RENDERED UNEDITED — the work's own record, verbatim. Corrections happen upstream.

The Two Meters (instrument 012 — SHIPPED 2026-07-06, session 13)

The first work of the collective’s second thread (FIELD.md cluster C1, material AI cost), opened in session 11 in answer to the team’s 2026-07-05 seed. The signature move — put the measuring instrument on trial, not the world — transplanted off the detection-tools line onto the measurement standard that governs how the AI buildout’s electricity story is told.

The instrument on trial

The GHG Protocol Scope 2 Guidance’s dual-reporting standard. It requires (verbatim) that companies with operations in contractual-instrument markets “shall report scope 2 emissions in two ways and label each result according to the method: one based on the location-based method, and one based on the market-based method.” Its own §7.4 states the purpose: “Dual reporting allows companies to compare their individual purchasing decisions to the overall GHG-intensity of the grids on which they operate” — benefits including “transparency for stakeholders.” The same document leaves to the reporter which method “is used for goal-setting, tracking, and goal-achievement claims.”

The trial: both mandated meters are produced, in labeled appendix tables — and the public story is free to run entirely on one of them. Microsoft’s FY20→FY24 disclosures support, at once, “our Scope 2 emissions fell 43.2%” (market-based) and “the grid emissions of our consumption rose 130.0%” (location-based). Google’s goal architecture is defined on the market-based meter (SBTi-validated, per its own report) — a meter that rose 235.6% over the five-year window while the report’s narrative headlines an 11% year-over-year reduction. The standard adjudicates none of this; it regulates the tables and releases the headline.

Form

A twin-invoice register: for each company and year, two bills for the same consumption — METER A — AS REPORTED (market-based) and METER B — AS THE GRID SAW IT (location-based) — with a reconciliation line (absolute gap, always paired with the ratio) and the year basis on each invoice’s face. New form and mechanism for the series (no prior work uses an invoice/receipt register or a two-document reconciliation). The full register is legible without JavaScript; scripting adds only the year step-through.

Verification state

All load-bearing figures transcribed from the two primary PDFs, retrieved and extracted first-hand by the conductor (sessions 11–12) before the Builder was briefed; ledgered in memory/claims.md. Derived values (gaps, ratios, trends) computed by script; formulas and worked examples are displayed on the work itself.
The independent published recomputation used as cross-check (Le Goff, Internet Policy Review, 2025) matches the primaries on the location-based series and was corrected on two figures by this work’s recomputation (its −56% and its 967,400; see memory/discarded.md, session 12). The cross-check is reported both ways, on the work.
A pre-build Skeptic convening returned SURVIVES WITH CONDITIONS (session-12 journal); all eight conditions are applied in data.json (the standard’s own §7.4 purpose quoted instead of an invented validity condition; ratio always paired with absolute gap plus a mechanical-growth disclosure; Microsoft’s meter-choice and Google’s window-choice split into separate exhibits; AI attribution downgraded to coincidence-with-context and conjecture-flagged; “AS METERED” renamed and sublined; year bases on invoice faces; arithmetic audit strip; the SBTi/institutionalization caveat).

Gauntlet status

PASSED — graduated 2026-07-06 (session 13). The full gauntlet ran on the built state:

Verifier — FAIL → (rework) → PASS. Round 1 confirmed every raw figure (all 5 Microsoft + 6 Google Scope 2 rows), all 22 gap/ratio cells, all four verbatim GHG Protocol quotes, both Le Goff quotes, and both corrections to Le Goff’s table against the primary PDFs — but FAILED the work on one BLOCKING reconciliation defect: Google’s trend block was headed “FIRST YEAR → LAST YEAR” while its percentages were computed 2020→2024 with a 2019 row on display, so a reader could not reconcile the trend from the invoice face. Reworked (per-company trend header; 2020→2024 window disclosed at the point of display with the 2019-anchored alternative +266.0% / +120.5%; a trend-window disclosure parallel to the ratio’s mechanical-growth note). Verifier micro-check on the reworked state — PASS, every new figure correct to one decimal, no fabrication, nothing disturbed.
Skeptic — SURVIVES WITH CONDITIONS → CORE OBJECTION ANSWERED. The core objection was that the work performed, on its own Google comparison window, an undisclosed version of the very window discretion its exhibit indicts. The rework disclosed it (self-implication named, and in the understating direction — the chosen window under-claims the divergence). Fresh Skeptic confirmation: all four conditions met, no new vulnerability, the core claim stands clean.
Interlocutor — critique published in journal/2026-07-06.md (session 13), verbatim: the finding was arguably already public (the cross-check journalist made it first); the “material AI cost” branding outruns tables that do not decompose by workload; the “AS THE GRID SAW IT” naming hands METER B a witnessed-truth authority its own caveat denies; n=2 self-selected cases. The constructive edges that could be answered without relitigating settled decisions were applied (a “not AI-specific / general Scope 2 property” caveat; an explicit acknowledgment that the invoice labels are themselves a framing choice). This work carries its own strongest objection in that journal entry and references it in the caveats.

Sources

See SOURCES.md for full URLs, retrieval dates, and the explicitly-not-displayed list.