Process lives here. Sections Ⅰ–Ⅸ describe what is common to every Grid print. They are revised as a document — the changelog at the foot is the record — and a change to a rule in them is announced by notice before it takes effect, like any other methodology change.
Specifications live one per series at /standards/instruments/<code>: class, basis, unit, close, hierarchy, normalisation, minimum observations, disclosures, revision history. A suspended series keeps its page. Nothing on a specification page repeats process; nothing here repeats a specification.
The changelog, the notices, the corrections, the participation rules and the terms are URLs, not chapters — /notices, /corrections, /participate, /terms — because a footnote has to resolve to a page, and an agent cannot search a chapter.
The Grid applies the hierarchy every price-reporting body applies (IOSCO PRA Principle 2.2): concluded transactions first, then bids and offers, then other market information. Where The Grid’s inputs sit is not a secret and not a footnote — it is stated per series on the specification page and, for the flagship series, on every print in the provenance stamp.
| Tier | What it is | Where The Grid sits today |
|---|---|---|
| 1 · Concluded transactions | Executed prices between counterparties. | None. Transactions: 0 on every print. No submitted transaction data has been received (participation v0). |
| 2 · Bids and offers | Firm, executable prices — parties prepared to transact at them. | GAP-H100: tier-2 bookable asks, availability-verified from two public venues. Token pricing (RWI paths): published per-token list prices, executable at the posted rate. |
| 3 · Other market information | Reported prices, indices, closes, filings, commentary. | DRAM spot quotes (TrendForce), equity closes (GCCX), interconnection-queue and filings data (GPI factor), OCPI (Ornn — cited, never originated), news and transcripts (context and records only — never a price input). |
The GAP-H100 print carries a provenance stamp written into the print of record at print time and read — never recomputed — by every surface: panel size, each contributing provider with its unchanged-days, transaction count, data tier, and judgment applied. In words, today it reads like this (contributors and days vary):
All contributing sources are public and named on the print — venue APIs for executable asks, provider price cards, a price-reporting source for DRAM spot, end-of-day equity closes, public filings and status pages, podcast and news capture for the records. The Grid takes no proprietary data, no insider information and no NDA-protected relationship. Two exclusions are stated plainly on the GAP-H100 specification: SF Compute’s clearing-price feed is not ingested (terms of service), and posted list prices are a secondary check, never an input.
No submission channel is open yet and no submission has been received. What a submission is, where it goes, how it is receipted and held in confidence, and the panel thresholds it must clear are stated at /standards/participate. Confidentiality (never attributed, never published individually, never named) will apply to submitted data; nothing on the site today is submitted data.
Every printed number is produced by a deterministic estimator from the day’s stored inputs, in code that is versioned per series (methodology_version is stamped on every print). No printed level, change, index or spread carries human or model judgment: the judgment field on the provenance stamp reads “none (arithmetic)” and there is no exception path that lets a value be set by hand. Where a rule would need judgment — an outlier disposition, a benching — the rule does not exist yet and the specification says so.
| Family | Estimator | Where specified |
|---|---|---|
| Levels (GAP-H100, R1Y) | Provider-pooled median: each provider's qualifying offers reduced to a median first (one vote per provider), then the median across providers. Sanity-bounded; class- and tier-filtered; per-GPU normalised at ingest. | GAP-H100 |
| Changes (every series) | Like-for-like only: the median of per-series changes across matched (provider, SKU) or single clean series present at both endpoints — never the difference between two daily medians of a changing basket. Fewer than three matched pairs: no change is published. One shared velocity module per runtime; no surface computes its own rate of change. | scripts/velocity.py ↔ dashboard/lib/velocity.ts |
| Composites (GPI) | Weighted sum (0.30 / 0.25 / 0.30 / 0.15) clamped to [0, 100]. Three factors are z-scored against their own 30-day history and mapped to 0–100 (50 = mean, ±3σ, clamped [5, 95]; fewer than two observations → neutral); Supply Chain Forward is a bounded mean of level signals, not a z-score — the specification states each factor as implemented. | GPI |
| Indices (GCCX, RWI-S/L) | Equal-weight rebased index (GCCX, panel frozen at base, carry-forward, corporate-action guard); cheapest-path cost rebased to 100 at first print (RWI). | GCCX · RWI |
| Spreads (GTS-1Y, Grid Basis, Inference Spread, divergence) | Subtraction of two stored legs of the same date, sign locked in the registry header; a spread prints only when both legs printed; a qualification gate (Grid Basis) withholds until the third-party leg has seven consecutive api days. | Grid Basis · divergence |
No number is produced by a language model. Models are used in two places, both outside the print: (1) classification and extraction for the records — claims, speakers, topics from transcripts and news, roster- constrained and audited; (2) the daily narration (THE CALL · THE ASSESSMENT · THE MOVE · THE SCORE), which is generated from code-owned inputs, with every delta pre-rendered and labelled by code, and then gated (§Ⅳ). A reader can always tell which is which: the numbers are on the print; the prose is around it.
| Element | Rule |
|---|---|
| The cycle | The daily cycle is anchored at 18:30 UTC (the scheduled fire). The print date is the cycle date, derived from the anchor — never from wall-clock time. Incident of record: the 2026-08-06 cycle started late, crossed midnight, and no 2026-08-06 print exists; the gap is filed as a correction of record and is never backfilled. |
| The print of record | One stored print per series per date, methodology version stamped, provenance stamped where the series carries it. Every surface — the briefing, /instruments, the daily report, the newsletter, /api/print, MCP — reads that print. No surface recomputes a printed value (single-provenance discipline). |
| Publication time | The daily appears with the cycle; the masthead states 04:30 AEST. The masthead issue number is continuous from the first daily (No. 1 = 2026-05-05) and is never reset. |
| Permanent URLs | https://thegridco.ai/api/print/<date> is the permanent citation URL for a print. The body for a past date never changes except for entries appearing in its corrections array; provenance is the stamp stored in the print, null for pre-stamp prints, never synthesised retroactively; continuity is computed as of the print date; the cite string carries no retrieval date. |
| Series status | A series on a dated print is printed (the stored value), not_printed (no row for the date — never carried forward), or suspended (on/after the effective date; notice reference, no value). From 2026-08-22 a fourth status, withheld, appears for GAP-H100 spot with its reason and stamp (§Ⅶ). |
| Immutability | Published history is immutable. Methodology breaks are annotated, never rewritten. Two annotated exceptions precede the rule, both before any external subscriber: the 2026-06-12 one-time recompute of the 11 May – 10 June GPI history (two corrupt factor inputs; pre-recompute series archived) and the 2026-06-29 forward correction of a v1.0 formula defect (thirty-one days of window-drift left in the series as honest history). From the first external subscriber onward the series stands as printed. |
| Assessed vs calculated | Every value is labelled by class — assessment · calculated · indicator · record — on its specification page and in the API body; the class is part of the print, not commentary. |
| The newsletter | The daily is sent from briefing@thegridco.ai to the subscriber list as the edition publishes; its content is the same print. An as-sent archive of the email itself is in development; the edition of record is the site and the permanent URL. |
The defect of record (corrections index 2026-08-14, six dailies): narration attributed the 7-day matched-pair delta — or its absence, rendered “unchanged” — to the printed GAP-H100 level, whose day-over-day move it was not. Both numbers are legitimate; conflating them is the defect. The prompt persuades; the gate guarantees.
Scope. Every sentence of the generated narration that names the instrument (GAP-H100 / assessed spot).
Test.A claimed delta in such a sentence must match either the print’s own day-over-day move or an explicitly labelled like-for-like figure (like-for-like · matched-pair · 7d · re-base / methodology-break disclaimers count as labels). A stillness claim (unchanged · zero-delta · held at · holds at · flat · static) is a claim the print did not move, and fails when it did. Ranges, prior-session references and yesterday-attributed figures are recognised as such (false-positive classes A–D, fixed 2026-08-16 against a 21-daily backtest and pinned by tests).
On failure.The generated prose is withheld and a deterministic, register-only issue publishes in its place — every number code-owned, the dated lead, the seven-segment register with dual-labelled deltas, and a one-line statement that today’s narration failed the print-vs-like-for-like delta gate. A trip marker is written and committed with the daily; the invariants step reads it, reds the run and sends the alert. A trip is never silent and never rewritten. From the 2026-08-16 cycle the gate writes a marker on every run (status clean, tripped or not evaluated, with the deltas it validated) so that a gate that ran clean and a gate that never ran can never look alike — an absent marker is itself a violation.
By construction.Since 2026-08-16 the GAP row’s deltas travel to the model pre-labelled by code (print delta · composition flag · like-for-like 7d) and the ASSESSMENT instruction requires the delta slot to be copied verbatim — a verbatim copy passes the gate by construction.
A correction restates what a value or a narration should have been on the information set of the day: a clerical error, a calculation error, a misapplication of the stated method. It is never a re-assessment on information learned after publication, and it never rewrites the print. The published value stands; the correction is filed as its own document with a reason code, appended to the affected daily as a CORRECTION NOTICE (annotation, never a silent rewrite), registered, and served inline in the permanent print URL’s corrections array. A defective method is not a correction — it is a change, by notice (§Ⅵ).
| Code | Meaning |
|---|---|
| methodology_effective_date | A methodology change governed a print before its announced effective date (ADR-003 §2 publish-before-effect); values stand, the effective date is corrected. |
| rebase_contamination | A declared panel re-base leaked into a derived print as a market move; rows rewritten under a chain-linked methodology (v1.1), pre-correction values archived. |
| narration_composition_as_price | Panel-composition change narrated as price behaviour; values stand, narration corrected by annotation. |
| narration_absence_as_stillness | Absence of a like-for-like measurement narrated as the print's stillness while the print moved on intra-panel repricing; values stand. |
| narration_matched_pair_as_session_delta | The 7-day matched-pair (like-for-like) figure narrated as the session's own move; values stand. |
| print_missing | No print of record exists for the date; the gap stands and is never backfilled. |
The full record — every correction, what was published, what stands corrected, whether a value was restated (two rows in the record; every other correction annotates) — is at /standards/corrections, the same registry /api/print/corrections serves.
filed date is the record. A stated time-to-file commitment is in development and will be published here when ruled.A methodology change is announced and dated before it affects a print — never discovered after. The one breach on record is the incident of record: the GAP v3 executable-ask panel governed the 2026-07-24 print one cycle before its announced 2026-07-25 date; the effective date was corrected on the corrections rail (2026-07-27) and the breach is named on the specification. Every notice since has run its period before binding.
| Step | What it is | Where it lives |
|---|---|---|
| Proposal (when the change is a new rule) | A dated notice stating the rule, what it would have done on the existing record, the mechanism, and its own proposed notice period. Nothing is built. | /standards/notices |
| Decision | A second dated notice: adopted (as proposed or modified), the notice period, the effective date, the representation choices, and the disclosure that rides with the series from publication. | /standards/notices |
| Notice period | The stated interval before the rule binds. Deviations from the specification's guidance are stated in the notice with the reason. | in the notice |
| Effective | The first print the change governs. Suspensions name their final print date; disclosures take effect on publication with no period. | in the notice · on the specification page · in the API body |
| Record | The specification page's revision history is updated; the disclosure rides in the derived-series registry (both runtimes) and appears on every surface that displays the series. | /standards/instruments/<code> |
| Notice | Change | Proposal | Decision | Notice period | Effective | Deviation stated |
|---|---|---|---|---|---|---|
| 2026-08-14 instrument actions | GAP-H100-R1Y and GTS-1Y suspended (zero variance); GAP-H100 spot reclassified assessment → indicator; provenance stamp added. | — | 2026-08-14 | five business days, 2026-08-17 → 2026-08-21 (suspensions); reclassification and stamp effective on publication | 2026-08-22 (final prints 2026-08-21) | Yes — §4: five business days against the specification's three-month negligible-impact cessation guidance; no contract, derivative or ledger entry references the series, zero external subscribers; holding disclosed-defective instruments in print for three months would prioritise procedure over honesty. |
| minimum-observation rule | GAP-H100 spot withheld when fewer than two contributors or all contributors static ≥ 7 days; withheld days disclosed with the stamp, never silent; sibling record. | 2026-08-14 (own notice) | 2026-08-16 — adopted unmodified | five business days, 2026-08-17 → 2026-08-21 | 2026-08-22 print | Yes — §3: same reasoning as the 2026-08-14 notice §4; the rule only withholds prints that would otherwise carry disclosed-defective information. |
| Commitment | Statement |
|---|---|
| Public source priority | No proprietary data, no insider information, no NDA-protected relationships. Value comes from structure, methodology and daily discipline — not access. |
| No investment positions | The Grid holds no investment position in the companies it covers, no position in compute or compute derivatives. Structural commitment. |
| Methodology transparency | Weights, formulas, band definitions and normalisation methods are published in this library. Changes are versioned and announced before effect. Historical values are labelled with their methodology version. |
| Permanence | Colour conventions are permanent. Published series values keep their original meaning indefinitely; breaks are annotated, never rewritten. |
| Prediction accountability | The Grid does not publish forecasts. Forecast series verify privately until a record exists (identity decision of record, 2026-07-24); public accountability is the dated record — the Ledger of others' dated claims, scored only when the date arrives, and the dated print archive. No forecast quality is claimed before a record supports it. |
| Separation of commercial activity | The only commercial arrangement that touches a surface — provider referral on /providers — is disclosed on that page (neutrality structure of 2026-07-24). No engagement client exists. No party's commercial relationship influences a printed level. |
| Ownership | No market participant holds equity in The Grid. |
| Complaints | One address (standards@thegridco.ai), stated response times, every complaint logged — /standards/terms §Ⅲ. |
| # | Segment | Coverage | Coverage note | Buyer decision | Cadence |
|---|---|---|---|---|---|
| 1 | Fabrication | [proxy] | revenue-level proxy — no fab-allocation or utilization visibility | diversify / demand-proof the silicon roadmap | monthly |
| 2 | Packaging & Assembly | [gap] | no direct CoWoS allocation or packaging-capacity series | secure packaging-gated allocation, or wait for the CoWoS ramp | event |
| 3 | Memory | [proxy] | DDR5 spot instrumented; HBM not instrumented (equity + DDR5 derivative only) | lock memory-dependent capacity now or hedge the exposure | weekly |
| 4 | Power & Cooling | [proxy] | forward pipeline only — no current power cost/availability | factor power lead time into where and when to site capacity | monthly |
| 5 | Networking & Interconnect | [gap] | equity colour only — no supply-chain instrument | monitor optics/interconnect for system-level lead-time surprises | event |
| 6 | Compute Supply | [instrumented] | GPU rental pricing (GAP-H100, 2-provider panel) + OCPI + incidents | lock, wait, or hedge the core procurement decision | continuous |
| 7 | Demand & Deployment | [proxy] | token pricing + hyperscaler capex — no direct deployment series | proof the need against efficiency gains, or time entry to a demand surge | weekly |
Rendered from the one segment definition both runtimes share (parity-tested). [instrumented] — a direct daily series stands behind the state. [proxy] — an adjacent series stands in (an equity, a derivative, a monthly release). [gap] — no live instrument; reported as a documented hole, never fabricated, and never a state: a [gap] segment carries no state at all. Networking & Interconnect is structural-only (two equities are not a supply-chain proxy) and carries no forced buyer verb.
Colour is buyer-side inverted and permanent: green favourable for the compute buyer (well-supplied), amber neutral (balanced / firm), red unfavourable (tight). This is the opposite of equity-market convention. The whole-market read is the GPI band (Abundant · Balanced · Tight · Constrained · Critical — see the GPI specification); “Tight” and “Balanced” appear at both scopes — the same condition at different altitude.
| Zone | Segments | Characteristic lead time to buyer-side pricing |
|---|---|---|
| Upstream — where physical constraints originate | Fabrication · Packaging & Assembly · Memory | 4–6 months · 3–6 months · 3–6 months |
| Midstream — where capacity is sited and integrated | Power & Cooling · Networking & Interconnect | 12–36 months · 3–6 months |
| Downstream — where buyers experience the effects | Compute Supply · Demand & Deployment | 1–4 weeks (pricing near real-time) · real-time to 1 week |
The propagation lag between segments is always quantified, never asserted. The connected map at /supply-chain holds the nodes and lagged edges and supersedes the three illustrative dependency chains printed in Grid Standards v1.2–v1.3.
| Tag | Class | Meaning |
|---|---|---|
| OBS | Observed | Directly captured from a public source with machine-readable provenance. |
| AST | Assessed | Editorial synthesis with stated confidence and reasoning basis. (Note: this editorial tag predates, and is distinct from, the Part B instrument class "assessment" — no series is in that class today.) |
| EST | Estimated | Model-derived value where observed data is unavailable; methodology stated. |
| Horizon | Format | Confidence |
|---|---|---|
| Today (0–24h) | Specific numbers with deltas | Very high |
| This week (1–7d) | Ranges with directional indicator | High |
| This month (1–4w) | Directional probabilities | Medium |
| This quarter (1–3m) | Above/below baseline with probability weights | Low–medium |
| Structural (3+m) | Narrative with causal attribution | Low on timing, high on direction |
Format communicates confidence; as the horizon extends, the format shifts from precise to probabilistic to directional. Since the identity decision of 2026-07-24 the longer horizons are not published as forecasts; the format governs how a recorded third-party claim or an editorial reading is written when one is quoted.
The daily market-condition word of v1.0–1.1 (the GMCC — clear / cloudy / rain / storm, a weather register) is retired from every published surface as of v1.2: it printed the same word for thirteen consecutive days while conditions underneath it moved. The internal classifier survives only as a computed input and is never surfaced. “CPI” was renamed GPI in v1.1 with no change to the calculation; “sub-indicators” are “contributing factors”. Historical editions that printed the retired words stand as published.
| Version | Change |
|---|---|
| v2.0 — 2026-08-17 | The library. The single-page Grid Standards v1.3 is reorganised into permalinked pages: this master document (process), one specification page per series from the A4 field template (with class, panel, normalisation, close, disclosures, revision history — suspended series included), the methodology notices rendered verbatim, the corrections of record over the API's registry, participation v0, and the terms with the citation and redistribution policy. Content brought to the 2026-08-14/16 state of record: no series in the assessment class; GAP-H100 spot an indicator with a provenance stamp; GAP-H100-R1Y and GTS-1Y suspended from 2026-08-22; the minimum-observation rule adopted, binding from 2026-08-22; the narration gate; the corrections reason codes. A documentation change, not a methodology change: no printed value, method or notice period changed with it. Rules that do not exist yet are marked in development. |
| v1.3 — July 2026 | Consolidation & instrument honesty. The standalone methodology page merges into this document (Part I Conduct · Part II Methodology). GAP classified as a posted-price composite with its exact leg composition documented (§3.1: RunPod posted rate · Vast.ai lowest marketplace asks); Grid Basis restated as cleared settlement vs posted list (§3.3). Second index-history annotation recorded (§2.6, 29 June 2026 formula correction). Retired weather/GMCC condition vocabulary documented (§2.7). Prediction-accountability wording brought to honest tense (record publishes as it accrues). [Reading note: the v1.3 GAP description covers the panel through 23 July 2026; from 24 July 2026 the v3 executable-ask panel governs — see the GAP-H100 specification.] |
| v1.2 — June 2026 | Surface & spine lock. The seven-segment spine becomes the comprehension frame, with a single per-segment state vocabulary (tight · firm · balanced · soft · well-supplied) and load-bearing [instrumented] / [proxy] / [gap] coverage tags. The legacy daily condition-classification table is retired; the GPI band scale is the published whole-market read. The supply-chain taxonomy moves from twelve layers to seven segments. The editorial spine is THE CALL · THE ASSESSMENT · THE MOVE · THE SCORE. |
| v1.1 — June 2026 | Terminology: CPI renamed GPI (Grid Pressure Index); "sub-indicators" are "contributing factors". Methodology brought current: GPI v2.0 z-score normalisation with four contributing factors (30/25/30/15). Added: index-history immutability and the one-time 12 June 2026 recompute annotation, GAP-H100 definition, like-for-like velocity methodology, OCPI qualification rule and Grid Basis suppression. |
| v1.0 — May 2026 | Initial publication of the open vocabulary. |