/status
This page is generated at build time. Nothing on it is live, nothing polls, and there is no
API behind it: it is a rendering of data/export/site.json, which is itself a pure
function of the archive. Two builds over the same ledger produce the same page.
- archive as of
- 2026-10-18T13:38:40Z
- source schema
0007_benchmark_superseded_index.sql- payload schema
- 2
- generator
scrollback/cmd/export
max(price_fact.valid_from, price_fact.valid_to, event.observed_at, model.last_seen, model_alias.last_seen) — the latest moment this archive READ a catalogue, over every layer that records one. event.occurred_at and availability.available_to are excluded: both are announced EFFECTIVE dates, they run months ahead of the page announcing them, and an as-of stamp in the future claims a reading that never happened.
no wall clock enters this file: every timestamp comes out of the archive, so two runs over the same ledger produce byte-identical files (`make export-verify`).
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.
the archive, by table
| table | rows |
|---|---|
model | 312 |
model_alias | 7440 |
offering | 3136 |
price_fact | 13105 |
benchmark | 13 |
benchmark_version | 13 |
benchmark_result | 67158 |
availability | 2606 |
event | 11094 |
suspect | 33 |
Not exported, ever: the observation ledger, the source snapshots and the fetch log. They are evidence, not product.
suspect rows
33 of 13105 price rows (0.25%) are marked: 17 by magnitude, 9 by rounding residue, 10 by curated incident.
Overlaps: 3 on magnitude and residue, 0 on magnitude and curation, 0 on residue and curation, 0 on all three.
33 == 17 + 9 + 10 − 3 − 0 − 0 + 0 — identity holds.
The rules, the reasons and the three incidents are on
/methodology.
the pools behind the headline
A pool is a set of benchmarks whose curves are pooled into one band. All three are published because the choice between them is an argument, and printing only the chosen one would hide it.
baseline_2026_08_15 — the headline pool
the 2026-08-15 set (ALE·WeirdML·ARC×2·Proof·DeepSWE) — the same six benchmarks the 2026-08-15 ad-hoc measurement pooled, so the comparison against it is about the pipeline rather than about a different mix
- benchmarks
- 6
- curves
- 198
- median
- 66 d
- IQR
- 49–90 d
- p10–p90
- 35–160 d
- full band
- 20–400 d
- per year
- ≈46×/yr at the median (IQR band 17×–179×)
- identity
- ale-bench 27 + arc-agi 37 + arc-agi-2 41 + deepswe 16 + proofbench 23 + weirdml 54 = 198 — holds
gate_no_arc_mirrors
everything past the gate, ARC mirrors excluded (§13-S20 open) — ARC-AGI and ARC-AGI-2 reach this archive only through Epoch's mirror; whether they may be pooled is S20, and it is open
- benchmarks
- 6
- curves
- 136
- median
- 62 d
- IQR
- 38–74 d
- p10–p90
- 22–109 d
- full band
- 17–400 d
- per year
- ≈60×/yr at the median (IQR band 31×–795×)
- identity
- ale-bench 27 + cursorbench 15 + deepswe 16 + proofbench 23 + the-agent-company 1 + weirdml 54 = 136 — holds
gate_all
everything past the gate, ARC mirrors included — the widest pool, printed beside the other two so the mirrors' effect on the band is visible rather than argued
- benchmarks
- 8
- curves
- 214
- median
- 63 d
- IQR
- 44–90 d
- p10–p90
- 24–143 d
- full band
- 17–400 d
- per year
- ≈57×/yr at the median (IQR band 17×–313×)
- identity
- ale-bench 27 + arc-agi 37 + arc-agi-2 41 + cursorbench 15 + deepswe 16 + proofbench 23 + the-agent-company 1 + weirdml 54 = 214 — holds
a pool's curve count is the sum of its benchmarks' curve counts; nothing else enters a pooled median.
the gate and the stale frontier
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.
1 of 214 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
S29: the availability filter stays out of the formulas; the surface MARKS a curve whose last step is held by a configuration that has left the market.