sweep every checker, and say what the step actually proves (Refs #2161)

[gHashTag]

tri gate-sweep selected files matching check_* or *gate* — so it measured the naming convention. Thirty checkers live in tools/; thirteen had never been swept, including every verify_*, the fuzzer and the generators. It now takes every non-private .py and lets the reader judge what belongs — the same call this campaign made for gate selection: choose by property, not by what someone called the file.

CRASH: 0 across all thirty.

The thirteen contained five correct skips (all fatal under --require — checked, because a skip that is not is a pass wearing another word), one generator raising FileNotFoundError on the catalog it reads, and one claim.

The claim

The workflow ran a step called:

Prove the trainer LEARNS (XOR 4/4 + nonlinear held-out >=90%, incl. deep 3-layer)

The tool is pure Python. No t27c, no .t27 spec, no iverilog. It generates backprop microcode, runs it on a bit-faithful GF-T interpreter, and asserts on the text of the Verilog it emits. It passes unchanged in a directory containing nothing but itself — which is how the scope was noticed, and is not something reading the step could reveal.

The tool's own docstring was honest and precise. The overclaim lived entirely in the step name — and the step name is the claim most people read. Nobody opens the tool; they read a green check and a sentence, and the sentence said the trainer learns, which a reader takes for the compiler, the RTL, or the board.

Renamed to what it proves, with the reason in a comment above it, and a what this does not establish block added to the tool naming the three claims it is not: that the compiler produces this microcode, that the emitted Verilog simulates, that any of it works on silicon.

An honest tool under an overclaiming name is the most durable kind of wrong number, because every individual artefact is accurate. §90 and §91.

Refs #2161

[github-actions]

📓 NotebookLM Notebook linked to this PR

• Notebook ID: b83263109fb055dc
• URL: https://notebooklm.google.com/notebook/b83263109fb055dc

This notebook contains session context, decisions, and artifacts for this work.

[github-actions]

PR Dashboard

Generated at: 2026-08-23 19:58:28 UTC

• sweep every checker, and say what the step actually proves (Refs #2161) #2598 sweep every checker, and say what the step actually proves (Refs #2161) | Created: 2026-08-23 | Updated: 2026-08-23 | 9 failing / 30 total
• Unbreak FPGA CI and correct two unbacked hardware claims #2081 Unbreak FPGA CI and correct two unbacked hardware claims | Created: 2026-08-11 | Updated: 2026-08-20 | 5 failing / 27 total

Summary

Status	Count
Total Open PRs	2
PRs with Failing Checks	1
PRs with All Checks Green	1
READY	0
FAILING	1
PENDING	0

Seal Status

• ⚠️ STALE -- sha256(compiler.rs)=76fdc57c6caa != manifest seal=87e5cbd3ad94.
  The committed NMSE numbers were certified against an older compiler.rs.
  Run scripts/reseal-check.sh locally for the two-step reseal command (advisory; not a merge gate).