CarbonLayer / evidence status

Evidence-Based Framework (Pending Final Coefficient Verification)

Methodology v1.0 is available as a framework for evidence boundaries and claim classification. Specific coefficients remain under verification. Read the Methodology v1.0 status or jump to the evidence-status glossary.

Framework status

Available framework; coefficients pending final verification.

Version
1.0
Coefficient verification
Pending
Carbon / water impact outputs
Unavailable

00

Evidence-status glossary

Use these labels to read the evidence behind a claim. Impact classifications remain measured, modeled, estimated, or unavailable; simulated describes prototype behavior and is separate from those impact classifications.

Measured

Directly observed or recorded evidence tied to a source and time. For CarbonLayer, an/api/inference receipt’s createdAt timestamp and request fields record API activity; they are not an environmental measurement.

Modeled

A value derived from assumptions, coefficients, or synthetic inputs rather than a direct observation. CarbonLayer’s seeded regional values passed through its synthetic grid-intensity calculation are modeled inputs, not measured per-inference emissions.

Simulated

Prototype behavior or output, distinct from impact classifications. POST /api/inference returns a simulated route target and latencyMs; modelProviderCalled: false means no model provider call is made. These latency values are prototype route metadata, not production measurements.

Unavailable

An impact result is withheld when required evidence is missing. CarbonLayer’s per-inference carbon impact is unavailable because serving-energy allocation and a verified emissions factor are absent. Unavailable impact data is not zero.

Reading this record

Impact classification

measured

Direct observation with a stated period and provenance.

modeled

A coefficient, assumption, floor, synthetic value, or attribution input.

estimated

A calculated result derived from named measured and/or modeled inputs.

unavailable

Required evidence, a defensible range, or a verified source is missing.

Receipt facts are kept separate from this legend. A real API response does not turn simulated output, route metadata, or an environmental estimate into a measured result. “Simulated” describes prototype behavior; it is not one of the four impact classifications.

01

Methodology v1.0 status

The Evidence-Based Framework (Pending Final Coefficient Verification) is available as Methodology v1.0. It describes the intended product boundary, evidence requirements, uncertainty treatment, and rules for withholding unsupported outputs.

It is a framework, not an impact result or independent verification. The researcher- reported summary is the current evidence input for this copy update; the full ledger and PDF are absent and the summary has not been independently verified here. No accounting basis is selected, and specific coefficients remain under verification.

Claim rule

Environmental figures must be labeled measured, modeled, estimated, or unavailable according to their evidence. Carbon and per-inference water remain unavailable until their inputs and allocations are selected, documented, and verified.

02

Purpose and product boundary

The scope is operational inference-impact modeling: relating a bounded inference dispatch to the workload and infrastructure evidence needed for an environmental estimate. This page does not certify a facility, model provider, grid mix, product claim, or compliance position.

Real recorded receipt facts

Request attribution, token/workload fields when supplied, a timestamp, and the persisted receipt record. These facts describe the request and its recorded lifecycle.

Simulated or unavailable

Model output and dispatch are simulated in the prototype. No model provider call, serving-stack energy observation, physical routing, facility meter, approved grid factor, or environmental result is established by that simulation.

The system boundary starts at the recorded inference dispatch and ends only when a defensible allocation to energy and environmental factors can be evidenced. Physical routing, provider/model telemetry, facility operation, and the embodied or lifecycle impacts of equipment are not represented when their evidence is absent.

The /api/inference decision receipt is separate from the cost-and-latency Playground. The Playground supports model-based scenarios and an optional direct provider call for one request when configured; it does not route production workloads or provide measured environmental impact.

03

Functional unit and activity data

Primary functional unit

One inference dispatch. This is the unit to which request attribution and a future supported allocation would attach.

Activity data

Token counts, input/output workload fields, model identifier, request timestamp, and any supported request metadata describe activity. A token count alone is not an energy value and cannot be converted into one without an approved factor.

Aggregation period

A future report must state the time period, aggregation boundary, inclusion/exclusion rules, and whether a range is per dispatch or time-bounded. No period or aggregation may be inferred from a single receipt.

No token-only inference

The method does not infer energy from token count alone. Missing serving-stack telemetry, provider/model data, or a documented allocation factor remains unavailable rather than zero.

04

Data sources and input register

Inputs are labeled by how they enter the method. A recorded field is not automatically an environmental measurement, and a modeled assumption is not observed data. The register below is the v1.0 evidence contract; missing source URLs and ranges are intentional disclosures, not completed inputs.

Input class

Receipt activity

Source / period

CarbonLayer API receipt
Request timestamp; exact period is receipt-specific

Treatment

Recorded fact; not environmental evidence

Input class

Serving energy

Source / period

Instrumented serving stack or provider telemetry
Not supplied

Treatment

Unavailable until meter/telemetry provenance and allocation are supplied

Input class

Grid/environmental factor

Source / period

Researcher-reported candidate ledger summary; full ledger not supplied
Accounting basis not selected

Treatment

Candidate values use different accounting bases; no grid coefficient is approved or published

Input class

Water factor

Source / period

AWS region-specific operator-reported withdrawal figures (summary only)
Final source and period verification pending

Treatment

Potential direct/evaporative withdrawal layer only; not a per-inference factor

Grid-source metadata is not approval of an environmental factor. Carbon impact remains unavailable until an accounting basis is selected, documented, and verified. The researcher-reported summary says candidate location-based, hourly, marginal-proxy, and lifecycle figures differ; the full ledger is absent, no basis is selected, and no candidate constants are reproduced here.

Water evidence boundary

The researcher-reported summary says region-specific AWS operator figures may support only a direct/evaporative withdrawal layer, pending final verification. Microsoft’s fleet average is not regional evidence. Indirect watershed impact remains unavailable pending validation, with no proxy substituted. The summary does not establish inference-to-kWh conversion, so per-inference carbon and water remain unavailable. The figures and conclusions are not independently verified by this task. No candidate water values or confirmed water-impact claims are published.

05

Freshness, provenance, and source validity

Each input used in a future calculation must be traceable to a source record and to the dispatch or aggregation period it represents. Retrieval time is not the same as observation time. A source can be current in the cache and still be unsuitable for a historical calculation or an approved environmental factor.

Observation period

When the underlying electricity, meter, workload, or water observation occurred. It must match the dispatch or reporting period; a retrieval timestamp cannot fill in a missing observation period.

Retrieval time

When CarbonLayer obtained a record, kept distinct from its observation time and from a reproducible source identifier.

Validity treatment

The applicable freshness window, expiry behavior, stale-data handling, and why an input is accepted or withheld for the stated period and boundary.

The current /api/grid-intensity contract exposes these separate fields:source category, fetchedAt, and expiresAt. Its cache window is five minutes and expired cache rows are discarded. These timestamps describe retrieval and cache validity, not the grid observation period. The contract has no observation-period or source-record identifier, and marks the response simulatedwith impactDataStatus: unavailable. A recent lookup therefore does not establish provenance or a measured environmental impact.

Classification and missing inputs

Measured requires a direct observation with a stated period and provenance. Seeded or synthetic coefficients and routing values are modeled. Estimated is reserved for a result calculated from named, supported inputs, with its assumptions and uncertainty shown. Without the required evidence, per-inference energy, carbon, water, baseline, and savings are unavailable; missing impacts are not zero and are not replaced by a proxy.

Modeled floor and reproducibility

The synthetic grid calculation uses a 25 gCO₂/kWh floor as coarse attribution. It is a modeled floor, not per-device amortization or a measured per-inference result, and it does not change the unavailable impact status. Disclosed constants can be classified as modeled while remaining unreproducible when their precise provenance or supporting inputs are not named and checkable; disclosure alone does not make them reproducible.

Source URLs not supplied

The researcher-reported summary is not accompanied by the full source ledger or source PDF. Its source-level claims cannot be independently checked from the material available for this update. No unsupported URLs or coefficients are added here.

06

Six-step calculation chain

Methodology v1.0 makes the evidence framework available. The stages below identify the evidence required for an impact result; they are not a verified set of coefficients or a claim that the missing ledger selected an accounting basis. Each impact-producing stage remains unavailable until its inputs and allocation are documented and verified.

STEP 01

Activity and dispatch record

Input
Receipt workload fields, token/workload metadata, timestamp, and supported request metadata.
Operation
Read the recorded activity and bind it to one inference dispatch.
Output
Bound dispatch activity
Unit
one dispatch + metadata
Source
CarbonLayer receipt contract
Evidence
Recorded receipt fact; not an impact measurement
STEP 02

Serving energy allocation

Input
Instrumented serving-stack energy, provider/model telemetry, and an allocation rule.
Operation
No evidence in the supplied summary establishes an inference-to-kWh conversion; no energy value is inferred from tokens.
Output
Energy attributed to the dispatch
Unit
kWh per dispatch
Source
Serving meter or provider telemetry — not supplied
Evidence
Unavailable
STEP 03

Facility overhead / PUE allocation

Input
A documented PUE ratio and its facility boundary, period, and provenance.
Operation
Apply PUE only if v1.0 requires it and the ratio is evidenced; no PUE ratio was applied.
Output
Facility-adjusted energy
Unit
kWh per dispatch
Source
Facility record — not supplied
Evidence
Unavailable; PUE was named but not measured or applied
STEP 04

Carbon attribution

Input
Energy output plus an approved regional/environmental factor with period and source.
Operation
Owner-supplied carbon factor and exact formula are unavailable.
Output
Operational carbon attribution
Unit
gCO₂e per dispatch
Source
Researcher-reported candidate sources; full ledger and source record not supplied
Evidence
Unavailable
STEP 05

Water attribution

Input
Energy output plus a documented cooling/process water-intensity factor.
Operation
Owner-supplied water factor, boundary, and formula are unavailable.
Output
Water attribution
Unit
L per dispatch
Source
Facility/process record — not supplied
Evidence
Unavailable
STEP 06

Range, classification, and publication

Input
All prior inputs, ranges, scenarios, provenance, and missing-data decisions.
Operation
Propagate ranges; retain assumptions; withhold point values, netting, or comparisons when v1.0 disallows them.
Output
A qualified range or an unavailable state
Unit
Per declared functional unit and period
Source
v1.0 governance record
Evidence
Unavailable until steps 02–05 are evidenced

07

Unselected inputs and publication status

The researcher-reported summary describes materially different carbon candidates by accounting basis, but the complete ledger and source PDF are absent. No basis has been selected or independently verified here, so candidate constants are omitted. The ledger also does not establish an inference-to-kWh conversion.

Grid accounting basis

unavailableNot selected

Energy per inference

unavailableUnavailable

Carbon per inference

unavailableUnavailable

Water per inference

unavailableUnavailable

Direct / evaporative withdrawal layer

unavailablePotential; pending verification

Indirect watershed impact

unavailableUnavailable; no proxy

Baseline / savings

unavailableUnavailable

Receipt fields

Request facts recorded separately

Carbon remains unavailable until the accounting basis is selected, documented, and verified. Water estimates remain unavailable until a defensible layer and allocation are verified. No environmental saving or combined receipt is implied.

07A

Hypothetical project-period comparison

Hypothetical example only

Every date, count, label, and latency value below is illustrative. It is not production evidence, proof of provider activity or physical routing, a measured latency result, or measured environmental impact.

Suppose a project selects February 3 through February 4, 2026, UTC. The through date is included in full, and the following midnight is the exclusive end. That makes the selected interval [2026-02-03T00:00:00Z, 2026-02-05T00:00:00Z), exactly 48 hours. The immediately preceding equal-duration interval is [2026-02-01T00:00:00Z, 2026-02-03T00:00:00Z), also 48 hours. The prior period ends exactly where the selected period starts, so the intervals are contiguous and do not overlap. Each interval includes its start and excludes its end.

Selected period · hypothetical

Feb 3–4, 2026 UTC

[2026-02-03T00:00:00Z, 2026-02-05T00:00:00Z)

48 hours

Immediately preceding period · hypothetical

Feb 1–2, 2026 UTC

[2026-02-01T00:00:00Z, 2026-02-03T00:00:00Z)

48 hours

For any supported measure, change means selected period minus previous period. For example, two selected-period records minus two previous-period records equals a change of zero. A change is available only when both periods contain records and the values needed for that measure are valid. If either period has no records, changes are withheld; missing or invalid values also leave the affected measure unavailable.

Persisted decision records and valid timestamps

Selected · hypothetical
2 records; 2 valid timestamps
Previous · hypothetical
2 records; 2 valid timestamps
Selected minus previous
0 records (2 − 2); 0 valid timestamps (2 − 2)

Counts by supported model label

Selected · hypothetical
claude-haiku: 2
Previous · hypothetical
claude-haiku: 1; cl-compact-v1: 1
Selected minus previous
claude-haiku: +1; cl-compact-v1: −1

Counts by supported route-region label

Selected · hypothetical
eu-north-1: 2
Previous · hypothetical
eu-north-1: 1; us-east-1: 1
Selected minus previous
eu-north-1: +1; us-east-1: −1

Persisted latency

Selected · hypothetical
Average 100 ms; minimum 50 ms; maximum 150 ms (2 valid values)
Previous · hypothetical
Average 200 ms; minimum 100 ms; maximum 300 ms (2 valid values)
Selected minus previous
Average −100 ms; minimum −50 ms; maximum −150 ms

Persisted provenance and degraded records

Selected · hypothetical
Live 1; synthetic 0; unknown 1. Known-source denominator 1 of 2; degraded 0 of 1 valid flags.
Previous · hypothetical
Live 1; synthetic 1; unknown 0. Known-source denominator 2 of 2; degraded 1 of 2 valid flags.
Selected minus previous
Live 0; synthetic −1; unknown +1; known-source denominator −1; degraded −1; valid-flag denominator −1

Read counts as simulated record metadata

A decision-record count is one persisted project-associated decision row, not a count of provider calls, retries, or fallback attempts. Model labels, route-region labels, latency, and provenance are persisted simulated record or route metadata. They do not establish provider/model activity, physical routing, provider latency, or environmental impact. Time coverage records the earliest and latest valid timestamps, along with the valid timestamp count. Model and route-region breakdowns count only supported persisted labels. “Live” is a persisted source label, not proof of a live provider call.

Missing and invalid fields remain visible as data-quality warnings

In this hypothetical selected period, one record has neither a source label nor a degraded flag. Missing or unrecognized source labels are counted as unknown and flagged; the source denominator counts only valid live or synthetic labels. The degraded count counts true values among valid true/false flags, so its denominator is one here; missing or invalid degraded values are flagged and excluded from that denominator. Missing record IDs, missing or unsupported model and route-region labels, invalid timestamps, and missing, negative, or non-finite latency values are also reported as data-quality issues, with affected fields excluded from their respective measures. A record with an invalid timestamp cannot be assigned to either period and is counted separately as unassigned.

Requested tokens and environmental results · unavailable

These values are unavailable for both periods and their changes. Unavailable does not mean zero.

Requested-token totals
Selected: unavailable · previous: unavailable · change: unavailable
InferenceRequest.tokens is absent from the confirmed database schema, so no total is read or substituted from another field.
Energy
Selected: unavailable · previous: unavailable · change: unavailable
Environmental evidence and a supported inference-to-energy allocation are missing.
Carbon
Selected: unavailable · previous: unavailable · change: unavailable
Required energy allocation and a verified emissions factor are missing.
Water
Selected: unavailable · previous: unavailable · change: unavailable
A verified water boundary, factor, and allocation are missing.
Baseline
Selected: unavailable · previous: unavailable · change: unavailable
A supported environmental baseline and its evidence are not available.
Savings
Selected: unavailable · previous: unavailable · change: unavailable
A supported baseline and comparable impact results are not available.

Missing evidence stays unavailable

No environmental amounts or deltas are calculated or implied here. Missing environmental evidence does not become an impact result.

08

Uncertainty and missing evidence

Uncertainty comes from serving energy, workload allocation, provider/model telemetry, physical routing, facility overhead, grid factors, water factors, observation periods, and the quality or staleness of each source. A future release must show input ranges or scenarios, the propagation method, the resulting output range, and which assumptions remain modeled.

Missing is unavailable, not zero

CarbonLayer does not substitute zero for missing energy, carbon, water, baseline, savings, SCI, PUE, or compliance evidence. Unsupported figures stay unavailable until evidence supports them.

Point-value and netting rule

Do not collapse a required range into a point value. Do not publish a baseline, saved comparison, or netted result when the inputs or v1.0 rules do not support it.

Measured

Direct observation with stated period, provenance, and an operating/instrumented serving stack or external source.

Modeled

Seeded coefficient, assumption, floor, synthetic grid value, or modeled attribution input.

Estimated

A calculated environmental output from named measured and/or modeled inputs, with its range and assumptions shown.

Unavailable

A required factor, range, source, boundary, or verification record is absent or cannot be defensibly formed.

09

Workload-Simulator Readiness Checklist

Before evaluating workload-simulator inputs, provide baseline evidence for latency, cost, carbon, and water. Supply the source, scope, and period for each baseline; do not treat simulator targets or illustrative values as observations.

  • Latency baseline

    Provide representative measured observations for the workload and measurement point, with the aggregation period and latency distribution used. Keep simulated route targets labeled as modeled.

  • Cost baseline

    Provide applicable provider/model pricing or billed-cost records, rate version, units, workload assumptions, and period. Distinguish invoice evidence from modeled prices.

  • Carbon baseline

    Provide the selected accounting basis, energy allocation evidence, factor source and version, location and time coverage, units, and formula. CarbonLayer currently publishes no per-inference carbon baseline or impact.

  • Water baseline

    Provide the water boundary and layer, facility/process record or verified factor with source and version, location and period, units, and allocation rule. CarbonLayer currently publishes no per-inference water baseline or impact.

Uncertainty context for every baseline

  • State applicable ranges or scenarios and assumptions, including workload allocation, provider/model telemetry, physical routing, facility overhead, and grid or water factors. Identify which inputs are observed and which are modeled.
  • Name the observation period separately from retrieval time, and provide source provenance, freshness window, quality, and known gaps.
  • Explain how uncertainty is carried through the calculation into an output range; keep unresolved evidence unavailable instead of substituting zero or an unsupported point value.

Evidence request, not current measurement

This checklist asks what evidence to provide; it does not mean CarbonLayer currently measures or validates these baselines. Per-inference carbon and water impact outputs remain unavailable, and missing evidence is not replaced with a proxy.

10

Revisions, exclusions, and reassessment

Methodology v1.0 is available as the public framework; coefficient verification remains pending. Future revisions must identify changes to the boundary, functional unit, source record, factor, allocation method, uncertainty treatment, or publication rules.

Any future estimates may be reassessed when required evidence and defensible input ranges become available. Reassessment does not turn a modeled value into a measured result, and no independent verification of the researcher-reported summary is implied.

  1. Methodology v1.0Available evidence framework. Specific coefficients and the accounting basis remain pending final verification; environmental impact outputs stay unavailable.

Excluded until supported

Embodied/lifecycle impacts, physical routing claims, facility-meter claims, provider model telemetry, unsupported PUE allocation, and unverified source factors.

Not a claim

This release is not an SCI score, compliance certification, independent audit, or statement that a particular facility or provider produced a measured impact.

11

Limitations, audit corrections, and source record

The current prototype reports request attribution, token count when supplied, timestamp, and simulated route and latency metadata. The route target and latency values are illustrative, not measured results or evidence of physical routing. Its receipt boundary remains explicit:simulationStatus: simulated, modelProviderCalled: false, andimpactDataStatus: unavailable. The API receipt is real; its simulated model and dispatch fields are not measurements.

  • No model provider call or request-level serving-stack measurement is made.
  • No physical routing, facility meter, PUE ratio, or edge-site claim is established.
  • Embodied emissions, water, SCI, compliance, baseline, and savings remain unsupported.
  • Seeded and synthetic values are modeled. Environmental results are estimated only when named, supported inputs and ranges support the calculation; otherwise energy, carbon, water, baseline, and savings are unavailable.

Prior site-audit corrections

In the referenced audit environment, robots.txt was a full disallow:User-agent: * and Disallow: /. The audit also found that the 404 pages were not byte-identical. These are audit corrections about site behavior, not methodology evidence and not proof of an environmental result.

Source register

Researcher-reported grid summary

The summary describes different location-based, hourly, marginal-proxy, and lifecycle candidates. The full ledger and source PDF are absent, no basis is selected, and the reported material is not independently verified here. Candidate constants are omitted.

Researcher-reported water summary

The summary says regional AWS operator-reported withdrawal figures may support a direct/evaporative layer after verification. A Microsoft fleet average is not regional evidence; indirect watershed impact and per-inference water remain unavailable. The figures are not independently verified here.

Repository seed and synthetic grid adapter

Internal modeled inputs and synthetic fallback behavior. These can explain prototype behavior but cannot establish a measured per-device environmental result.

Full ledger and source PDF

Neither is attached to this task. This page uses only the researcher-reported summary, does not reproduce candidate constants, and does not treat the summary as independent verification.

For the current receipt contract, read the API docs. For the concise boundary note, read the training and carbon math. Those pages describe prototype behavior; they do not provide evidence for unavailable impact values.