Methods · data

Every number on this site, and where it comes from.

Every chart reads a data file in the ledger. Each entry below gives the source, the cutoff date, the method, the assumptions, and how to show it is wrong.

Launch tempo

Used on: Home (the record and the tape); The idea; Ledger · Fig. 2

Source: Hugging Face public models API, pulled 2026-09-01 (22:08 UTC). ↗
Method: Lab list of 42 organisation ids in labs.json; 6,977 repositories read. Pipeline tags kept: text-generation, image-text-to-text, any-to-any. Quantizations and derivative repositories excluded (GGUF, AWQ, GPTQ, MLX, FP8, INT4 and similar). Families deduplicated by stripping size and variant suffixes; each family dated by its earliest repository creation and scored by its peak likes. Tiers: flagship 1,000+, major 500 to 999, broad 250 to 499. Window 2024-09-01 to the pull date.
Assumptions: Likes are a popularity proxy, not a capability measure. The lab list is a judgement call. ModelScope is not scanned, so labs releasing only there are invisible. Family deduplication is heuristic.
Uncertainty: Likes accumulate over time, so recent months are undercounted at any fixed threshold. Counts are floors, not estimates. Rows count releases, not evaluation campaigns; tier and modality variants of one model stay separate.
Falsification: Name a qualifying release the pull missed, or a counted family that is a derivative. The pull re-runs without credentials.

Coverage

Used on: Home (the record); The idea · Fig. 1; Ledger · Fig. 1

Source: Model cards, technical reports, release posts and independent evaluation reports, walked in the rubric’s source hierarchy.
Method: Rows are open-weight releases judged frontier-capable: listed as frontier or notable by Epoch AI, ranked near the frontier on a public index, or a flagship release from a frontier lab (the rule is recorded per row in the data file). API-only comparison rows are excluded. Every yes or partial cell carries a URL. The four labels on the site map to the rubric's public tiers: nothing found = T0, model card only = T1, developer evaluation = T2, independent evaluation = T3; the rubric also keeps an internal 0 to 5 scale. 19 of 39 rows coded as of 2026-09-01; the rest render as unreviewed, never as T0.
Assumptions: Independence ranks above a lab’s own disclosure: a lab-run CBRN evaluation is T2, an independent methodology-public evaluation is T3. "No source found" is an observation, not proof of absence.
Uncertainty: Two passes: one coder with a stated source walk per row, then an independent second reviewer who opened every T3 source URL and spot-checked T1 rows. Coding notes record every ambiguous call.
Falsification: Write to info@pacificcompute.org with the evaluation we missed and its source.

Campaign cost estimate

Used on: For funders

Source: Project planning band of $3 to $4 per H200-hour.
Method: $200,000 buys about 50,000 to 67,000 H200-hours at that band; the planning envelope for one major launch.
Assumptions: Rental pricing in Singapore in 2026; a difficult model may run well above the figure and a clean model well below it.
Uncertainty: An estimate, campaign-dependent. Unused compute rolls into the next launch.
Falsification: Published campaign rows on the ledger, once they exist, will show actual GPU-hour bands.

Cite this dataset

Pacific Compute Ledger, version 17df23f, 2026-09-01. pacificcompute.org/ledger.