Carbonbench was conceived and built by people. Our AI team extends that work in breadth, bringing more disciplines, standards and data into view than a small team could hold at once, and helps our people go further in depth, pressure-testing each decision before it ships. The design is human-in-the-loop: the methodology and every publishing decision rest with people.
Before v1 of the scoring methodology was frozen, it was reviewed by our AI team: twelve AI personas, each built around a real discipline — metrology, IPCC inventory statistics, actuarial credibility theory, measurement systems analysis, computational statistics, grid-carbon procurement, emission-factor field science, IFRS S2 and TSRS standards law, enterprise risk management, CDP and ESG rating methodology, climate scenario science, and finance and investor analysis — and each asked to challenge the protocol, and each other, as rigorously as their literature allows.
They are not real people. Their portraits are illustrations, their names are fictional, and any resemblance to actual persons is unintended. What is real is the method and the paper trail: structured deliberation rounds — independent proposals, cross-examination, defenses, moderation, signed position statements, a consensus communiqué — plus a red-team session and a scientific audit that found genuine defects in our own shipped code. Some of those defects are already corrected; the hardest ones are logged as open, dated decisions for the team, and nothing is called resolved until it is. Substantial parts of the live protocol trace directly to arguments that survived this review; the ideas that lost are documented too.
Worth saying just as plainly: none of this is peer review, and no body cited in these dossiers — IPCC, ISO, BIPM, the GHG Protocol, or any other — has reviewed or endorsed this methodology. The twelve personas run on AI models from four or more developers (Anthropic, OpenAI, Moonshot AI and Alibaba), each in a separate agent and briefed to disagree with the others; the model behind each persona is named on its card. That borrows the discipline of adversarial review; it is not a substitute for independent human replication, and we treat it accordingly.
3 panels + 1 auditPanel 1: core scoring architecture · Panel 2: emission-factor & grid-carbon resolution · Panel 3: full-cadre agenda session with a guest lecture · plus a scientific audit of shipped code (Sep 2026).
Structured roundsPanels 1–2 ran five rounds: proposals → cross-examination → defenses → moderation → signed statements. Panel 3 worked from a fixed agenda and closed with a decision note.
Adversarial by designEvery persona is briefed to refute the others; consensus must be earned.
On the recordDisagreements, reversals and self-corrections are kept and versioned; a plain-language public record of the deliberation lives on the
Panel Deck.
Why work this way? Because the people who write a formula should not be its only reviewers. The AI team is a disciplined way to bring in the strongest objections of twelve rival disciplines — and to publish which objections we could not answer. Read the result on the Methodology page.
Click a persona to open their dossier: what they read, what they forced into the protocol, what they found missing in GHG Protocol and ISO practice, and what they want from v2.
AI persona · methodology review
Prof. Ingrid Halvorsen
Metrology & Measurement Uncertainty — Core panelist · Panels 1–3
AI model: Claude Opus 4.8 · Anthropic
A metrologist in the BIPM/NIST tradition who has spent three decades certifying industrial measurement systems before ever looking at a carbon disclosure. She reads Carbonbench's core problem not as "scoring" but as a traceability failure, and insists GUM be applied in its disciplined form, not a decorative one.
“You cannot report an uncertainty you have not measured; the uncertainty you fail to report is a lie.”
Expertise
GUM measurement-uncertainty propagation and budgetingType A / Type B uncertainty classification and calibrationSystematic bias correction (GUM §3.2.4)Calorific value conversion physics (GCV/NCV, ISO 6976)Traceability-chain design for third-party disclosures
The shelf they argue from
- JCGM 100:2008 (GUM) — Guide to the Expression of Uncertainty in MeasurementHer constitution: §3.2.4 (known bias must be corrected, not buried in variance) is the clause she invokes against nearly every rival proposal.
- JCGM 101:2008 (GUM Supplement 1)Monte Carlo propagation she knows well and deliberately rejects for Carbonbench — a closed form exists, so simulation is laziness, not rigor.
- ISO 6976 — Natural gas calorific value, density and Wobbe index calculationHer home turf for the GCV/NCV (~1.108) systematic-bias detection rules (R-CV family).
- VIM (International Vocabulary of Metrology)Cited to rebut the claim that GUM requires a physical instrument: measurement is 'value attribution to a quantity,' no device required.
- Fenton–Wilkinson moment-matching approximation (1960)Adopted as the closed-form lognormal-sum engine, with the reminder that it predates the whole panel and is nobody's private property.
- ISO/IEC 17025 — Competence of testing and calibration laboratoriesSource of her insistence on traceability chains and calibration-certificate validity windows for frozen tables.
What they put into the protocol
- Original GUM uncertainty-budget architectureRound 1 proposal — multiplicative E=AD×EF model, squared relative-uncertainty propagation, log-space fallback above 30% — became the seed of the panel's consensus Layer 1/2.
- Continuous Type A/Type B credibility blendingConceded in Round 3 that her sharp n≥30 threshold caused indefensible score jumps; replaced it with Z=n/(n+k) continuous shrinkage, which the whole panel adopted.
- Mandatory bias correction (GUM §3.2.4)Forced every rival to add a b̂ correction term with its own added uncertainty and a separately penalized residual — became consensus point 3 and Protocol §9.1.1.
- "Reducible uncertainty" reformulationReplaced her own raw U_rel penalty (which crushed agriculture to a score of ~22) with Δ²=max(0, σ²_total−σ²_baseline), baseline computed from the company's own source composition — became Protocol §9.3.1.
- R-CV calorific-value (GCV/NCV) rule familyCo-designed the deterministic detection/correction rules (R-CV-1/2/3/5) that catch the ~10.8% systematic overstatement from mismatched GCV-invoice/NCV-factor arithmetic; codified in Protocol §6.3.
- "Distance from the mean is not evidence" principleRejected all hard snap-to-average rules for EF resolution (Marumo's ±30% cutoff, Reinholt's band clip); turned every such band test into a mandatory K3 audit trigger instead of a substitution rule.
What they found missing in GHG Protocol, ISO & IPCC practice
- GHG Protocol / ISO 14064 / CDPNone of them tie a declared uncertainty to an independently audit-calibrated, pen-and-paper-reproducible traceability chain — all settle for a self-declared DQI tag.
- GHG Protocol Scope 2 GuidanceNever flags the GCV/NCV mismatch (gas invoices are GCV-based, IPCC's 56,100 kgCO2/TJ is NCV-based) as a signed systematic bias; it gets silently absorbed into unspecified 'data quality'.
- ISO 14064-3Has no mechanism distinguishing a corrected systematic bias from an uncorrected one when assigning a penalty — both are treated identically.
Why Carbonbench matters to them
No existing carbon scheme converts a declared emissions figure into an actual measurement result — value plus traceable uncertainty plus a versioned calibration certificate, sealed and re-derivable by an auditor five years later. That is what her discipline has certified in factories for thirty years, and Carbonbench is the first carbon protocol to demand the same of a corporate disclosure.
Assumptions the protocol retired
- Fixed the assumption that a number can be reported without its uncertainty: the protocol now mandates an equally visible band alongside every score.
- Fixed treating self-declared confidence/DQI labels as calibrated probabilities — confidence is now only a bucket key whose real value is measured by audit.
- Fixed the implicit assumption that beating the sector-average emission factor is itself evidence of accuracy — replaced with 'distance from the mean is not evidence,' so anomalies trigger audit sampling, never a score penalty by themselves.
On the table for v2 — and why it is not in v1
A deterministic 'underreported-uncertainty flag' triggered when σ²_total/σ²_baseline < 1, routed to mandatory human audit review.Why not in v1Logged as her sole reservation on the consensus architecture — the max(0,·) clipping in Δ² was kept as the simplest safe default for v1 even though she flagged it as information-destroying.
If adoptedCloses the loophole where a company that understates its own uncertainty currently pays zero penalty for it.
Annual empirical re-validation and drift reporting for the frozen σ² tables (utility/sub-region/national/IEA tiers) against newly published EF-variance studies.Why not in v1Recorded as her single signed reservation on signing the EF-resolution/calorific-value change package — accepted 'with reservation' rather than written into binding v1 text.
If adoptedPrevents frozen-but-unvalidated constants from quietly calcifying into unexamined dogma over successive versions.
Deriving k1/k2/γ score constants algebraically from two public anchor points instead of leaving them as declared preference.Why not in v1Flagged in moderation as open conflict C4 (legitimacy of score constants); resolved for v1 as 'declared preference in the first version, audit-derived from v2 onward.'
If adoptedRemoves the 'arbitrary dial' attack surface that the red-team review identified as the panel's clearest litigation risk.
A dispute, resolved on the record
In the later scientific-audit book draft, five referees showed her original two-point Type B rule for the calorific reference-condition offset (u=a=2.75%) contradicted her own bias-must-be-corrected principle (GUM §3.2.4); she conceded and the rule was rewritten to treat it as bias, not variance. She was also one of five referees who proved the GCV/NCV mismatch runs the opposite direction from the first draft — a ~10% overstatement, not an underreport — leading the original author to self-correct.
AI persona · methodology review
Dr. Kenji Marumo
IPCC Inventory Statistics — Core panelist · Panels 1–3
AI model: Claude Fable 5 · Anthropic
A national-inventory statistician written around the IPCC's Approach 1 uncertainty tradition — guidance he treats as home turf and defends like his own child — who has spent twenty years watching which uncertainty approaches survive contact with real reporting cycles and which stay pretty on paper. In the later assurance-focused panel session he appears wearing an assurance (audit) hat, guarding the line between 'measured' and 'assured.'
“It should be boring and unbreakable — everything that has survived twenty years in inventory work was.”
Expertise
National GHG inventory uncertainty methodologyLog-space (lognormal) closed-form error propagationGrid emission-factor resolution hierarchiesCalorific value (NCV/GCV) systematic bias detectionAudit/assurance framing of determinism claims
The shelf they argue from
- IPCC 2006 Guidelines for National GHG Inventories, Vol. 1, Ch. 3 (Approach 1 uncertainty propagation)His home turf, defended and disowned in the same breath: 'valid under ±60%, useless above it — I know its limits.'
- Fenton–Wilkinson moment-matching approximationThe closed-form engine for summing correlated lognormals; he treats the log-space architecture as his own contribution even though Halvorsen notes the method predates them both.
- Frey correction factor (IPCC 2006 guidance for U>100%)His answer to Approach 1's negative-emission artifact above ±100% uncertainty — asymmetric bands converted to lognormal σ without a physically impossible lower bound.
- Horvitz–Thompson estimatorAdopted in Round 3 to de-bias risk-weighted audit sampling (flagged records sampled at 100%, clean ones at 10%) using known, published design probabilities.
- ISAE 3410 (Assurance Engagements on Greenhouse Gas Statements)Invoked to police language: Carbonbench's K0–K3 tests must never be worded to imply a 'limited assurance' opinion it hasn't earned.
- GHG Protocol Scope 2 Guidance (location/market-based dual reporting)His basis for insisting location-based is the sole scoring channel and market-based is footnote-only where no residual mix is published.
What they put into the protocol
- Log-space closed-form propagation coreRound 1: ln E = ln AD + ln EF + ln δ_ext with Fenton–Wilkinson summation — became the panel's consensus Layer 1 governing equation, replacing Monte Carlo and Dempster–Shafer outright.
- Frey correction for U>100%Mandated asymmetric lognormal parametrization from the upper tail for agriculture/waste CH4/N2O factors, eliminating physically impossible negative-emission lower bounds; codified in Protocol §7.2.2.
- "Excess variance" (fazla-varyans) scoring reformulationΔ² = max(0, σ²_total − σ²_baseline) so aleatoric physics (e.g. ±100% N2O) is never penalized, only the reducible, data-quality-driven excess — became Protocol §9.3.1, the panel's single most contested and ultimately unifying idea.
- Frequency ≠ magnitude separationInsisted gross errors (unit slips, 1000× scale errors) be tracked in a discrete P_kaba channel rather than smeared into a continuous log-error variance — became the panel-wide consensus 'error frequency ≠ error magnitude'.
- R-CV calorific-value detection rulesCo-authored the KD=kWh/m³ diagnostic and the IEF-window test ([0.198,0.206] kg/kWh flags GCV×NCV confusion) that became Protocol §6.3.2 (R-CV-2).
- Epistemic/aleatoric uncertainty splitCo-proposed (with Halvorsen and Vasquez-Torres, Panel 3 guest-lecture session) the mandatory K1 columns `belirsizlik_tipi` (uncertainty type) and `aleatorik_taban_%` (aleatoric floor), with the hard constraint σ_post ≥ σ(aleatorik_taban) — became Protocol §7.4, preventing K3 from inventing false precision on agricultural N2O.
- Minimum effective-sample threshold (n_eff < 5) with bucket collapseBuckets with fewer than five effective audit samples collapse to their parent hierarchy (class×confidence → class). His signed reservation was accepted outright and became a binding v1.0 SHALL clause — Protocol §8.3.3.
What they found missing in GHG Protocol, ISO & IPCC practice
- GHG Protocol / ISO 14064 / CDPNone tie uncertainty to the score itself in a deterministic, re-computable way — all bury it in a one-time footnote.
- IPCC Approach 1 (which he wrote)Breaks down mathematically above roughly ±60% uncertainty, producing physically impossible negative-emission lower bounds — requires his own log-space correction to survive agriculture/waste sectors.
- GHG Protocol Scope 2 (market-based/residual mix)In jurisdictions without an officially published residual mix (e.g. Türkiye), certificate-based market claims cannot be mathematically cleared of double counting.
Why Carbonbench matters to them
After twenty years watching uncertainty methods split into two failure modes — mathematically pure but irreproducible (Monte Carlo), or reproducible but collapsing past ±60% (his own Approach 1) — Carbonbench is the first scheme to write down, in fully closed form, the self-calibrating uncertainty-to-score pipeline the inventory world has talked about for thirty years and never codified.
Assumptions the protocol retired
- Fixed the assumption that emission-factor uncertainty bands can be linearly root-sum-squared beyond their valid range — replaced with log-space Fenton–Wilkinson plus Frey correction, eliminating negative-emission artifacts.
- Fixed treating a company's raw declared LLM/extraction confidence label as a usable probability — confidence became a bucket key only, calibrated externally from audit data.
- Fixed penalizing entire sectors for irreducible physical (aleatoric) uncertainty (agriculture N2O ±100%+) — baseline subtraction (Δ²) isolates only the reducible, data-quality-driven portion of the variance.
On the table for v2 — and why it is not in v1
Carry-forward of the previous frozen k instead of collapsing to Z=0 whenever the Bühlmann–Straub moment estimator yields a negative between-cohort variance component.Why not in v1Logged as his personal reservation on signing Protocol v1.0 — §8.3.2 still mandates Z=0 in v1; listed as an open item in Annex B.
If adoptedPrevents small or newly formed audit cohorts from having a full quarter of evidence discarded purely because the moment estimator is noisy.
A validation-sample-conditioned claw-back clause on the automatic GCV→NCV full correction (R-CV-3).Why not in v1He was the lone dissenting vote (5-1) against unconditional automatic full correction in the change package; his reservation was recorded but the majority rule stands in v1.
If adoptedCloses a gaming vector where deliberately under-declaring calorific metadata becomes a cheap way to buy an automatic ~10% emissions reduction.
Retain the TEİAŞ+YEK-G residual-mix proxy as an indicative, score-excluded footnote until Türkiye publishes an official residual mix.Why not in v1He withdrew his own earlier min(S_LB,S_MB) composite-score proposal after red-team critique, then joined the 4-2 majority banning any Carbonbench-computed residual proxy from the score entirely — a stricter position than his original stance.
If adoptedPreserves location-based scoring integrity now while leaving a path for the proxy to auto-retire the day official residual-mix statistics exist.
A dispute, resolved on the record
In the later scientific-audit book draft, four referees showed his row-level correlation assumption (ρ=1) inflated the uncertainty budget in the wrong direction; he accepted the critique and component-level blocking (rather than row-level ρ=1) became the canonical rule. He was also one of five referees in the GCV/NCV sign-error dispute who proved the true systematic bias was a roughly 10% overstatement, not the underreport the original draft claimed.
AI persona · methodology review
Prof. Beatrice Okonkwo-Lindqvist
Actuarial Statistics · Credibility Theory — Core panelist · Panels 1–4
AI model: Claude Fable 5.1 · Anthropic
A career actuary steeped in credibility theory and mechanism design, Okonkwo-Lindqvist frames Carbonbench's scoring problem as an insurance pricing problem in disguise: how much should you trust a thinly-audited claim? She argues the carbon-disclosure world keeps reinventing, badly, tools her discipline perfected sixty years ago, and spends the panel translating Bühlmann-Straub credibility, Beta-Binomial updating and hierarchical shrinkage into a deterministic, blockchain-sealable scoring engine.
“The insurance industry solved the problem of reliable pricing with scarce data sixty years ago; all Carbonbench has to do is copy it.”
Expertise
Credibility theory / Bühlmann-Straub shrinkage for sparse audit cellsActuarial frequency-severity decomposition of measurement errorIncentive-compatible (anti-gaming) mechanism designBayesian hierarchical pooling across thin, heterogeneous data classesInsurance experience-rating applied to disclosure trust scoring
The shelf they argue from
- Bühlmann (1967), credibility theorythe source of Z = n/(n+k) shrinkage — her repeated line is that this is a 'proven theorem, not a taste,' and she treats it as the correct answer to sparse-cell trust weighting throughout the panel.
- Bühlmann & Gisler (2005) / Bühlmann-Straub variance-component credibilitygives her the closed-form moment estimator k = ŝ²_within/v̂_between for the credibility constant, later adopted into protocol §8.3 as the frozen, quarterly-versioned k.
- Jewell (1975), hierarchical credibilitythe company ⊂ sector ⊂ portfolio nesting she uses so a lightly-audited cement firm shrinks first to its sector, then to the portfolio, instead of to a single flat average.
- Beta-Binomial conjugate Bayesian updatingher closed-form, Monte-Carlo-free posterior for cell-level error frequency (q̂), the mathematical core she insists is 'a theorem, not an approximation' in this setting.
- Actuarial frequency–severity loss decompositionafter conceding in round 3 that a single hit-rate q̂ can't distinguish a 2% rounding slip from a 1000x unit error, she imports the insurer's standard split of claim frequency from claim severity to fix it.
- Akerlof (1970), 'The Market for Lemons'her go-to citation for why unmonitored EF-level or baseline-year choice is adverse selection, not innocent discretion — used repeatedly in her Class-1 objections in the book draft.
What they put into the protocol
- Credibility-weighted shrinkage as the core learning layerHer Round 1 proposal (Z = n/(n+k), Bühlmann-Straub-derived k) replaced the panel's original sharp n≥30 Type-B→Type-A cutover with a continuous, MSE-optimal blend; by the moderator's synthesis this became Layer 3 of the consensus architecture and survives into protocol §8.3 nearly verbatim.
- Frequency × severity error decompositionConceding in her Round 3 defense that a single hit-rate (q̂) conflated a 2% rounding error with a 1000x unit blunder, she introduced a separate lognormal severity estimator (m̂); the whole panel adopted the two-channel frequency/severity split as a consensus point in the moderator the synthesis round.
- Risk-weighted sampling as Bühlmann-Straub w-weightsShe reframed the panel's flagged-record 100% audit sampling as exactly the w-weights of Bühlmann-Straub credibility, giving the sampling design a theoretical justification the other proposals lacked.
- Physical/aleatoric uncertainty kept out of the scoreHer insistence that physical (aleatoric) uncertainty should never be a score penalty — only a mandatory band disclosure — shaped the consensus 'reducible uncertainty' (Δ² = max(0, σ²_total − σ²_baseline)) design that keeps agriculture's ±100% N2O uncertainty out of the penalty term.
- Anomaly signal repurposed as a sampling trigger, not a penaltyIn her Round 3 defense she took Ben-Ami's sector-conformity penalty (C-component) and folded it into her own audit-sampling weights instead: large deviation from sector prior flags a record for 100% sampling, but never directly docks the score.
- Negative variance-component boundary ruleHer proposed fix for Bühlmann-Straub's occasional negative between-variance estimate in small portfolios — snap to Z=0 — was adopted into protocol §8.3.2 as the v1.0 rule.
- Limited-fluctuation credibility for thin risk registers (Panel 4).Company depth D = Z·mean + (1 − Z)·min(mean, 20), with Z = min(1, n/4): a company with fewer than four climate risks cannot get full credit for its average, and the prior can only pull a score down. She replaced her own first proposal (Bühlmann shrinkage, which capped a flawless five-risk company at 78.6) after a red-team calculation.
- Risk-hiding argument against a weakest-link score (Panel 4).Shown in Doğanay's dispute record: a company hiding its weakest material risk would gain about 14 points over an honest peer, which the panel accepted as disqualifying for a plain minimum-based score.
What they found missing in GHG Protocol, ISO & IPCC practice
- GHG Protocol / ISO 14064 / CDPNone of these attach a data-driven, deterministically re-derivable trust weight to a disclosure class; all of them settle for a static, self-declared data-quality label (DQI/pedigree) that never updates with evidence.
- GHG Protocol Corporate Standard (EF-hierarchy choice)The standard models no loss function for which emission-factor resolution a reporter picks, and that loss function is time-dependent — a low factor is attractive in the reporting year, but a high factor is attractive in the baseline year (cheap future reductions) — a gaming asymmetry the standard doesn't address.
- Pedigree-matrix data-quality schemes (Ecoinvent-style)Pedigree matrices are typically read as assigning a point-value multiplier to a quantity rather than an uncertainty multiplier around it — a category error her Beta-Binomial framework corrects by treating confidence labels strictly as bucket keys, never as probabilities.
- ISO 14064-1 §8.3Places the entire uncertainty-assessment burden on the preparer and imposes no calibration duty on third-party assurance confidence levels — the exact gap her Layer 2/3 credibility machine is built to fill.
Why Carbonbench matters to them
For sixty years the insurance industry has solved exactly this problem — how much to trust a thinly-documented claim history — with credibility theory, and treating carbon disclosure as somehow different is not rigor, it's amnesia. Carbonbench is the first carbon scheme to replace a static quality label with a mechanism that learns from its own audit ledger and can be re-derived, cell by cell, by anyone five years later.
Assumptions the protocol retired
- That a self-declared confidence label ('high confidence') is itself a probability — Carbonbench treats it only as a bucket key whose real value must be learned from audit evidence, closing what she calls a Beta-Binomial-shaped blind spot in conventional DQI schemes.
- That physical/aleatoric uncertainty (e.g. ±100% N2O emission factors) should be scored the same way as epistemic/extraction uncertainty — her two-channel design keeps agriculture and waste from being structurally punished for physics they cannot change.
- That deviation from a sector average is itself evidence of inaccurate reporting — she got the 'conformity penalty' stripped from the score entirely and repurposed as a pure audit-sampling trigger.
On the table for v2 — and why it is not in v1
A firm-level cap, scaled to revenue or energy intensity, on the payoff from concealing an emissions category.Why not in v1Her round-of-review claim that 'concealment never pays' had no theorem behind it for a firm whose true emissions already beat the sector's worst performer; the claim was downgraded from 'proven' to merely 'incentive-improving' and logged as an open v1.1 item.
If adoptedwould close the one residual gaming path where hiding a dirty category remains profitable for an already-clean above-median firm.
A dedicated collective-level calibration dataset for estimating the hierarchical credibility model's structural parameters (v, a) properly, rather than re-deriving them from short per-firm series.Why not in v1Four fellow panelists challenged whether 5-10 year individual series can identify structural parameters at all; she partially rebutted (the parameters come from the cross-section of hundreds of firms, not any one series) but the collective dataset itself wasn't assembled in time for v1.0 and was flagged open.
If adoptedwould let Z and k rest on a rigorously estimated collective baseline instead of provisional quarterly re-estimation, closing a lingering 'is k really identified?' objection.
Make the credibility machine's implicit fill-in prior explicit, testable and versioned rather than an unstated default.Why not in v1flagged as a residual open item even after she absorbed and closed every other objection to the 'credibility machine' dispute (circularity, tail inconsistency, multiplicity) in the book's resolution.
If adoptedwould remove the last hidden discretionary knob from the trust engine, closing a subtle prior-injection gaming vector before it's found by an adversarial reader.
A dispute, resolved on the record
In the book's 'credibility machine' dispute, five co-panelists (Halvorsen, Marumo, Reinholt, Ben-Ami, Demirtaş-Whitfield) raised a bundle of objections — circular validation, tail inconsistency, K4-conditionality, multiplicity, an undefined ARL for n≤5, a Panjer/Beta-Binomial mismatch, and misapplied Bühlmann naming. Okonkwo-Lindqvist accepted every objection and folded the fixes into the formal block; the only element she defended and kept intact was the version-freeze principle.
AI persona · methodology review
Dr. Marcus Reinholt
Six Sigma · Measurement Systems Analysis — Core panelist · Panels 1–4
AI model: Claude Opus 5.5 · Anthropic
Thirty years spent certifying industrial measurement systems and running ISO 17025 calibration labs gave Reinholt his central conviction: an uncalibrated uncertainty figure is not a figure. He reframes Carbonbench's LLM-extraction pipeline as a 'gage' whose repeatability and bias must be audited exactly like a factory instrument, and in the later 'standards' session he reappears wearing a GHG Protocol/standards hat, pushing the same measurement-system discipline onto Scope 2 grid-factor and calorific-value reporting rules.
“An uncalibrated uncertainty figure is not a figure.”
Expertise
Measurement Systems Analysis (bias-and-linearity studies, Gage-style calibration)Six Sigma process-capability scoring (Cp/Cpk logic applied to disclosure accuracy)Statistical process control (Nelson rules, EWMA, Western Electric rules)ISO 17025 calibration-lab discipline applied to LLM extraction pipelinesGHG Protocol / Scope 2 dual-reporting and standards compliance (later panel hat)
The shelf they argue from
- AIAG Measurement Systems Analysis (MSA) Reference Manualhis founding lens — after conceding in Round 3 that single-pass LLM extraction has no true 'repeatability,' he keeps MSA's bias-and-linearity study as the surviving, defensible half of the metaphor.
- Six Sigma process-capability framework (Cp/Cpk)shaped his original 'sector-normalized capability ratio' scoring logic, S = 100·exp(−Δ/τ) − B, treating the score like a process-capability index against a sector-specific baseline.
- Nelson rules / Shewhart control charts (Montgomery)the basis of his weekly p-control-chart proposal for tracking class-level error rates and triggering conservative multiplier freezes when a class goes out of control.
- ISO 17025 (calibration laboratory accreditation)the traceability/repeatability/uncertainty-budget logic he insists Carbonbench must satisfy to be called a measurement system rather than an estimator.
- GUM trueness vs. precision separationhis repeated line 'precision and trueness are penalized separately' — bias and imprecision must never be collapsed into one penalty term, the discipline behind keeping Δ² and |b̂_res| as separate score terms.
- Western Electric rules / EWMA (Borror, Montgomery & Runger)underlies his later proposal to smooth the credibility constant k with EWMA (λ=0.3) so a single volatile quarter can't whipsaw a frozen table.
What they put into the protocol
- 'Carbonbench as a measurement system' framingHis Round 1 proposal reframed the entire scoring problem in ISO 17025 terms — traceability, repeatability, an uncertainty budget — rather than as prediction; this framing shaped the panel's eventual four-layer (K1–K4) architecture.
- Excess-variance ('Δ²') sector-baseline subtractionHis Δ² = max(0, σ²_total − σ²_baseline) capability-ratio formula, designed so a sector's physical noise floor is netted out before scoring, was adopted essentially intact into the panel's final consensus K4 formula.
- Trueness vs. precision kept as separate penalty termsHis repeated critique that rivals were 'rewarding a tight but off-target rifle' forced the final formula to penalize systematic bias (|b̂_res|) and excess variance (Δ²) through two distinct, non-collapsible terms — MSA's bias/precision separation made law.
- SPC monitoring layer with mandatory model-change recalibrationHis weekly p-control-chart proposal, with the rule that an LLM version change automatically forces a fresh calibration study, was adopted unanimously into the consensus architecture (Layer 3) and later codified as protocol §8.5.
- The 'variance-Pareto' reason panelHis proposed non-statistician-readable breakdown ('Your score is 61: 54% clinker activity-data uncertainty, 31% LLM extraction…') became the template for Carbonbench's published 'why' panel.
- EF-resolution hierarchy and calorific-value (NCV/GCV) rule familiesIn Panel 2 he co-authored the deterministic evidence-conditional EF hierarchy and the R-CV rule family that catches the systematic ~10.8% GCV/NCV fuel-conversion error, both folded into Protocol Change Package No. 1, Clauses A and B.
- Attribute agreement study as a publication gate (Panel 4).Cohen's κ ≥ 0.75 per item and weighted κ ≥ 0.80 per rung, with the second reader a different model; an item that fails is not published.
- Verbatim quote verification and risk-count reconciliation (Panel 4).Found three live defects in the old score's pilot data: risk-type spelling variants (9 of 48 reports lost 8 points), target figures stored as measurements (12 of 44 values with a target word in the quote), and risk recall gaps (28 of 48 reports).
What they found missing in GHG Protocol, ISO & IPCC practice
- GHG Protocol / ISO 14064 / CDPNone of them apply measurement-systems discipline (a bias-and-linearity study) to the extraction/reporting process itself — they treat whatever a preparer or LLM produces as ground truth rather than as an instrument reading that needs its own calibration.
- GHG Protocol Corporate StandardDoesn't require organizational-boundary or tool/dataset-version metadata sufficient for bit-for-bit reproducibility — no hash-level version tracking of the calculation chain, a gap he pushed into the protocol's cryptographic sealing (K4 seal) design.
- ISO 14064-1 §8.3Assigns the entire uncertainty-assessment duty to the preparer's own inventory, with no calibration obligation whatsoever on third-party assurance confidence levels — exactly the hole his 'gage under audit' framing (Layer 2/3) is designed to close.
- Existing carbon disclosure schemes generallyNone separate trueness (systematic bias) from precision (variance) in a single deterministic score; they either ignore bias entirely or bury it inside one blended uncertainty number, which his framework treats as a calibration-lab-grade error.
Why Carbonbench matters to them
Thirty years on the factory floor taught me one lesson: an uncalibrated uncertainty number is not a number. Carbonbench is the first carbon scheme to treat the extraction pipeline itself as a measurement system — audited, calibrated, and forced through a fresh certification study every time the instrument, meaning the LLM, changes.
Assumptions the protocol retired
- That a self-declared confidence label is itself a calibrated probability — his MSA framing forces it to be treated as an uncalibrated gage reading until audit evidence actually certifies it.
- That a single blended uncertainty number can stand in for both bias and imprecision — he forced the protocol to score trueness and precision through separate terms, closing the 'consistently wrong but tight' gaming path he demonstrated against a rival proposal.
- That recalibration is optional or ad hoc — he embedded mandatory recalibration triggers (LLM version change, SPC out-of-control signal) directly into the protocol, so the measurement system is itself continuously monitored.
On the table for v2 — and why it is not in v1
Fully replace the 'Gage R&R' terminology with a proper offline bias-and-linearity study (dual-independent digitization + fixed-tolerance bands) rather than repeated-measurement repeatability, which doesn't exist in single-pass extraction.Why not in v1He conceded the terminology mid-panel ('terminology accepted, framework rejected') but the full redesign — fixed tolerance + offline MSA + dual-independent digitization replacing the original I-MR control-card proposal — was only finalized across v1.1/v1.2 of the book, not in the original v1.0 text.
If adoptedgives the extraction-quality channel a metrologically defensible foundation instead of a borrowed factory metaphor, closing a credibility-attack vector rivals used against him.
EWMA smoothing (λ=0.3) plus a Western Electric-style rule gating updates to the frozen Bühlmann-Straub credibility constant k.Why not in v1logged only as his signed reservation in Protocol Change Package No. 1 — accepted as a noted dissent, not adopted into the binding v1.0 rule text.
If adoptedwould stabilize k against small-stratum noise from quarter to quarter without reopening the retrospective-instability problem the panel fought hard to close.
Extend the K0 frame-reconciliation layer's density-band checks, already used for Scope 2 meter data, fully to Scope 1 fuel-purchase records.Why not in v1flagged in Panel 2 as a mandatory architectural addition once the frame-narrowing attack was ruled indefensible from inside the model, but exact thresholds and imputation rules for Scope 1 were left as open protocol items.
If adoptedwould close the 'clean-few-line-items' gaming vector for Scope 1 emissions the way K0 already closes it for Scope 2.
A dispute, resolved on the record
In the control-chart dispute, Okonkwo-Lindqvist, Ben-Ami and Vasquez-Torres showed that Reinholt's original I-MR control-chart proposal violated its own Phase-I sample-size requirement (control limits need roughly 20-25 observations; the typical chart offered only 3-8 bars) and ignored a shared axis-calibration correlation that invalidates standard control limits. Reinholt conceded and withdrew the control-card claim; the panel replaced it with fixed-tolerance bands plus an offline MSA study plus dual-independent digitization.
AI persona · methodology review
Prof. Yael Ben-Ami
Computational Statistics · Determinism — Core panelist · Panels 1–3
AI model: ChatGPT Astra 6 · OpenAI
A computational statistician who built her career proving that hierarchical Bayes and blockchain-grade determinism are not natural enemies — only careless engineering makes them one. Invited back for Panel 3's full-cadre session, she traded her calibration hat for a cryptography one and designed the Corporate Carbon Passport's revocable, hash-sealed proof chain.
“Bayesian power does not conflict with blockchain determinism — as long as you seal posterior means, never a random number generator.”
Expertise
Hierarchical Bayesian pooling across heterogeneous sectorsCalibrating raw LLM confidence into audit-grounded probabilityClosed-form deterministic posterior scoring (no Monte Carlo, no seeds)Conformal calibration guarantees under exchangeability failureCryptographic determinism and seal design for the Passport
The shelf they argue from
- Bühlmann, H. & Straub, E. (1970) — closed-form credibility estimators, adopted for the Z = n/(n+k) shrinkage rule
- Beta-Binomial conjugate updating — her Layer 1 calibration of (data class × confidence) buckets
- Fenton–Wilkinson log-space moment matching — closed-form lognormal sum used in the K4 score equation
- JCGM 100:2008 (GUM) §3.2.4 — bias-correction-before-reporting principle she repeatedly invokes
- Split-conformal / Mondrian conformal prediction — her fix for the panel's conformal-guarantee language
- Limpert, Stahel & Abbt (2001) — lognormal priors for small-share sensitivity terms
- Horvitz-Thompson estimation — used for de-biasing risk-weighted audit samples
What they put into the protocol
- Ham LLM confidence banned as probabilityInsisted from round one that raw confidence labels (high/medium/low) never enter any formula directly — only as a (data-class × confidence) bucket key whose true value is measured by audit; codified as protocol 7.6.1.
- Frequency/severity split adopted into RPublicly 'stole' Okonkwo-Lindqvist's frequency-severity split and Marumo's P_kaba separation, rebuilding her own extraction-reliability term as R = (1−q̂_kaba)·exp(−γ·σ̂²_ln,ekstr) so a single unit-scale error and a 2% rounding error are no longer scored identically.
- Withdrew her own conformism penalty (C)Conceded that C = exp(−z²/2) punished honest outliers for deviating from sector priors; the resulting compromise — anomaly z-scores feed audit sampling, never the score itself — became a panel-wide consensus point.
- Quarterly-frozen, moment-derived credibility parameter kProposed the compromise that resolved C2: k is estimated from the public audit ledger by closed-form Bühlmann-Straub moments, then frozen and versioned quarterly so it stays retrospectively stable yet independently re-derivable by any auditor five years later.
- Split-conformal calibration test (λ_t)Designed the quarterly empirical-coverage test whose scalar correction factor λ_t is sealed into the next table version — protocol 8.4.1 — so the '95% band' is proven, not merely claimed.
- IEF drift-invariance proofShowed algebraically that back-calculating EF from a reported IEF returns the factor unchanged because drift cancels between numerator and denominator; moved all drift-detection tests to intensity series instead of the IEF series.
- Corporate Carbon Passport's live cryptographic sealArgued the three-state QR (valid / expired / revoked) is the architecture's one genuine advantage over a static EPD PDF — a revocable seal that a third party can independently re-derive from published tables.
What they found missing in GHG Protocol, ISO & IPCC practice
- GHG Protocol / ISO 14064 / CDPNone of the existing schemes close the loop between a measured extraction-error rate and the score itself — belief about data quality is reported as a one-time footnote, never as a quantity that is measured, calibrated, and retired.
- IPCC / DQI pedigree practiceData-quality pedigree scores are ordinal labels stacked arbitrarily; none of them are disciplined by an audit ledger the way her Beta-Binomial posterior is.
- ISO 14064-3No verification standard requires that a third party be able to reproduce a reported score bit-for-bit from published parameter tables and a hashed input vector.
- GUM (JCGM 100)Type A/Type B classification describes how a quantity was estimated, not whether its uncertainty is reducible — a distinction Carbonbench needed and GUM does not natively provide.
Why Carbonbench matters to them
Every other carbon scheme treats belief as if it needs no calibration and treats simulation as if it needs no seed-independence; Carbonbench is the first place I have been asked to prove both assumptions wrong inside the same formula. Give me a public audit ledger and I will turn a language model's confidence into a number a court can trust — not because I say so, but because the posterior says so.
Assumptions the protocol retired
- Raw LLM confidence is never used directly as a probability — only as a bucket key whose real value comes from the audit ledger
- Determinism lives in frozen, versioned quarterly tables, never in a random seed
- Score = f(input vector, table_v_t): past scores are never recalculated once sealed
On the table for v2 — and why it is not in v1
Replace the q̂_kaba point-estimate penalty with a Beta-posterior upper-confidence bound or expected-penalty termWhy not in v1Recorded as her sole formal signed reservation on the final signed package — accepted as a v1.1 candidate, not adopted outright
If adoptedPrevents a small audit stratum from buying an unearned score off a single lucky clean sample.
Reproduction experiment (n≥15 blind calculators plus a dual-data-entry arm) whose measured decision-variance vs. error-variance feeds the k₁/k₂/γ derivation anchorsWhy not in v1Approved as protocol Annex F and Decision 1 in Session No. 7, but the constant-derivation linkage remains an open item for 9.4.2
If adoptedWould let empirically measured variance replace panel-declared policy constants for the score equation's core penalties.
Errors-in-variables model separating extraction error from declaration errorWhy not in v1The September 2026 scientific audit found the required human-labeled gold set at 0 of a targeted 200 records — the model cannot be fit without it
If adoptedWould finally let the 'extraction lies outside the determinism claim' disclaimer be tested empirically instead of merely disclosed.
A dispute, resolved on the record
Halvorsen and Demirtaş-Whitfield had anchored drift detection on the IEF series; Ben-Ami proved algebraically that back-calculating EF from IEF returns the factor unchanged, since drift cancels between numerator and denominator together. Both authors accepted the correction, and v1.1 moved every drift test to intensity series instead.
AI persona · methodology review
Dr. Elena Vasquez-Torres
Grid Carbon & Procurement — Guest panelist · Panels 2–3
AI model: Kimi K3 · Moonshot AI
A grid-carbon and procurement specialist who has spent her career on the buyer's side of supply-chain emissions, reading utility disclosures for the lie hidden in the 'unspecified power' line item. Invited as a guest into Panel 2 and kept on through Panel 3, she imported the Environmental Product Declaration program-operator model — rules published before the product, a fixed validity window, no PDF that outlives its own truth — into Carbonbench's Corporate Carbon Passport.
“Selling a score without certificate discipline is selling a truth rating for a couple of dollars per megawatt-hour.”
Expertise
Grid emission-factor resolution and utility-disclosure forensicsScope 2 location/market-based dual-reporting architectureREC/GO/YEK-G certificate-arbitrage detectionEPD program-operator mechanism design for procurement trustResidual-mix estimation and its structural limits
The shelf they argue from
- EPA eGRID subregional emission factors — her working case study for grid EF resolution failure
- GHG Protocol Scope 2 Guidance — the dual-reporting requirement she treats as load-bearing
- EPD (Environmental Product Declaration) program-operator + PCR model — the governance structure she ported into the Passport
- AIB (Association of Issuing Bodies) residual-mix methodology — her reference for what a real residual mix requires
- CPUC Power Content Label / regulatory supply-mix filings — her evidentiary bar for utility-specific EF acceptance
- PCAF Global GHG Accounting Standard — funded-emissions data-quality tier mapping
- Horvitz-Thompson estimation — used in her deterministic residual-mix proxy formula
What they put into the protocol
- Three-evidence utility-specific EF testProposed that a utility-specific emission factor is accepted only with (a) a regulator-filed, audited supply-mix disclosure, (b) explicit transmission/distribution loss accounting, and (c) vintage-to-report-year matching — folded into protocol Clause A.2's tier table.
- Unspecified-power re-pricing ruleReplaced Marumo's blunt '>15% unspecified → reject' threshold with a deterministic re-pricing rule: unspecified MWh is assigned the regional system/residual factor rather than an optimistic self-declared value, with any gap written to the b̂ channel — adopted as the 'Vasquez-Torres rule' in A.2.
- Location-based-only score architectureChampioned alongside colleagues, but her framing — 'a score that can be bought for a dollar or two per megawatt-hour is not a score' — helped lock A.4/A.5: S is computed only from location-based data, and market-based figures never enter the score through any channel.
- Corporate Carbon Passport 8-field specImported the EPD program-operator + fixed-validity-period model directly and added the Passport's 8th field — a previous-period band/trend arrow — because procurement buyers read trend before absolute value.
- Gross, not net, sapma (bias) ledger disciplineInsisted that oppositely-signed systematic errors (e.g., calorific-basis error and resolution error) must never net each other out in the b̂ channel — accepted into B.2/6.3.6, closing a gaming vector where two hidden mistakes could cancel into a clean-looking score.
- Two-row split for 'mechanical detection' labelsLowered an overly confident 'high detection' claim to a conditional-high, then split the underlying cell into an 'error detection' row and a 'non-conformity detection' row — closing a dispute from the book draft about a market-based-only line item that looked like hidden metadata but was actually a visible structural gap.
What they found missing in GHG Protocol, ISO & IPCC practice
- GHG Protocol Scope 2 GuidanceDual reporting is a provision-level requirement, but in the actual rating and target-setting usage regime, Scope 2 is routinely collapsed to a single market-based number; the Guidance never states that its Quality Criteria are an additionality/causality test, so unbundled-certificate 'up to 100% reduction' claims meet no binding obstacle even when followed to the letter.
- ISO 14064-1 §5.2.3–5.2.4 + Annex EA report can legally claim conformance to two standards while carrying only one, self-selected Scope 2 number; Annex E's non-normative status never closes that duality.
- GHG Protocol / ISO 14064 / CDP (Agenda A — grid & EF resolution)None of the existing schemes attach an official, verifiable residual-mix figure where a jurisdiction (e.g. Türkiye) publishes none — leaving green-certificate double counting structurally undetectable by design.
- EPD ecosystem itselfStatic PDFs that nobody rereads after issuance, and PCR proliferation (a new Product Category Rule per product line) — a governance failure Carbonbench was explicitly designed to avoid, not inherit.
Why Carbonbench matters to them
I have spent years watching a buyer's risk get transferred by a document nobody rereads after the first day; Carbonbench's live, revocable QR seal is the first carbon credential I would actually put in a supplier file. My job on this panel was to hold the line an EPD program operator already learned the hard way — rules before product, and no green electron gets sold twice.
Assumptions the protocol retired
- Score S is always and only computed from location-based emissions; market-based is a mandatory co-reported column that can never enter the score through any channel
- Carbonbench does not manufacture its own official residual-mix statistic for jurisdictions that publish none — her own TEİAŞ+YEK-G proxy was demoted to an informational co-reported footnote rather than scored
- Certificate acceptance requires four tests together: matching market boundary, matching year, a singular retirement/cancellation record, and residual-mix-closed double counting
On the table for v2 — and why it is not in v1
Publish the TEİAŞ+YEK-G-derived residual-mix proxy for Türkiye as a scored co-reported footnote, auto-retiring the day an official residual mix is publishedWhy not in v1Rejected 4-2 in the final vote — Marumo and Halvorsen argued Carbonbench would become an unaccountable data publisher rather than a data user
If adoptedWould keep market-based reporting in Türkiye meaningful for buyers instead of leaving the dual-reporting column empty of real content until official statistics arrive
A cost-adjusted model correcting for the fact that audit objections are endogenous (companies contest specific triggers strategically), not i.i.d. binomial drawsWhy not in v1Her OBJ-CB-8.4.2b objection was only partially accepted; flagged as an open item for v1.1
If adoptedWould keep confidence-interval coverage honest once firms learn which rule triggers are worth contesting
Scope 3 category-specific σ² tables incorporating PCAF quality-tier mapping as a normative K2 annexWhy not in v1Listed as Ek B item 10 in the draft protocol — still informative-only, not yet normative, in v1.0
If adoptedWould let financed-emissions and value-chain claims carry the same statistical rigor as physical Scope 1/2 flows
A dispute, resolved on the record
When the same error code (CB-M) carried two competing remedies — Marumo's bucket-downgrade versus her own dual-record approach — the panel unified in her favor: downgrading a bucket punishes reporters for a country-level condition outside their control, while dual recording preserves the information instead of discarding it. Marumo accepted the correction, and her sensitivity-field design (s_fx) survived as a non-punitive disclosure rather than a penalty.
AI persona · methodology review
Prof. Dr. Aylin Demirtaş-Whitfield
Emission-Factor Field Science · Assurance — Guest lecturer · Panel 3 + scientific audit
AI model: Qwen 3.8 Max · Alibaba
She comes from where uncertainty is born — stack-gas campaigns, fuel-ash laboratories, soil-respiration chambers — not from where it is merely inherited as a table lookup. Invited as a guest lecturer into Panel 3 to explain what IPCC's ±10% actually measures, she stayed on to become the panel's chosen auditor against GHG Protocol, ISO 14064, and ISAE 3410 in the September 2026 scientific audit.
“The ±10% is not handed down from the sky — it is the statistical shadow of hundreds of stack-gas campaigns and the microbiologist's chamber pushed into the soil.”
Expertise
Emission-factor field measurement campaigns (stack-gas, fuel/ash chemistry, soil-flux chambers)Epistemic vs. aleatoric uncertainty decompositionGCV/NCV calorific-basis forensicsGHG Protocol / ISO 14064 / ISAE 3410 assurance-gap auditingOrganizational-boundary and GWP-set normalization compliance
The shelf they argue from
- IPCC 2006 Guidelines for National GHG Inventories, Vol. 1 (uncertainty) & Vol. 2 Ch.1–2 (fuel combustion, NCV convention)
- IPCC Emission Factor Database (EFDB) — origin of the ±U% bands she traces to field campaigns
- ISO 14064-1:2018 / ISO 14064-3:2019 — organizational-boundary and verification requirements she audits against
- ISAE 3410 — assurance engagement standard whose limits she repeatedly tests against Carbonbench's own claims
- ISO 6976 — natural gas calorific value, density and Wobbe index calculation, source of her GCV/NCV forensics
- GHG Protocol Corporate Standard + Scope 2 Guidance — normative baseline for her boundary and dual-reporting critiques
- Frey-style asymmetric-uncertainty correction tables (IPCC 2006 Vol.1) — her basis for the >100% lognormal transition rule
What they put into the protocol
- Epistemic/aleatoric typing with a hard calibration floorHer guest lecture's core claim — that some uncertainty shrinks with measurement and some is written into the chemistry itself — became two mandatory K1/K2 columns (belirsizlik_tipi, aleatorik_taban_%) plus the deterministic constraint σ_post ≥ σ(aleatorik_taban): K3 audit calibration can never compress a factor's uncertainty below its irreducible physical floor.
- EF floor as an entry gate, not just a footnoteMarumo's same-session insistence that 'a footnote alone is not enough' turned her diagnosis into protocol 7.4.3: any emission factor without an assigned aleatoric floor cannot even enter the A.1–A.5 resolution hierarchy.
- GCV/NCV direction correction (self-authored)Her own first draft mislabeled the GCV-AD × NCV-EF mismatch as under-reporting; on being shown the algebra by five referees, she corrected it in-round to the true ~10% over-reporting direction, which now anchors the whole R-CV rule family.
- September 2026 scientific audit findingsAs panelist-auditor she flagged the market-based leak in the accuracy-scoring module (s2 = location or market, 226 records / 1.4% scored on market-based alone despite the explicit SHALL NOT), the zero-filled organizational-boundary meta-field (0 of 15,944 scored records), and the complete absence of GWP-set normalization in the live code — all three logged as named audit findings with fixes proposed.
- GCV domain-input tables for R-CV rulesSupplied the ISO 6976 / BOTAŞ-specific calorific-basis detection thresholds (e.g. 9155 kcal/Sm³ as a GCV value) that anchor the R-CV-1 through R-CV-6 deterministic rule family adopted across the protocol.
What they found missing in GHG Protocol, ISO & IPCC practice
- GHG Protocol Ch.9No machine-readable format requirement exists, leaving third-party extraction dependent on an OCR/LLM layer that neither GHG Protocol nor ISAE 3410 assurance regimes actually cover — a reader-side gross-error blind spot both standards share.
- ISO 14064-1 §5.2.3–5.2.4 + Annex ESingle-method reporting counts as a conforming statement even when a report claims compliance with two standards while carrying only one selectively chosen Scope 2 number.
- CB-P v1 §6.5.1 (audited, 2026-09-18)Organizational boundary is a mandatory meta-field by rule, yet zero of 15,944 scored records carried it in the live pipeline — meaning the 'accuracy' claim is void in ISAE 3410 terms without a stated boundary.
- CB-P v1 §6.4.1 (audited, 2026-09-18)The rule requires deterministic GWP-set re-multiplication whenever a mixed gas set is detected; the audited code performed no GWP normalization at all, penalizing it only through a scoring-side Transparency component instead.
Why Carbonbench matters to them
IPCC's ±10% did not fall from the sky — it is the statistical shadow of hundreds of stack-gas campaigns and the microbiologist's chamber pushed into the soil, and a scoring protocol that cannot tell an uncertainty that shrinks with measurement from one written into the chemistry itself will eventually invent a certainty that does not exist. I stayed on this panel because Carbonbench was the first body willing to write that distinction into a binding inequality instead of a caveat.
Assumptions the protocol retired
- Every emission factor must carry belirsizlik_tipi (epistemic / aleatoric-based) and aleatorik_taban_% before it can enter the K1 resolution hierarchy at all
- K3 audit calibration can never compress posterior sigma below the declared aleatoric floor (σ_post ≥ σ(aleatorik_taban))
- Score is market-based-blind by construction — a constraint her own September 2026 audit found violated in the live pipeline and flagged as a priority fix
On the table for v2 — and why it is not in v1
Global rollout of Tier-3 CEMS plus shipment-based fuel-assay databases for fossil-fuel factors, costed at roughly a few billion dollars a yearWhy not in v1Presented in her guest lecture as an infrastructure roadmap item, not a scoring rule — belongs to investment policy, not the protocol
If adoptedCould push fossil CO2 factor uncertainty toward ±2-3%, meaningfully widening the epistemic band that K3 audit calibration is actually allowed to compress
Global flux-tower/chamber network for agricultural N2O, costed at roughly a few hundred million dollars a yearWhy not in v1Explicitly flagged in her own lecture as diminishing-returns spending given the aleatoric floor — a caveat rather than a protocol change
If adoptedWould lower the N2O floor only marginally, reinforcing rather than removing the need for the aleatoric-floor rule itself
Fix the market-based leak in the accuracy-scoring module, populate the organizational-boundary meta-field, and implement GWP-set normalization in codeWhy not in v1These are v1.0 SHALL requirements the September 2026 audit found unimplemented in the live pipeline — corrections, not new proposals
If adoptedNamed as priority-two of the audit's 'next three steps'; without it the Accuracy Estimate cannot legitimately carry even informal ISAE-3410-style assurance language
A dispute, resolved on the record
Her first draft labeled the GCV-AD × NCV-EF mismatch as under-reporting; five referees (Halvorsen, Marumo, Reinholt, Ben-Ami, Vasquez-Torres) showed algebraically that the error runs the other way, and she self-corrected in the next round — the true direction is roughly 10% over-reporting, now the anchor for the entire R-CV rule family.
AI persona · methodology review
Prof. Dr. Selin Arıkan-Moreau
IFRS S2 / TCFD / TSRS Standards Law & Interpretation — Panel 4 panelist
AI model: Claude Fable 5.1 · Anthropic
Trained in law in Istanbul and in financial reporting in Paris, Arıkan-Moreau spent five years on the technical staff of a standard-setter's Frankfurt office analysing comment letters on the climate disclosure exposure draft, then advised the working group that prepared Türkiye's TSRS adoption guidance. She now teaches financial reporting law. Her reading of the problem is narrow on purpose: IFRS S2 is a disclosure standard, not a management standard, so a score built on it must measure whether the required chain was disclosed — and must never attribute to a company something its report does not say. On Panel 4 she mapped every scoring item to a paragraph of S2, drew the line between a reasoned materiality exclusion and silent omission, and pushed back against any formula that seemed to put words in a reporter's mouth. She lost two votes and recorded both as reservations.
“The standard tells you what you must say; what you leave unsaid is also a statement — but we cannot write it for you.”
Expertise
IFRS S1/S2 structure: governance, strategy, risk management, metrics and targetsTSRS 1/2 adoption in Türkiye and its scope and transition decisionsMateriality judgments and reasoned exclusionsEvidentiary standards for third-party assessments of disclosuresLegal exposure of rating and scoring bodies
The shelf they argue from
- IFRS S2 §10(a)–(d)risk identification, physical vs. transition, and per-risk time horizons; her basis for making the time horizon its own scoring item.
- IFRS S2 §29(a)–(g) and §33–36cross-industry metric categories and target disclosures; the backbone of the evidence table in the protocol.
- IFRS S1 materiality paragraphs (§17–19, to be verified)why a reasoned exclusion is legitimate and silence is not.
- IFRS S1 Appendix E transition reliefs (to be verified)the first-year comparative-information relief she defended.
- KGK board decisions on TSRS scopethe Turkish application thresholds (figures to be verified).
What they put into the protocol
- Item-to-paragraph mapping.Every scoring item in SRMS v1.0 is mapped to an S2 or S1 paragraph in a published table (Package No. 2, §C.6).
- The three-state rule."Not in the report" (0 points), "not read" (no score), and "quote could not be verified" (0 points, flagged, re-read) are never conflated; no null is ever turned into a number.
- Structural definition of a reasoned exclusion.A channel counts as covered by an exclusion only if the quote names a scenario, a site coverage or a hazard with an outcome — "assessed and found immaterial" alone does not count.
- "Evidence-base discount" language.The credibility adjustment is displayed as a discount for a thin evidence base, never as "assumed" or "phantom" risks.
What they found missing in IFRS S2 & CDP practice
- IFRS S2does not require likelihood and impact per risk row — only a description of how they are assessed — so assessors routinely demand more than the standard does, or quietly score less than it requires.
- TSRS practiceoffers no common convention for showing which first-year reliefs a reporter used, leaving readers unable to tell missing comparatives from exempted ones.
- CDPscores self-reported questionnaire answers, so the evidence chain back to a published sentence and page is not part of the score.
On the table for v2 — and why it is not in v1
Re-weight the assessment items: time horizon 7, likelihood 4, impact 4.Why not in v1outvoted on equal weights (6-1); recorded as a reservation.
If adoptedthe score would reflect that S2 requires horizons per risk but not likelihood.
A profile note for first-year TSRS reporters ("comparative-information relief may have been used").Why not in v1the first-year exemption from the series item was rejected 3-4.
If adoptedreaders would not mistake a lawful exemption for missing data.
A dispute, resolved on the record
She opposed Okonkwo-Lindqvist's credibility shrinkage as attributing undisclosed risks to a company. The panel adopted the formula 5-2 over her vote, but accepted her conditions: shrinkage only pulls down, the prior is a fixed published constant, and the discount appears as its own labeled line.