Every 7 days the grams on this network work the model catalog in cohorts — each family's lead member at rung 0 on every engine first, then its quantization ladder, then the other members, then derivatives — and publish what they proved, per hardware class, with how many machines contributed. No number here is from a model card. Publishers are listed by what the network proved of their builds, never by popularity.
2026-W41
Open now. Grams take the least-covered artifact for their own hardware class first, so two machines of one kind never measure the same thing while a third kind waits.
code.reason, code.write, faithful_tool_result, instruct.follow, lang.bn, lang.en, lang.gu, lang.hi, lang.kn, lang.ml, lang.mr, lang.pa, lang.ta, lang.te …
per engine: how many artifacts this season proved, on how many hardware classes, which capabilities came back, and how many artifact-slots the fleet currently cannot reach at all (no Apple Silicon for MLX, no schedulable card for vLLM, no cluster for the speech pods). An engine with zero proved and a large unreachable count has never been shown to work on this network -- a statement about the network, not about the models.
away 183 h · last session ended without being asked to — power loss, a hard reset, or a crash · a laptop on battery — most likely asleep
expects 27 h for the rest (own samples) · blocked: budget
away 19 h · last session ended without being asked to — power loss, a hard reset, or a crash · this machine sleeps (60 min during its last session) and is on battery — most likely asleep
expects 61 h for the rest (own samples) · trailing rate says 634 h
away 501 h · last session ended without being asked to — power loss, a hard reset, or a crash
expects 39 h for the rest (own samples) · blocked: engine
away 23 h · last session ended without being asked to — power loss, a hard reset, or a crash · this machine sleeps (37 min during its last session) and is on battery — most likely asleep
expects 53 h for the rest (own samples) · trailing rate says 1500 h · blocked: fetch
away 40 h · last session a supervisor bounce: a new session started while the last one was alive · this machine sleeps (27 min during its last session) — most likely asleep
expects 9 h for the rest (own samples) · blocked: disk
away 160 h · last session a supervisor bounce: a new session started while the last one was alive · this machine sleeps (979 min during its last session) — most likely asleep
expects 6 h for the rest (own samples) · trailing rate says 4056 h · blocked: engine
away 407 h · last session a supervisor bounce: a new session started while the last one was alive
expects 29 h for the rest (own samples) · blocked: engine
no priced estimate yet · blocked: pace
away 40 h · last session a supervisor bounce: a new session started while the last one was alive · a laptop on battery — most likely asleep
no priced estimate yet · blocked: pace
away 25 h · last session ended without being asked to — power loss, a hard reset, or a crash
expects 7 h for the rest (own samples) · blocked: engine
away 1 h · last session ended without being asked to — power loss, a hard reset, or a crash
expects 2 h for the rest (own samples) · blocked: pace
one row per gram: what it is, what its hardware can never reach, how far it has got, and its OWN priced estimate for the rest. There is no single fleet ETA, because a gram only proves what it can obtain.
8
cells signed as gated, unloadable, or not published upstream. Each one is a measurement every other machine of that class is spared.
Medians from the grams own clocks, carried across windows so a class whose machines are asleep still plans on what the network paid to learn.
Tokens actually served, cumulative. Demand decides how deep a proof goes and where weights sit — never what exists.
Past seasons
The season files are public JSON, built by an open aggregator from signed results anyone can re-verify — data/catalog/seasons on GitHub. Cite them; they carry the signatures. Back to the evidence table →