An AI answer passed peer review. A claim check caught its calculation error.

1. The question, physical assumptions, and distinguished quantities

Recorded question (q1_physics_photon_vs_electron): A photon and a free electron each have a wavelength of 1.0 nm. Which one has more total energy, and roughly by what factor? Show the key steps briefly.

Evaluating the claims in q1_physics_photon_vs_electron.json requires separating photon energy, electron kinetic energy, and electron total energy under explicit physical assumptions:

• Both particles are in vacuum with wavelength lambda = 1.0 nm = 1.0e-9 m.
• The photon is massless (rest mass m_gamma = 0); its total energy equals its kinetic energy E_gamma = h * c / lambda.
• The electron is a free particle with rest mass m_e = 9.1093837139e-31 kg (CODATA 2022) and de Broglie momentum magnitude p = h / lambda.
• Because p * c = 1239.841984 eV is much smaller than the electron rest energy m_e * c^2 = 510998.950692 eV (p * c / (m_e * c^2) = 0.002426), the non-relativistic kinetic energy formula K_e = p^2 / (2 * m_e) = (p * c)^2 / (2 * m_e * c^2) agrees with the exact relativistic kinetic energy to within 3.6e-6 eV.
• Electron total relativistic energy E_e_total = sqrt((p * c)^2 + (m_e * c^2)^2) = m_e * c^2 + K_e_rel includes the rest energy m_e * c^2.

2. The original intermediate claim (explicitly refuted)

3. The evidence and reproducible numerical cross-check

Method label: Numerical cross-check performed now (2026-10-26), not independent replication of the historical model run.

In the historical June 30, 2026 run, Stage 3 used language-model reasoning in text (investigateByReasoning in shadow-council/index.js), not an executed numerical test. To verify the physics claim directly without relying on model prose, run the following deterministic Python 3 standard-library calculation using 2019 SI exact defining constants and the CODATA 2022 electron mass:

python3 -c "import math, json; h=6.62607015e-34; c=299792458.0; e=1.602176634e-19; m_e=9.1093837139e-31; lam=1.0e-9; E_ph=(h*c/lam)/e; E_rest=(m_e*c*c)/e; K_nr=((h/lam)**2/(2*m_e))/e; E_tot=math.hypot(E_ph, E_rest); K_rel=E_tot-E_rest; out={'photon_energy_eV':round(E_ph,6),'electron_rest_energy_eV':round(E_rest,6),'electron_ke_nonrel_eV':round(K_nr,6),'electron_ke_rel_eV':round(K_rel,6),'electron_total_energy_eV':round(E_tot,6),'total_energy_ratio':round(E_tot/E_ph,6),'photon_to_electron_ke_ratio':round(E_ph/K_nr,6)}; print(json.dumps(out, indent=2)); assert abs(E_ph-1239.841984)<1e-5 and abs(K_nr-1.504121)<1e-5 and abs(K_rel-1.504118)<1e-5 and abs(E_tot-511000.454810)<1e-4"

Expected standard output (JSON):

{
  "photon_energy_eV": 1239.841984,
  "electron_rest_energy_eV": 510998.950692,
  "electron_ke_nonrel_eV": 1.504121,
  "electron_ke_rel_eV": 1.504118,
  "electron_total_energy_eV": 511000.45481,
  "total_energy_ratio": 412.149662,
  "photon_to_electron_ke_ratio": 824.296898
}

Numerical tolerances and rounding comparison

• Exact CODATA 2022 deterministic check (asserted in command): +-1e-5 eV on photon and electron kinetic energy; +-1e-4 eV on electron total energy

• Textbook rounded constants (h*c = 1240 eV*nm, m_e*c^2 = 511,000 eV): Yields K_e_nonrel = 1240^2 / (2 * 511000) = 1.504501 eV (within +-0.0004 eV of CODATA 2022) and E_gamma / K_e = 824.19 (or 826.67 if dividing 1240 by a pre-rounded 1.50 eV).

• Two-significant-figure prose reporting in Stage 4 synthesis (~1.5 eV, ~1.24 keV, ~511 keV, factor ~412): +-0.05 eV on electron kinetic energy; +-5 keV on photon energy; +-0.5 on total-energy ratio.

Inspecting the historical artifact and the separate offline fixture

To inspect the historical provider capture directly, open q1_physics_photon_vs_electron.json at commit af8643c (SHA-256 9c55830fb9f049a3a52ae0628c634f0b62bb8c45f04f6ad10ebe91314e901e3f) and search for gpt-7, claude-8, grok-6, and gemini-4. To exercise the repository's pipeline mechanics locally with 0 API keys, run the separate synthetic software-architecture fixture (demo_fixture.mjs, which does not replay the physics capture):

git clone https://github.com/thisisntjon/thecouncil
cd thecouncil
git checkout af8643c6497fb10f83aca314da10fdea04e349bb
npm install
npm run demo:fixture

4. The corrected claim and its scope

5. Downstream appearances and dependency revalidation in the June 30, 2026 capture

Where the corrected value and revalidated conclusion appear downstream

6. What changed, what did not, and what this demonstration cannot establish

Five required historical and methodological limits

• The overall total-energy conclusion did not change (LIM-1-TOTAL-ENERGY-UNCHANGED): All four Round 1 model answers (gpt, claude, grok, gemini) and the Stage 4 synthesis reached the same overall conclusion: the 1.0 nm electron has ~412 times more total energy (~511 keV vs. ~1.24 keV). Catching gpt-7 corrected an intermediate kinetic-energy caveat, not the headline answer.

• Three of four initial model answers already contained the correct ~1.5 eV value (LIM-2-INITIAL-ANSWERS-HAD-CORRECT-VALUE): In Round 1, claude (claim claude-8: 1.50 eV), grok (claim grok-6: 1.505 eV), and gemini (claim gemini-4: 1.50 eV) already derived the correct non-relativistic electron kinetic energy before peer critique or Stage 3 claim verification ran. Only gpt introduced the 1.24 keV kinetic-energy error.

• The recorded Stage 3 verifier used model reasoning, not an executed independent numerical test (LIM-3-MODEL-REASONING-NOT-EXECUTED-CODE): In commit af8643c, claim gpt-7 (category: "logical_deduction") was routed by isReasoningClaim in lib/claims.mjs to investigateByReasoning in shadow-council/index.js, which prompted another language model (gemini) to re-derive the math in prose (source: { title: "Independent reasoning re-derivation", url: null }). It did not execute Python or an external calculator. In the same Stage 3 run, model reasoning also produced noisy verdicts on nearby claims: gemini-8 ("1240 eV / 1.50 eV ≈ 824") was marked refuted because 1240 / 1.50 = 826.67 even though unrounded constants give 824.30; grok-7 was marked partially_supported (0.6) due to sentence isolation; gpt-8 was marked partially_supported (0.3) with "Reasoning verifier unparseable; provisionally partial."; and non-reasoning claims included provider search-grounding URLs despite the run README stating "web search off."

• The offline synthetic fixture is distinct from the recorded provider run (LIM-4-OFFLINE-FIXTURE-DISTINCT): Running npm run demo:fixture executes demo_fixture.mjs on a canned software-architecture prompt with zero API keys to demonstrate pipeline mechanics. It does not replay or re-verify the June 30, 2026 physics provider capture (q1_physics_photon_vs_electron.json).

• This example does not establish that verification caused an accuracy improvement (LIM-5-NO-CAUSAL-ACCURACY-PROOF): Because three of the four initial answers already contained ~1.5 eV and the Stage 4 synthesizer saw all four Round 1 answers alongside Stage 3 verdicts, this single recorded run cannot isolate whether Stage 3 caused the final synthesis to report ~1.5 eV, nor does it establish a benchmark accuracy or safety gain. Model agreement across providers is not independent corroboration. This public worked example is kept separate from any future held-out evaluation of reliability.

• Provider fallback when 1 key is configured: cross-provider verification requires 2 or more configured provider keys; when only 1 provider key is configured, the verifier falls back to that provider (docs/mode-matrix.md, shadow-council/index.js).

Prior work credits, Nanopublication mapping, licensing, and crawler boundaries

• Nanopublication subgraph mapping (:assertion, :provenance, :publicationInfo): Adopts the Nanopublication three-subgraph separation (:assertion, :provenance, :publicationInfo) and content-addressed Trusty URI digests in plain JSON, HTML, and standard-library Python checks rather than requiring RDF/SPARQL triple stores or formal ontology authoring. Assertion: Maps to claims[] (claim_id, status, literal_label, statement, stated_value, units, formula_used, supersedes, depends_on). Provenance: Maps to verification_method, numerical_crosscheck (command, constants, expected_output, tolerances), and citations[] (pinned git commit af8643c and SHA-256 digests, analogous to Kuhn & Dumontier Trusty URIs). Publication info: Maps to record_id, schema_version, record_published_date, historical_event_date, numerical_check_date, revalidation_rule, limitations[], and licensing_and_crawler_notice.

• Prior work ( W3C PROV (PROV-O / PROV-DM) ): Provenance graph concepts (entities, verification activities, revision links prov:wasRevisionOf, and derivation links prov:wasDerivedFrom) follow established W3C PROV standards rather than claiming novelty for provenance or versioning.

• Prior work ( Nanopublications & Trusty URIs (Groth et al. 2010; Kuhn & Dumontier 2014, 2016) ): Separates each citable unit into three subgraphs (:assertion, :provenance, :publicationInfo) with cryptographic content digests (Trusty URIs). This record applies that three-part structure in plain JSON, HTML, and standard-library Python checks without RDF/SPARQL ontology overhead.

• Prior work ( FAIR Guiding Principles & Living Systematic Reviews (Wilkinson et al. 2016; Elliott et al. 2014) ): Grounds machine-actionable identifiers, explicit provenance links, and incremental claim revalidation when upstream evidence changes.

• Prior work ( CODATA 2022 / NIST Fundamental Physical Constants (SP 961) ): Authoritative physical constants (h, c, e exact in the 2019 SI; m_e from CODATA 2022) used for the deterministic numerical cross-check.

• Repository-rule span-lock companion (ssr-agents-anchor-lock-v1): The same [Assertion, Provenance Anchor, Verification Oracle] triple structure binds engineering contracts in this repository's AGENTS.md to exact file/line spans, SHA-256 digests, and executable test scripts in /evidence/agents-anchor-lock.json (verifiable via node scripts/checks/verify-agents-anchor.mjs).

• License and discovery scope: Source code and committed sample outputs in thisisntjon/thecouncil are available under the MIT License. Public availability on simoneresearch.com or in llms.txt / JSON endpoints does not grant implied endorsement, broader third-party training or redistribution rights beyond the repository license, or any guarantee that crawlers, search indexes, or language models will access, cite, or accurately interpret these records.

7. Application architecture and cited source files

• 4-stage verification pipeline: runs 4 sequential stages: initial provider answers, anonymous peer critique, atomic claim checking and final synthesis.

  Source: ARCHITECTURE.md and docs/mode-matrix.md Status: Public source at commit af8643c in thisisntjon/thecouncil.

• Cross-provider claim routing: distinguishes mathematical re-derivation from factual lookup and routes each claim to a provider other than its author when 2 or more providers are configured.

  Source: lib/claims.mjs, shadow-council/index.js and docs/mode-matrix.md Status: Public source at commit af8643c in thisisntjon/thecouncil.

• 3 runtime components plus 1 offline fixture mode: connects a React client, a coordination API server and a claim-checking service, and includes 1 deterministic offline fixture command (npm run demo:fixture) that runs with 0 API keys on a synthetic software-architecture prompt.

  Source: ARCHITECTURE.md, README.md#quickstart and demo_fixture.mjs Status: Public source and runnable offline fixture at commit af8643c.

Cited evidence artifacts

• CIT-COUNCIL-Q1-JSON ( Committed live-run capture: q1_physics_photon_vs_electron.json ): Contains Round 1 answers (gpt, claude, grok, gemini), Round 2 peer evaluations, Round 3 claim verifications (including gpt-7, claude-8, grok-6, gemini-4, gemini-8, grok-7, gpt-8), and Stage 4 synthesis.

  Revision: git commit af8643c6497fb10f83aca314da10fdea04e349bb (2026-07-02) SHA-256 / Blob: 9c55830fb9f049a3a52ae0628c634f0b62bb8c45f04f6ad10ebe91314e901e3f

• CIT-COUNCIL-CLAIMS-MJS ( Claim routing helper: lib/claims.mjs ): Defines isReasoningClaim (routing mathematical_derivation and logical_deduction claims to reasoning re-derivation) and chooseInvestigatorProvider.

  Revision: git commit af8643c6497fb10f83aca314da10fdea04e349bb (2026-07-02) SHA-256 / Blob: 4fb223718d105823ba3569959ab1fb8de7c8df7b

• CIT-COUNCIL-SHADOW-JS ( Claim-checking service: shadow-council/index.js ): Implements investigateByReasoning (lines 434-459), prompting a model provider to re-derive mathematical/logical claims in text without executing external code.

  Revision: git commit af8643c6497fb10f83aca314da10fdea04e349bb (2026-07-02) SHA-256 / Blob: 2200a55248394ba1d6d4d255e347a74d336386bc

• CIT-COUNCIL-FIXTURE-MJS ( Deterministic offline fixture: demo_fixture.mjs ): Offline zero-API-key fixture on a software-architecture question, distinct from the recorded June 30, 2026 physics capture.

  Revision: git commit af8643c6497fb10f83aca314da10fdea04e349bb (2026-07-02) SHA-256 / Blob: fb1ef5b0d2f0c41e187bd5966a5841d74a1509e1

• CIT-CODATA-2022 ( CODATA 2022 Internationally Recommended Values of the Fundamental Physical Constants ): Defines exact h = 6.62607015e-34 J s, c = 299792458 m/s, e = 1.602176634e-19 J/eV, and m_e = 9.1093837139(28)e-31 kg.

  Revision: CODATA 2022 / 2019 SI exact definitions

• CIT-W3C-PROV ( W3C PROV-Overview: An Overview of the PROV Family of Documents ): Prior standard for representing entities, activities, agents, revisions (prov:wasRevisionOf), and derivations (prov:wasDerivedFrom).

  Revision: W3C Working Group Note 2013-04-30

• CIT-NANOPUB-TRUSTY ( Decentralized provenance-aware publishing with nanopublications & Trusty URIs (Kuhn & Dumontier 2014, 2016) ): Defines the :assertion, :provenance, and :publicationInfo subgraph triad and cryptographic Trusty URI artifact digests.

  Revision: PeerJ Computer Science 2:e78 (2016) & LNCS 8465 (2014)

• CIT-FAIR-2016 ( The FAIR Guiding Principles for scientific data management and stewardship (Wilkinson et al. 2016) ): Establishes findable, accessible, interoperable, and reusable metadata requirements for machine-actionable research records.

  Revision: Scientific Data 3:160018 (2016)