/methodology

The index ranks currently purchasable configurations by task dollars per score point on one named job. Two measurements sit under that table. F1 asks what a fixed capability costs over time. F2 asks how often the cheapest way to buy that capability changes hands. Everything below is derived from the archive named on /status; nothing on this page is estimated, and where a number does not exist it says so.

F1 — the cost of a fixed capability

Hold a capability fixed by picking a score threshold on one benchmark. Walk the release-date axis and keep the running minimum cost per task among configurations that clear the threshold. Every strictly cheaper qualifying release is a new step. That staircase is the curve, and its half-life is the number.

Thresholds are not a grid of round percentages. They are exactly the distinct scores the benchmark's costed points achieved, so the scan is the whole population of curves rather than a sample — and any fixed step width would be an unmeasured constant smuggled into the headline. Adjacent thresholds usually collapse onto the same shape; the label is the most demanding of them, because every lower threshold buys at least that capability.

the gate, printed with its numbers

a staircase carries a half-life only with at least 4 steps AND at least 90 days between its first and its last; anything shorter is a curve whose whole life is briefer than the number it would claim.

A threshold that fails prints the condition it failed and the number it failed against, and its staircase is still drawn.

the headline is a band

the headline is a BAND, never a single number: this pool's curves run from 25–394 days depending on the threshold, so a lone median would be a choice presented as a measurement.

Two half-life definitions are published for every curve: the endpoint definition (first and last step, the one the band is made of) and a log-linear least-squares fit over every step. The frontier panel also prints the same curve with its first step dropped and with its last step dropped, because one revised endpoint can move the number by more than half its value.

what the corpus is, and what it excludes

a published cost of $0 is not a measured spend of zero, it is a price nobody can buy — the ':free' hazard class — and one such point would pin a threshold's frontier at $0 forever. They are excluded and counted, never plotted.

'imputed_local' is a price nobody can buy; WeirdML's local and quantized runs are invisible through the mirror and may sit on the frontier edge.

every measurer is its own channel: two measurers never share a frontier or a dominance timeline.

the x axis is the MODEL'S RELEASE DATE, not the day anyone measured it or could have bought it (§5-F1 trap 6).

purchasability

OFF by decision (S29): the filter does not enter F1, F2 or the event fold, and the surface marks the stale frontier instead.

1 of 155 valid curves (gate_all: every curve past the gate, ARC mirrors included) is held by a retired config at the corpus's now; every holder has availability data

weirdml, score ≥ 0.2318: Gemini 1.5 Flash 002 (google/gemini-1.5-flash-002), at 2026-07-31.

price_fact.valid_to IS NULL and availability.available_to IS NULL mean OPEN AT archive_as_of, not open today. A model with no availability row is UNKNOWN, not available.

cost per task of a fixed capability

• frontier step
• no cheaper configuration in the corpus since — not a measurement past that date
• last step held by a configuration that has left the market

selected ale-bench score ≥ 1162.45 → 98 d · 4 steps · 358 d span · 12.5× cheaper

F2 — how often the frontier changes hands

a KICK is the moment a configuration became dominated: something at least as strong arrived at a strictly lower cost, or something strictly stronger arrived at no higher cost. Overlapping score intervals suppress the event, and a victim is kicked once.

the window ends on the NEWEST KICK, the corpus's own now — a wall-clock anchor would make the same database print a different frequency every quiet week. Frequency is counted in EVENT-DAYS, not kicks: twelve kicks in one afternoon are one day the channel spoke.

54 of 66 configs are dominated (82%) — ale-bench, measured by Epoch AI.

dominated is not the same as bad: a dominated configuration can still be the right purchase for a reason this archive cannot see.

2 points excluded for a published cost of $0, counted over the cost F2 compares, which is per-task where the benchmark publishes one and the suite total otherwise. That is a WIDER set than F1's y axis, so this figure can exceed f1.corpus.zero_cost_excluded over the same corpus; the two are not a contradiction, they are two scopes.

saturation: the evidence, not a label

S28 (19 Aug 2026): NO saturation label is published in Phase 1a and no curve is suppressed. The evidence is printed instead, because the threshold that would separate 'saturating' from 'saturated' has not been measured, and a guessed threshold whose documented consequence is DO-NOT-PLOT would delete benchmarks from charts on the strength of a number nobody chose.

revisited at T+90 (≈2026-11-13); until then a threshold is set only if it derives from a measurement.

• aider-polyglot — not measured: no threshold produced a staircase past the gate
• ale-bench — unbounded above (a rating, not a fraction); not defined: headroom needs a ceiling, and this benchmark has none
• critpt — not measured: no threshold produced a staircase past the gate
• exploitbench — not measured: no threshold produced a staircase past the gate
• osworld-2 — not measured: no threshold produced a staircase past the gate
• surface-evolver-bench — not measured: no threshold produced a staircase past the gate

data quality: rows this archive does not stand behind

S31 (19 Aug 2026): price_fact is NOT corrected, filtered or deleted — the archive records what the catalogues actually published. Doubt is this separate, derived claim, and it is presented per INCIDENT (seller × upstream commit), never by blaming the seller whose arrival made an old broken row visible again.

31 of 6702 price rows (0.46%) carry a suspect mark. They are published rather than deleted, and every consumer of the dumps is expected to filter them before computing anything.

Identity, checked on every export so it can actually fail: union == M + R + C − MR − MC − RC + MRC, with each rule set built in its own pass over the ledger so the equation can actually fail — 31 == 15 + 9 + 10 − 3 − 0 − 0 + 0, holds.

Overlaps: 3 rows fire on both magnitude and residue, 0 on magnitude and curation, 0 on residue and curation, 0 on all three.

100% precision is a MEASURED PROPERTY of this corpus, not a guarantee. A future error with a mid-sized residue would land in the empty band between the benign artefacts and the real ones, and would want a re-measurement rather than a silent pass.

the curated incidents

An incident is a seller and one upstream commit instant — never the seller whose arrival made an old broken row visible again. A months-old wrong row prints a fresh headline every time a new, blameless seller lists the same model, so attributing the fault to the entering seller names the wrong party.

licence and attribution

The derived dataset is published under CC-BY-4.0. Attribute as: SCROLLBACK (mirc.tech), CC-BY-4.0.

CC-BY-4.0 covers this derived dataset. Upstream attribution is a condition: the measurements in benchmark_result derive entirely from Epoch AI's CC-BY-4.0 benchmark corpus (the descriptive benchmark and benchmark_version rows are this archive's own curated registry, not Epoch data); the price layer derives from the LiteLLM (MIT) and models.dev (MIT) catalogues, and NO price_fact row cites a vendor page; official vendor DEPRECATION pages back 44 availability rows and the 44 event rows derived from them. See DATA-LICENSE.

• Epoch AI, Benchmarking Hub (CC-BY-4.0) — the sole source of every measurement in benchmark_result, all 2343 rows of it. The descriptive benchmark and benchmark_version tables are this repository's own curated registry, not Epoch data; attributing them upstream would be over-attribution.
• LiteLLM model catalogue (MIT), excluding the repository's enterprise/ directory, which this project does not read.
• models.dev catalogue (MIT).
• Official vendor deprecation pages (developers.openai.com, platform.claude.com) back 44 availability rows and the 44 event rows derived from them — not the layers, which are 1554 and 3588 rows. Only dates and retirement kinds are reproduced; no page is republished, and no price row cites a vendor page.
• The WeirdML CSV carries no licence statement upstream. It is archived locally and is never redistributed; it produces no derived row. The quality axis comes from Epoch's CC-BY-4.0 corpus, including the WeirdML measurements Epoch mirrors under that licence.
• The raw archive is not part of the dataset. Archived source bodies are stored for measurement and reproducibility and are not redistributed; the evidence handles in the dumps name archive rows, not resolvable joins.
• The dumps republish upstream's own citation strings. benchmark_result.evidence_url is Epoch's citation, verbatim, on 1245 of 2343 rows. Of those, 336 point at arcprize.org and 318 point at htihle.github.io; event.source_url carries 32 of arcprize.org and 48 of htihle.github.io. None of those URLs was fetched to build this dataset: the strings travel exactly as Epoch published them, and they are citations rather than data collected from those pages.

The full text travels with the dumps as DATA-LICENSE, served beside them — CC-BY-4.0 §3(a) asks that the notice reach whoever takes the data, and a file that is only in the repository does not. Source code is separate and is MIT.