicar niveles epistémicos de respuesta, distinguimos formalmente entre datos clínicos empíricos, simulaciones biofísicas e hipótesis exploratorias, evitando que especulaciones computacionales se hagan pasar por medicina comprobada.

Derivación Formal: El marco de verificación formal se fundamenta en un predicado epistémico explícito definido sobre el conjunto total de afirmaciones públicas. Para cada oración c dentro de Claims_public, nuestra tubería exige la existencia de un registro r en Catalog_allowed tal que el resumen criptográfico SHA-256 coincida con lock(c), el conteo de retracciones sea cero y el nivel epistémico pertenezca al conjunto admisible {T0, T1, T2, T3}. En la generación con un solo modelo, definimos ε_i como la probabilidad de que el linaje i no detecte un parámetro numérico alucinado o una extrapolación indebida. Si dos revisores comparten arquitecturas y datos de entrenamiento idénticos, sus modos de falla están correlacionados, por lo que P(F_2 | F_1) supera ampliamente a ε_2. Al convocar un consejo contradictorio integrado por modelos estructuralmente independientes y ortogonales, la probabilidad conjunta de falla se reduce al producto de los errores marginales, P(F_1 ∩ F_2 ∩ F_3) ≈ ∏ ε_i, impidiendo que sesgos individuales se confundan con consenso científico.\n\nSin embargo, la revisión estadística no puede otorgar garantías de error cero en variables clínicas críticas. Por ello, acoplamos el consejo multifamiliar a un autómata determinista implementado en src/scripts/site.js y auditado por verificadores en Python. Este verificador actúa como una máquina de estados finitos que analiza la sintaxis, valida cotas numéricas frente a instantáneas JSON inmutables y frena la compilación si surge algún término no registrado. Como el verificador posee una tasa de falsos negativos idénticamente nula para invariantes numéricos y sintácticos, la probabilidad compuesta de escape resulta ser P_escape = P(⋂ F_i) · (1 − Cov(V_det)). Al alcanzar una cobertura unitaria Cov(V_det), la tasa de escape de afirmaciones cuantitativas no respaldadas es exactamente cero, cimentando la síntesis científica en bases rigurosas.

Frontera Traslacional: El alcance traslacional de la verificación determinista trasciende la calidad del software; repercute de manera directa en cómo la literatura biomédica se interpreta en el ámbito clínico real. En la fibrosis quística, donde se han identificado más de dos mil variantes en el gen regulador de la conductancia transmembrana, generalizar resultados sin definir fronteras genotípicas precisas provoca confusiones de enorme gravedad. Los moduladores de moléculas pequeñas han transformado el pronóstico de pacientes con al menos un alelo Phe508del o mutaciones de compuerta responsivas, pero quienes portan mutaciones sin sentido o variantes de procesamiento siguen sin poder beneficiarse de estas terapias. Un resumen algorítmico que mezcle respuestas terapéuticas entre genotipos dispares genera esperanzas infundadas y confusión clínica en comunidades que evalúan protocolos experimentales.\n\nNuestra arquitectura computacional impone una regla estricta: toda afirmación terapéutica debe explicitar los criterios de elegibilidad genética y enlazar con registros de ensayos clínicos o prospectos regulatorios oficiales. En nuestro sistema, los verificadores automáticos auditan las clasificaciones genotípicas, garantizando que las variantes de terminación prematura de Clase I y las mutaciones poco comunes se reporten con total precisión. Al distribuir estas herramientas como paquetes abiertos y reproducibles, permitimos que equipos externos auditen sus propios procesos con el mismo estándar. Erradicar las alucinaciones en la redacción científica no equivale a prometer una cura; asegura que cada avance hacia nuevos tratamientos descanse sobre hechos verificables que médicos y familias puedan consultar con absoluta confianza.

Horizonte terapéutico: Cierra el ciclo de 24 monografías en los cuatro Pilares: mostrando cómo cualquier investigador puede descargar nuestros paquetes cf-skills y SKILL.md para ejecutar las mismas verificaciones en sus propios agentes. Investigación independiente, no asesoramiento médico.

Act V · Primary Sources, Scite Audit & Companion Tools

Primary Evidence & Verification Ledger

Each monograph names its primary literature anchor. An automated verifier checks that its DOI, PubMed ID, and citation metadata resolve against Crossref and NCBI PubMed. Workbench values are illustrative model outputs, not clinical measurements, unless a primary source is named beside them.

  1. Primary Literature Receipt: Middleton et al. (2019), New England Journal of Medicine 381(19):1809-1819 — Elexacaftor–Tezacaftor–Ivacaftor for Cystic Fibrosis with a Single Phe508del Allele (PMID 31697873). (DOI 10.1056/NEJMoa1908639 · PubMed 31697873 · PDB 8EIQ).
  2. Citation Context (Scite): Inspect how later literature cites 10.1056/NEJMoa1908639 on Scite.ai · Crossref + Scite verified (data/folio-citations.cache.json · fetched 2026-11-08 · 2,320 citing · 109 supporting · 1,820 mentioning · 5 contrasting).
  3. Open-Source Computational & Epistemic Engine: OpenAlex, Scite Smart Citations & cf-evidence-loop Bayesian Synthesis (DOI 10.1162/qss_a_00146 · Scite Report).
  4. Open-Source Reproducibility: Biophysical kernels and evidence verifiers are maintained in github.com/senseiewok/cf-research and github.com/senseiewok/cf-skills.
For builders: method, companion SKILL.md, Python verifier, and edition history (REV 1.1)
EDITIONREV 1.1
PUBLISHED2026-11-03
REDATED2026-11-04
VOLUMEVOL. 03
EPISTEMIC SCOPET0/T1/T2 · Sourced Epistemic Governance & Verification Model
PORTABLE AI-AGNOSTIC COMPANION SKILL & VERIFIER cf-reproducible-evidence-verification-SKILL.md · Kernel src/scripts/site.js

Load this companion skill into Claude Code, GitHub Copilot, Cursor, Antigravity, or local Ollama models (Qwen 3.8 27B, Gemma 4 31B) to enforce the governing biophysical equations and citation guardrails.

View raw SKILL.md source
---
name: cf-reproducible-evidence-verification
description: AI-agnostic epistemic verification skill for Answer Tiers (T0-T3), SHA-256 primary evidence locks, Scite retraction gates, and multi-lineage council falsification.
license: CC-BY-4.0
---

# L.06 — Reproducible Claim Checkers & Source-Grounded Council Loops

## Governing Biophysical / Mathematical Law

∀ c ∈ Claims_public: ∃ r ∈ Catalog_allowed : SHA256(r) = lock(c) ∧ Retractions(r) = 0 ∧ Tier(c) ∈ {T0..T3}

## Primary Evidence Receipt

- **Citation:** Middleton et al. (2019), New England Journal of Medicine 381(19):1809-1819 — Elexacaftor–Tezacaftor–Ivacaftor for Cystic Fibrosis with a Single Phe508del Allele (PMID 31697873).
- **DOI:** https://doi.org/10.1056/NEJMoa1908639
- **PubMed PMID:** https://pubmed.ncbi.nlm.nih.gov/31697873/
- **Structural Reference (RCSB PDB):** https://www.rcsb.org/structure/8EIQ
- **Answer Tier:** T0/T1/T2 · Sourced Epistemic Governance & Verification Model

## Verification Rules

- Classify every scientific or medical statement into Answer Tiers T0 (refuse clinical advice), T1 (pinned primary fact), T2 (sourced model with caveats), or T3 (open hypothesis).
- Enforce all five anti-hallucination rules (AH-1 through AH-5) via reproducible scripts before publication.
- Never allow a single model lineage to self-approve high-risk scientific prose without independent review and reproducible verification.

## Safety & Clinical Boundary

Research synthesis and biophysical modeling only — never medical advice, clinical guidance, or a treatment promise.

REV 1.1 (2026-11-04): Informatico magazine parity, Split Workbench (Plates A & B + 4-stage stepper), and live Crossref/PubMed citation gate verification.

SEAT 1 · EDITORIAL & IA ARCHITECTURE Gemini 4 Argon + Gemma 4 31B

Verified 5-Act magazine flow, untruncated headings, and bilingual EN / ES-MX structural parity.

SEAT 2 · ADVERSARIAL CLAIM CHALLENGER Qwen 3.8 27B + Ornith 1.5 35B

Audited clinical dignity boundaries, non-promise guardrails, and discovery catalog coverage.

SEAT 3 · CITATION & IDENTITY GATE check_folio_citations.py

Verified DOI 10.1056/NEJMoa1908639 against Crossref and NCBI PubMed metadata.

SEAT 4 · BIOPHYSICAL KERNEL src/scripts/site.js

Coupled governing relation ∀ c ∈ Claims_public: ∃ r ∈ Catalog_allowed : to interactive Plates A & B.

Sensei Ewok Research Lab · 4 Pillars · 40 Five-Act Monographs

Browse all 40 CF & Biophysics Monographs

40 monographs