Methodology changelog

v0.7.1 — 2026-09-06 (current)

Citation integrity + disclosure alignment — no index math change

No weight, tier map, trailing window, or Indicative threshold changed, and no published index level moves as a result of this entry. Every item below is either a correctness fix to issue numbering or an alignment of published text to the behavior of the production compositor that generated the live print. A separate reconciliation is in progress between that production code and the engine repository, which currently differ; no further print will ship until that is resolved. Version note: this is the published methodology version. The engine repository tracks its own software changelog and records this same change as its v0.6.1 engine note; the two sequences are deliberately separate, and public citations should use the methodology version (v0.7.1).

• Issue-number epoch corrected. The exporter derived issue_no from a 2026-01-05 epoch, which computes 35 for the 2026-09-04 print — a print published and cited as issue 14. The exporter could not reproduce its own live print, and re-running it would have silently re-minted every citation and /v/I/{n} permalink. Corrected to the 2026-06-01 relaunch epoch: issue_no = ((as_of − 2026-06-01) // 7) + 1, which reproduces both the live print (14) and the 2026-07 archive (6). Regression tests now pin the epoch against the published trail.
• Issue-number crosswalk disclosed. The published trail is non-monotonic. 2026-05 (issue 21) was stamped under the retired epoch; 2026-06 and 2026-08 carry a literal issue 1 that is not derivable from any single weekly epoch; the artifacts establish those values but not why they were stamped that way, so their provenance is recorded as unknown. Frozen archives are not rewritten — they remain canonical as published, and this crosswalk is how a citation is reconciled against the current formula.
• Cadence clarified. The issue counter advances weekly from 2026-06-01 while issues ship on a monthly deep-dive rhythm. An issue number is a week index, not a count of published issues: issue 14 does not imply fourteen prior issues.
• Indicative gate text corrected. Published methodology described the gate as "≥30 comps per component per 7 days." That described the pre-RT-20 gate, which was real: until 2026-05-29 SWII went Indicative when its transceiver comps or FS SKU count fell below per-component minimums. RT-20 replaced that rule with a weight-adjusted total — SWCI 20, SWII 20, SWPI 10, SWFI 30 — and flipped published Indicative flags in that print. The published text was never updated to follow. This entry corrects the text, not the gate: no index's Indicative flag changes value now.
• "Equal-weight" claim corrected. Copy described the composites as equal-weighted. The engine composites on explicit unequal weights (SWCI weights frontier GPU 0.35 and HBM 0.25 against 0.04 for server DDR4 RDIMM). Copy now says weighted and points to the published component_weights.
• Component drop-out and renormalization disclosed. A component leaves the composite when it yields no value for the window, and surviving weights are renormalized. SWPI and SWFI publish the applied weights as component_weights_used; SWCI and SWII do not yet carry that field. Yielding no value is distinct from reporting no comparables — a component can carry a value forward on zero new comps and keep its full weight, and can log comps yet yield no value. As of the 2026-09-04 print the renormalized share is SWCI 20%, SWII 30%, SWFI 30%. No reweight was performed — this is existing behavior, now stated plainly rather than silently applied.
• MoM interpretation limits published. Where components are mechanically carried rather than re-measured each window — 80% of SWII by weight — a MoM figure can move without an underlying transaction. In the 2026-09-04 print SWII's series is eleven identical points followed by a single step, and both volume-varying components reported zero new comparables. That delta cannot be attributed to current transacted-price movement. Separately, SWII is Indicative because its weight-adjusted comp count is 2.75, below the threshold of 20.

v0.7 — 2026-07-01

Period comparison switched from WoW to MoM

• MoM replaces WoW. The period delta published alongside each index level is now month-over-month (28-day lookback) rather than week-over-week (7-day). The comparison window is the current rolling 7-day median vs. the same 7-day window 28 days prior. Field name in indices.json changed from wow_pct → mom_pct. Rationale: secondary GPU markets can go 8+ days without a credible completed transaction; a 7-day comparison amplified collection-density noise rather than genuine price signal. The 28-day window stabilizes medians across collection gaps and matches the natural frequency of physical infrastructure pricing, analogous to ISM/Markit PMI methodology. First MoM delta published with Issue 2. The date 2026-07-13 was carried in this entry as that issue's as_of, but no published artifact for 2026-07-13 exists in the record (the adjacent prints are 2026-07-10, issue 1, and 2026-07-11, issue 6), and the weekly derivation would give 7 for that date. The date is therefore unverified and is not part of the citation crosswalk; cite Issue 2 by issue number, not by that date.

v0.6 — 2026-06-03/04

PRA hardening + multi-vendor sourcing

• SWCI frontier component multi-sourced. The 35%-weight Frontier GPU OEM component was discovered to be 100% single-vendor (Supermicro via Thinkmate). Dell publishes no public frontier-node price (quote-only); CDW doesn't list frontier nodes. Two independent integrator sources added (Exxact, Bizon — B200/H200/MI300X configurator list prices). frontier_n_sources 1 → 3.
• Single-source guard. New single_source_warning flag: fires when a component at ≥15% weight rests on fewer than 2 independent price vendors. Was true at introduction; cleared by the multi-sourcing. Permanent — if a vendor feed dies, the index re-flags itself.
• Per-accelerator normalization. Frontier node prices now also published per accelerator (frontier_per_accelerator_value), commensurable with the per-card legacy GPU component. Rack-scale SKUs (GB200 NVL72) excluded.
• Price-primitive disclosure (SWCI + SWII). Every component tagged with the kind of price it carries (ebay_transacted / oem_list / channel_list / cloud_quoted / stress_proxy). A list price is not a transacted price; the data now says which is which.
• Condition-consistency (SWCI legacy GPU). The secondary GPU basket now assesses a consistent used-working spec, excluding premium (New/Refurb/Open Box) and parts/broken tiers. Measured effect −2% to +10% per SKU, direction varying — the prior mix was composition noise. Disclosed as gpu_condition_basis.
• Source-tier relabel. frontier_gpu and nvswitch_oem relabeled tier_1_oem_direct → tier_2_integrator_channel_survey — OEM-direct frontier pricing is quote-only; what we survey is the integrator channel, and the label now says so.
• SWII vendor-concentration guard + interconnect survey. New vendor_concentration_warning (fires at ≥40% single-vendor weight); correctly flagged FS.com at 60% at introduction. Three independent vendors added — ColfaxDirect (NVIDIA/Mellanox InfiniBand NDR/HDR/XDR at real part numbers), QSFPTEK, 10Gtek. Survey status is per-component and data-driven (≥2 vendors AND no vendor >75% of priced comps). All three former FS-only components now qualify; FS.com single-vendor exposure 60% → 0%. Per-component vendor shares published.
• Open data. Per-index history CSVs (provenance-labeled rows), machine-readable methodology.json, frozen monthly archives, /cite formats, and stable issue permalinks (/v/I/{issue}).

v0.5 — 2026-06-02

Comp integrity + multi-year history

• GPU comp composition cleanup. Secondary-market GPU ingest now excludes multi-unit lots, parts/non-working listings, and datacenter cards priced below a physical-implausibility floor. A100 80GB PCIe intra-window dispersion fell 8.55x → 4.29x — the median now tracks single working cards, not listing mix.
• Multi-year sold-comp backfill. First-party transaction records extend GPU comp history to mid-2023 for key SKUs (A100 80GB PCIe: 5 months/19 comps → 25 months/74 comps). Backfilled rows are provenance-tagged and pass the same cleanup.
• Cluster Valuator dataset. Per-SKU secondary medians + empirically measured vintage decay curves (A100 80GB PCIe ≈ −54%/yr, V100 32GB PCIe ≈ −36%/yr, H100 ≈ flat), cited condition/liquidation anchors, confidence bands, coverage-tier flags. Condition factors are cited anchors — eBay condition tags proved too noisy to derive them, and the methodology says so.
• Depreciation cross-validation dataset. Disclosed accounting useful lives from 13 SEC filers vs SpotWire-measured secondary decay, with useful-life-change events and filing URLs as receipts.

v0.4 — 2026-05-29

• Component weights for SWPI and SWFI revised to research-derived values with sensitivity analysis (RT-7)
• Per-component n_comps (volume) now exposed in indices.json and on the public site (RT-11)
• EDGAR text-mining lead-time regex widened — SWPI flips to LIVE (FU-1)
• SEC EDGAR access refactored to use marketkit shared library

v0.3 — 2026-05-29

• SWFI composite goes LIVE — first composite to drop the Indicative badge (P3-T8)
• SWPI upgraded from spot-power-only to full diffusion (4/5 components active, P3-T5)
• BLS PPI scraper covering 9 series across SWPI and SWFI cohorts (P3-T4, P3-T6)
• EDGAR mining for SWFI cohort: Vertiv, Trane, JCI, Rockwell, Honeywell, Dover, Emerson (P3-T7)

v0.2 — 2026-05-28

• SWCI compositor lands — first composite computed from real SpotWire data-engine comps
• Historical YoY backfill for SWCI basket: 71 source-attributed monthly points across 17 SKUs (P1-T7)
• EDGAR mining for SWPI cohort: Quanta, MasTec, GE Vernova, Eaton, Hubbell (P3-T3)
• EIA-860M, LBNL Queued Up, FS.com, gridstatus — five new data sources powering SWPI (P3-T1, P3-T2, P2-T1, P2-T4)
• SSD pipeline unstranded; DDR5/DDR4 RAM scraper added to SWCI basket (P1-T1+T2)

v0.1 — 2026-05-27

• Initial four flagship composite indices defined: SWCI, SWII, SWPI, SWFI
• All four start as v1 hand-curated mock values, labeled Indicative