New readers of FlashProve should start at the project's README and the short pages under guide/. This folder holds the long pages. On the pages of limits (limits.md, limits-further.md, limits-all-rows.tsv, known-errors.md), unless an entry says otherwise, its results were not seen on the released programs, and a fault of the code, told in the present tense, holds of the released programs (what was run on these programs, and what was not). Each page's head tells its rule in full; the head of limits-all-rows.tsv has two exceptions. The other pages mark such a result, with † or in words. figures.tsv has a rule of its own, told below.
Pages cite rows of the table of limits by id. To read one, search for its id at the start of a line:
grep -P '^SO-15\.7\t' docs/limits-all-rows.tsv
- Using the programs: quickstart.md, verifying.md, operators-duties.md, service.md, faults.md.
- Limits and results: limits.md, limits-further.md, limits-all-rows.tsv, known-errors.md, what-is-not-hidden.md, how-hiding-is-checked.md, results.md, figures.tsv.
- Reference: architecture.md, protocol.md, model.md, format/, glossary.md, sources-notes.tsv.
- The fork's research: research-roadmap.md, research-evidence.md, and qwen-port.md, the status of the Qwen3.8-27B port (no part of that model can be proved end to end today).
figures.tsv has most of the figures that these pages print, and others, one row each.
An empty programs cell means the timed runs, on the released programs; any other cell says what
ran instead, and no time, rate or memory peak carries over from such a row. The rows of the runs
that timed the serving scripts' check stand apart: their runs were on the released programs, one of
them with one constant of the scripts set to a value that is not this repository's, and are none of
the timed runs. Rates are cut, not rounded, their means too; in the rows of the timed runs a memory
peak, disk in use and a largest size are rounded up, the lower end of a range of them cut, and
times, and the means of times and sizes, are rounded to the nearest. In these rows a standard
deviation (sd), a percentage and a multiple against stock vLLM are rounded to the nearest too. 1 GB
is 10^9 bytes, 1 MB 10^6, 1 MiB 2^20 and 1 GiB 2^30; the "80 GB" in the GPU's name is its maker's
figure. An id tells nothing of the weights: read the column weights. The timed runs measure speed
and are no test of weight hiding. "Other programs" and "other sources" in a cell:
docs/glossary.md.