Methodology Decision Records (MDR)¶

Provenance registry: constants.json (canonical, machine-checked by scripts/check-constants.mjs). Every coefficient the v1.2 energy model ships is justified by at least one MDR below; the registry's methodologyDecisions.ids index mirrors this document and the gate verifies the references in both directions.

Posture: this methodology is aligned with the GHG Accounting Principles, the GHG Protocol Corporate Accounting and Reporting Standard (revised edition), the GHG Protocol Scope 2 Guidance (2015), ISO 14067:2018, ISO 14064-1:2018 and ISO 14044. "Aligned with" is deliberate — nothing here is certified, compliant or verified, and no third-party assurance exists. Estimates produced from these constants are illustrative estimates, framed as a climate contribution — an estimate never claims to compensate emissions on the user's behalf.

Edition note: the GHG Protocol Corporate Standard and Scope 2 Guidance are under revision (ISO partnership announced September 2025; Scope 2 consultation October 2025–January 2026). Only published editions are cited as the basis of a decision; draft documents are tracked but never relied upon.

Record format: Date, Owner, Standard (named edition with clause reference), Decision, Alternatives considered, Rationale, Review trigger.

MDR-13-01: Functional unit declaration¶

Date: 2026-11-11
Owner: Bloomy Earth engineering
Standard: ISO 14067:2018 §5.2 (functional unit) — aligned with
Decision: The functional unit is one AI inference interaction as characterized by the caller: a token-usage record (modelId, tokensIn, tokensOut, cachedTokensIn), or one generated image, or one legacy counter-based response. Declared unit = 1 request; reference flows are input/output tokens, grid electricity (Wh) and CO₂e (g).
Alternatives considered: A time-based unit (per minute of session) and a per-conversation unit were rejected — neither is observable at the MCP boundary, and both make the estimate depend on client behavior the estimator cannot characterize.
Rationale: ISO 14067 requires the functional unit to be defined before coefficients are admissible. The caller-characterized request is the only unit the server actually sees; declaring it explicitly keeps every coefficient's scope comparable. Results are not comparable across different functional units, scopes or coefficient versions, and public copy must say so.
Review trigger: Any change to what estimate_ai_footprint accepts as a request, or any Phase 17 public-comparability copy change.

MDR-13-02: Epistemic tier definitions¶

Date: 2026-11-11
Owner: Bloomy Earth engineering
Standard: ISO 14064-1:2018 uncertainty principles; GHG Protocol Corporate Standard (revised edition) transparency and accuracy principles — aligned with
Decision: Three strict, scope-matched tiers: MEASURED — a primary source reports this exact quantity at this exact scope (same hardware, phase and quantity); CALIBRATED — produced by a recorded transformation, regression or fit from measured data; HYPOTHESIS — a Bloomy assumption not derivable from a primary source. A measured input never makes a modelled output MEASURED.
Alternatives considered: A two-tier measured/assumed split was rejected because regression-derived values (e.g. fitted coefficients) are neither direct measurements nor free assumptions; an untyped pedigree score was rejected as unauditable.
Rationale: Tier semantics are the credibility contract: the gate enum-rejects anything outside the three tiers, and each tier's definition is decidable against a source reading. Recording epistemic status per entry lets readers weigh every number without trusting the model.
Review trigger: Any proposed tier definition change re-labels every entry and the public tier table — annual methodology review, or earlier if a source publication changes scope semantics.
Clarification 2026-11-11: A model-specific measurement is MEASURED when the entry publishes the model identity and applies no transformation, and a bucket that references it does not turn it into a bucket average. A bucket-level claim built by averaging, substituting a different model or applying a transformation remains CALIBRATED. C_IN_MASSIF_MOE stays MEASURED on this basis: it is the Llama-3.1-70B H100 prefill point at batch 64, published as a model-specific measurement and referenced by the bucket directly.
Correction 2026-11-12: The 2026-11-11 clarification is withdrawn. A quantity readable only off a figure axis with no printed unit is not a published quantity, so a model-specific point cannot be MEASURED on that basis. C_IN_MASSIF_MOE (now C_IN_MASSIF) is HYPOTHESIS from revision 3, and the general rule stands: MEASURED requires the source to publish the quantity at the entry's scope (CORR-10, CORR-11, MDR-13-17).

MDR-13-03: Recalculation and materiality policy¶

Date: 2026-11-11
Owner: Bloomy Earth engineering
Standard: GHG Protocol Corporate Standard (revised edition) base-year and recalculation concepts; ISO 14044 comparability requirements — aligned with
Decision: The registry is versioned (version + effectiveFrom + change log) and reviewed annually and on every new edition of a primary source. If a shipped coefficient moves ≥ 5 %, or any boundary/assumption/functional-unit decision changes, a new registry version is published with a dated methodology-change note and a Phase 14 re-pin. Below 5 %, the new edition is recorded in the change log but the pinned value stays until the next methodology release. A retraction or withdrawal of a cited source triggers immediate review regardless of magnitude.
Alternatives considered: Continuous re-pinning (rejected: public outputs would churn without notice) and fixed annual-only updates (rejected: source retractions and >5 % moves cannot wait).
Rationale: A materiality threshold plus a dated-version mechanism keeps published estimates stable while guaranteeing that significant evidence changes reach the product promptly and visibly. Every change records old value, new value, source edition, date and MDR id.
Review trigger: Annual review; publication of a new edition of any cited source; any source retraction.

MDR-13-04: System boundary definition¶

Date: 2026-11-11
Owner: Bloomy Earth engineering
Standard: GHG Protocol Scope 2 Guidance (2015) §4.3 (boundary/method concepts); ISO 14067:2018 §5.3 (system boundary) — aligned with
Decision: The model boundary is operational electricity for inference only: GPU + non-GPU server + datacenter overhead (PUE). Electricity is classified as Scope 2 location-based at the AI provider's boundary; the Bloomy estimate is a consumption proxy, not the user's corporate inventory.
Alternatives considered: A full cradle-to-grave product footprint (rejected: embodied-energy data for AI serving hardware is not openly licensed at usable scope) and a GPU-only boundary (rejected: it would systematically understate operational energy).
Rationale: The boundary names exactly what the coefficients measure, so no number is read as broader than it is. Excluded categories are declared in MDR-13-11 rather than silently missing.
Review trigger: Any proposed widening of included flows (embodied energy, water, network) is a new decision record and a new registry version.
Amendment 2026-11-11: The three-term composition (GPU + non-GPU server + datacenter overhead/PUE) is implemented by the shipped C_OUT_* coefficients: each records gpuRequiredCount × perGpuValue + serverEnergyPerToken with includesServerEnergy: true and includesPue: false. Coefficients are IT energy; PUE is applied exactly once by the consumer to the summed prefill + decode per-request IT energy, (T_in × C_in + T_out × C_out) × DATACENTER_UPLIFT. This keeps the locked boundary implemented rather than narrowed.

MDR-13-05: Allocation method¶

Date: 2026-11-11
Owner: Bloomy Earth engineering
Standard: GHG Protocol Corporate Standard (revised edition) completeness principle; ISO 14067:2018 allocation concepts — aligned with
Decision: Shared infrastructure is allocated as follows: batch sharing at BATCH_SIZE = 64 (inherited from the EcoLogits reference configuration, see MDR-13-12), provider-level PUE for datacenter overhead, multi-GPU scaling via the EcoLogits memory model, and no training amortization allocation to inference requests.
Alternatives considered: Per-request exclusive-GPU accounting (rejected: serverless serving does not dedicate GPUs per request) and revenue-based allocation of overhead (rejected: not decidable from the observable request boundary).
Rationale: Allocation choices are recorded, not implied: a reader can reproduce any coefficient's denominator. Completeness requires that the allocation basis be stated whenever shared resources contribute.
Review trigger: Published serving-stack evidence showing materially different batch or fleet sharing; change of the EcoLogits reference configuration.
Amendment 2026-11-11: Multi-GPU scaling is the EcoLogits memory model applied to total parameters (gpuRequiredCount = 2^ceil(log2(ceil(GPU_MEMORY_OVERHEAD × totalParams × QUANTIZATION_BITS / 8 ÷ GPU_MEMORY_GB)))), while the regression and latency terms are evaluated on active parameters. Both bases are recorded per bucket in derivation.anchorTotalParamsB and derivation.anchorActiveParamsB, so the allocation is reproducible.

MDR-13-06: Grid factor method — location-based¶

Date: 2026-11-11
Owner: Bloomy Earth engineering
Standard: GHG Protocol Scope 2 Guidance (2015) §5.1 (location-based method) — aligned with
Decision: One named grid factor ships: Ember world lifecycle carbon intensity, 2025 annual average = 458.29 g CO₂e/kWh (Ember Yearly Electricity Data 2026, republished by Our World in Data under CC BY 4.0). The edition and access date are pinned in the registry entry; revisions follow MDR-13-03.
Alternatives considered: A regional lookup table (deferred to v2 as GRID-01 — per-region data is available but multiplies the provenance surface before the model ships) and an undisclosed "average grid" number (rejected: unnamed factors are unauditable).
Rationale: The location-based method maps supply-mix onto consumption; a generation-based world average is an acknowledged proxy — the scope limitation (not consumption-based, imports/exports not reallocated, global rather than regional) is recorded in the registry assumptions block rather than hidden.
Review trigger: Publication of Ember's 2026 dataset (expected ~June 2027); any per-region capability decision.

MDR-13-07: Coefficient precedence framework¶

Date: 2026-11-11
Owner: Bloomy Earth engineering
Standard: ISO 14067:2018 §6 (data hierarchy) — aligned with
Decision: Per-bucket coefficient precedence: C_out takes the EcoLogits regression value (α = 1.1665273170451914e-6, β = −0.011205921025579175, γ = 4.052928146734005e-5, f_E(P_active, B) = α · e^(β·B) · P_active + γ) evaluated at the bucket's representative active parameters and BATCH_SIZE = 64, recorded as CALIBRATED, then cross-checked against ML.ENERGY/TokenPowerBench measurement points. C_in takes TokenPowerBench/ML.ENERGY measured phase-aware data where available; otherwise HYPOTHESIS (superseded 2026-11-12 — see Correction below). Model routing parameters come from the EcoLogits model catalog.
Alternatives considered: Single-paper coefficients per bucket (rejected: no one openly-licensed paper covers all buckets at serving scale) and a vendor-figure web scrape (rejected: proprietary and unauditable).
Rationale: The EcoLogits regression is fitted on ML.ENERGY measurements, so it is a recorded transformation of measured data — a CALIBRATED synthesis, not a weaker source — and it is recomputable from the pinned formula and inputs.
Review trigger: A new EcoLogits regression fit or a superseding measured benchmark changing the cross-check agreement.
Amendment 2026-11-11: The C_out chain is gpuRequiredCount × gpu_energy + server_energy at the bucket's anchor model, recomputable from the registered parameters array: ALPHA_REGRESSION, BETA_REGRESSION, GAMMA_REGRESSION, BATCH_SIZE, GPU_MEMORY_OVERHEAD, QUANTIZATION_BITS, GPU_MEMORY_GB, LATENCY_ALPHA, LATENCY_BETA, LATENCY_GAMMA, SERVER_POWER and SERVER_GPUS. Every registered parameter named in a coefficient's formula is declared in that entry's derivation.inputIds.
Correction 2026-11-12: The C_in clause is withdrawn as written. The premise that TokenPowerBench/ML.ENERGY publishes measured phase-aware per-input-token prefill data under an open licence is false: the prefill figure reports energy per stage on an axis with no printed unit, the per-token figures are not phase-separated per input token, and the record carries the arXiv non-exclusive distribution licence with an AAAI 2026 all-rights-reserved notice rather than an open licence (CORR-10). No cited source supplies a per-input-token prefill energy at the bucket anchors, so the C_in lane ships as a documented HYPOTHESIS under MDR-13-17 in the same revision-3 release (CORR-11). The C_out clause is unaffected.

MDR-13-08: Cache multiplier treatment¶

Date: 2026-11-11
Owner: Bloomy Earth engineering
Standard: ISO 14044 data-quality requirements (precision, completeness) — aligned with
Decision: The cached-prefill lane is evidence-first: a measured prefix-caching energy saving from an openly-licensed primary source becomes CALIBRATED; absent such a source the multiplier ships as HYPOTHESIS with rationale and an upgrade path. Unsourced cache-saving figures circulated in early drafts are dropped and must not reappear.
Alternatives considered: Adopting the input document's illustrative cache multiplier as fact (rejected: no primary-source measurement supports it) and removing cached-token accounting entirely (rejected: the field exists in the request shape and ignoring it would overstate energy).
Rationale: Degrading an unsourced number to a labeled hypothesis preserves model expressiveness while keeping the honesty contract: no number ships that looks measured when it is not.
Review trigger: Publication of a measured prefix-caching energy study, or first-party measurement with KV-cache enabled/disabled.

MDR-13-09: Data hierarchy and quality¶

Date: 2026-11-11
Owner: Bloomy Earth engineering
Standard: ISO 14067:2018 §6 (data hierarchy); ISO 14044 data-quality dimensions (temporal, geographic, technological representativeness; precision; completeness) — aligned with
Decision: Data priority: measured primary data first, values derived from measured data by a recorded transformation second, Bloomy hypotheses last. Secondary emission factors must be openly licensed, dated and edition-pinned; proprietary or non-commercial-licensed figures are inadmissible regardless of quality.
Alternatives considered: Quality-score pedigree aggregation (rejected: scores are invented precision where the sources either support a measurement or they do not) and "best available number" pragmatism (rejected: it silently admitted NC-licensed figures the license screen forbids).
Rationale: The hierarchy is what makes the tier labels meaningful at entry level, and the openness rule keeps every cited number reproducible after the repository split (Phase 16 license screen consumes the per-entry license field).
Review trigger: A required quantity only available under an inadmissible license, or a new primary source superseding an edition pin.

MDR-13-10: Uncertainty treatment¶

Date: 2026-11-11
Owner: Bloomy Earth engineering
Standard: ISO 14064-1:2018 §6.3 (uncertainty assessment) — aligned with
Decision: Uncertainty is recorded only when the source supplies it. Otherwise the entry carries an explicit unquantified marker with a reason and an upgrade path. The gate never accepts an invented confidence interval, and none is derived from a pedigree score the source did not publish.
Alternatives considered: Attaching symmetric ± percentages to every coefficient for appearances (rejected: fabricated precision) and omitting uncertainty language entirely (rejected: readers could misread HYPOTHESIS values as tight measurements).
Rationale: Source-reported-only uncertainty is the difference between an honest range and marketing. An unquantified marker is informative in itself: it tells the reader the source did not quantify, which is the truth for most current AI-energy measurements.
Review trigger: Any source beginning to publish uncertainty bands for a pinned quantity (upgrade path is a dated registry change per MDR-13-03).
Amendment 2026-11-11: The current release records every entry as unquantified: no cited source published a numeric interval at the entry's scope, and D-06 forbids inventing one. The required record shape is {kind, reason, upgradePath} with kind ∈ {unquantified, reported}; an entry upgraded to reported must also carry a numeric value, a non-empty unit and a non-empty sourceLocator.

MDR-13-11: Exclusions declared¶

Date: 2026-11-11
Owner: Bloomy Earth engineering
Standard: GHG Protocol Corporate Standard (revised edition) boundary-setting; ISO 14067:2018 completeness/disclosure concepts — aligned with
Decision: Excluded from the estimate and disclosed as excluded: embodied/grey energy, hardware manufacturing and end-of-life, water consumption, network and storage energy, training amortization, and idle capacity beyond the declared batch/PUE assumptions.
Alternatives considered: A nominal "full lifecycle" multiplier over operational energy (rejected: unsourced at the required scope) and silent omission (rejected: undisclosed exclusions make an estimate look more complete than it is).
Rationale: A declared-exclusion list converts "what's missing" into part of the credibility contract; the total estimate is systematically conservative in scope, which the public copy states alongside the functional-unit comparability limit.
Review trigger: Any inclusion of a currently-excluded flow is a boundary change requiring a new MDR and a new registry version.

MDR-13-12: Batch size assumption¶

Date: 2026-11-11
Owner: Bloomy Earth engineering
Standard: ISO 14044 technological representativeness; EcoLogits reference configuration (CC BY-SA 4.0 methodology) — aligned with
Decision: BATCH_SIZE = 64 ships as a stated assumption inherited from the EcoLogits reference configuration — not a measurement of production serving. The registry assumptions block records it with the note that actual batch sizes vary by provider.
Alternatives considered: Treating 64 as measured reality (rejected: production batches are unknown and variable) and re-fitting the regression at other batch points (deferred: measured batch sweeps show energy per token falling steeply before flattening, which is exactly why the assumption must be explicit rather than hidden in a coefficient).
Rationale: Making the inherited batch a named, reviewable parameter keeps the calibrated coefficients reproducible (f_E(P_active, 64)) and prevents a silent scope mismatch between the regression's denominator and any provider's real fleet behavior.
Review trigger: First-party serving telemetry, or a measured study bounding production batch distributions.
Correction 2026-11-12: The numeric spread is removed: the sweep figures come from a source whose licence is not open (CORR-10), and MDR-13-09 forbids embedding all-rights-reserved figures regardless of quality. The BATCH_SIZE = 64 decision and its review trigger are unchanged.

MDR-13-13: Market-based exclusion¶

Date: 2026-11-11
Owner: Bloomy Earth engineering
Standard: GHG Protocol Scope 2 Guidance (2015) — market-based method and contractual instruments scope; location-based-only decision recorded against it — aligned with
Decision: The model computes location-based electricity emissions only. Market-based reporting, renewable energy certificates, power purchase agreements and other contractual instruments are not modelled, and the exclusion is disclosed rather than left implicit.
Alternatives considered: Dual location/market reporting (rejected: a consumer-facing estimator has no access to provider instruments, so any market number would be fabricated) and silently presenting the location figure as "the" footprint (rejected: the method basis would be unstated).
Rationale: Scope 2 guidance permits both methods; choosing one and naming the other as excluded is the transparency principle in practice, and it prevents users from reading the estimate as their own supply-chain position.
Review trigger: Outcome of the Scope 2 market-based-method consultation tracked under the standard revision — method definitions may change but the exclusion posture stays decision-gated.

MDR-13-14: Datacenter uplift factor¶

Date: 2026-11-11
Owner: Bloomy Earth engineering
Standard: GHG Protocol Corporate Standard (revised edition) conservatism principle — aligned with
Decision: The 1.30 datacenter uplift ships as an explicit conservative HYPOTHESIS with an evidence gap recorded: the EcoLogits provider PUE table (Anthropic 1.09–1.14, OpenAI 1.20, Google 1.09, Mistral 1.16) tops out at 1.20, so no provider reaches 1.30. The value is not presented as sourced, and 1.20 is not silently adopted either.
Alternatives considered: Silently adopting 1.20 (rejected: it would match a documented source while being labelled a hypothesis — an inconsistency between label and basis) and dropping the uplift (rejected: cooling and overhead are real; excluding them would bias low, the opposite of the conservatism posture).
Rationale: Where the evidence gap is acknowledged, a labeled over-estimate with an upgrade path is more credible than a precision number with no basis. The registry entry cites the counter-evidence table as the locator.
Review trigger: Publication of provider PUE data reaching 1.30, or a decision to move to provider-specific PUE (which would be a new MDR superseding this one).

MDR-13-15: Registry versioning and change log policy¶

Date: 2026-11-11
Owner: Bloomy Earth engineering
Standard: GHG Protocol Corporate Standard (revised edition) consistency principles; ISO 14067 comparability conditions — aligned with
Decision: The registry carries version and effectiveFrom ISO dates plus the corrections array as its structured change log. Every coefficient change records old value, new value, source edition, date and MDR id; edition pins (e.g. the Ember 2026 dataset for GRID_FACTOR_EMBER_2025) live in the entry's locator and citation fields.
Alternatives considered: Git history as the change log (rejected: not renderable by Phase 17's public pages and not gate-checkable) and a free-text CHANGELOG (rejected: the corrections array is the same data in a machine-shaped form).
Rationale: Consistency over time is a comparability precondition: a dated version lets any historical estimate be attributed to the exact coefficient set that produced it, which the recalculation policy (MDR-13-03) triggers on.
Review trigger: First population of the corrections array (plans 13-02/13-03); any registry version bump requires this policy to be exercised end-to-end.
Amendment 2026-11-11: A same-day release is discriminated by the top-level revision integer: version and effectiveFrom stay 2026-11-11 for a same-day remediation of an unshipped registry, and revision 2 marks the CORR-07 denominator change. A version-date change remains the discriminator for any release on a later date.

MDR-13-16: Bucket representation and naming¶

Date: 2026-11-12
Owner: Bloomy Earth engineering
Standard: ISO 14067:2018 §6 (data hierarchy) and ISO 14044 technological representativeness — aligned with
Decision: The four routing buckets are keyed on total parameter count with fixed anchors (PETIT ≤10B on Llama-3.1-8B; MOYEN >10B–≤50B on Mixtral-8x7B; MASSIF >50B–≤200B on Llama-3.1-70B; GÉNÉRIQUE >200B or unclassifiable, extrapolated above Llama-3.1-405B). Mixture-of-experts composition is not a routing dimension: it is recorded per bucket in derivation.anchorModel and named in the label anchor clause (MOYEN is the MoE-anchored bucket; PETIT, MASSIF and GÉNÉRIQUE name dense anchors). The _MOE suffix is retired from the MASSIF id and label because that bucket is anchored on a dense model.
Alternatives considered: (a) re-anchoring MASSIF on an MoE — rejected at this revision because the four boundaries and anchors were accepted by the 2026-11-12 review and re-anchoring would change C_out; (b) keeping _MOE as a legacy routing label — rejected, a label asserting a composition the bucket does not implement is a false public claim; (c) adding MoE as a fifth routing dimension — deferred, no measured coefficient set supports it.
Rationale: A bucket name is a public claim; the defect found was representational, not arithmetic — the only MoE anchor in the set sits in MOYEN.
Review trigger: Any proposal to re-anchor a bucket, any routing change that makes MoE composition material, or a new bucket proposal.

MDR-13-17: C_in basis — prefill fraction hypothesis¶

Date: 2026-11-12
Owner: Bloomy Earth engineering
Standard: ISO 14067:2018 §6 (data hierarchy); ISO 14064-1:2018 transparency — aligned with
Decision: C_in ships as HYPOTHESIS in every bucket, anchored on the paired C_out at the same anchor model: C_IN_<bucket> = PREFILL_FRACTION × C_OUT_<bucket> with PREFILL_FRACTION = 0.25, the conservative upper bound of the stated hypothesis band [0.10, 0.25]. Every C_in is strictly below its C_out; the check is recorded in each entry's derivation.sanityBound and enforced by the gate group CONSTANTS-CIN-BOUND (plan 13-08). No batch-sweep interpolation is claimed: the measured sweep the previous revision read is total energy per token, not a per-input-token prefill quantity, and its source is not openly licensed (CORR-10).
Alternatives considered: (a) keeping the three interpolations — rejected 2026-11-12, not a reading the source supports, and it inverted the phase cost relationship by 3.6×–5.7×; (b) importing the demoted ~3.5× illustration — rejected, D-12 forbids importing it, and the resulting 4.0 ratio is a consequence of the phase argument rather than an input; (c) pricing input tokens at zero — rejected, the request shape carries tokensIn and ignoring input energy understates the estimate; (d) first-party prefill measurement — deferred as the upgrade path.
Rationale: Prefill is compute-bound at high utilisation while decode re-reads the active weights for every token and is memory-bandwidth-bound, so an input token cannot cost more than an output token; with no admissible per-input-token measurement the honest representation is a labelled hypothesis bounded by physics and enforced mechanically rather than asserted in prose.
Review trigger: Publication of a per-input-token prefill energy at any anchor model under an admissible licence, or first-party measurement.