Skip to content

fix(gates): the last gate gets a control — 21 gates, 0 uncovered (#2492) - #2528

Merged
gHashTag merged 2 commits into
masterfrom
fix/backprop-control
Aug 23, 2026
Merged

gHashTag merged 2 commits into
masterfrom
fix/backprop-control

Conversation

@gHashTag

Copy link
Copy Markdown
Collaborator

21 gates, 0 with no control in any form. The count that opened this campaign at 4 of 12 is at zero, across a set two-thirds larger and selected by property rather than by name.

The last one

gft_backprop_microcode.py carried sixteen asserts — XOR trains to 4/4, held-out clears 90%, the emitted Verilog carries the ports it claims — and nothing showed that any of them could go red. Sixteen assertions and no evidence they assert anything.

Three planted cases now:

  • a sign flip in the shared multiplier stops XOR converging → assert acc == 4 fires
  • a renamed port trips the emitter assertion
  • the clean tree stays green, so the two above cannot pass for free

Each demands the assertion's own message, because an assert delivers its verdict through a traceback: exit 1 means both "XOR did not train" and "somebody typed a name wrong".

Building it corrected #2524's null result

That PR measured "do any CI-invoked tools fail ONLY through assert?" — answer zero, no gate is invisible to the selector — and I reported it as if it settled the matter. It does not.

The mutation operators score this gate 0/0 silent, 0/0 loud, 0/0 invert. Its verdicts are asserts, which no operator recognises, so every verdict in this gate is invisible to three of the four questions. The selector sees the file (via a ternary returning a class label — #2524's accident); the operators see none of its sixteen assertions.

Two different questions — "is this file classified as a gate?" and "can its verdicts be broken and noticed?" — and I answered the easier one, found nothing, and moved on.

The documented trap, reproduced

The port-rename plant spelled its needle literally, so the first occurrence of that string in the file became the control's own source line, and str.replace(.., 1) edited the harness instead of the target. The case then reported the gate as blind when nothing had been planted at all.

check_duplicate_agreement.py carries a comment warning about exactly this, written after it happened there. I had read that comment. The fix is to assemble the needle — "input [31:0] x" + "0i" — so it exists as no literal anywhere.

Named and left

The boundary column reads 5/31. Those sites are arithmetic internals — encodings, magnitude comparisons — where moving a comparison is a numerical change rather than a verdict change. A different kind of surface, larger than anything else outstanding, and not this campaign's question.

Refs #2492

@github-actions

Copy link
Copy Markdown
Contributor

📓 NotebookLM Notebook linked to this PR

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

@github-actions

Copy link
Copy Markdown
Contributor

PR Dashboard

Generated at: 2026-08-23 10:45:29 UTC

Summary

Status Count
Total Open PRs 7
PRs with Failing Checks 3
PRs with All Checks Green 4
READY 3
FAILING 3
PENDING 0

Seal Status

  • ⚠️ STALE -- sha256(compiler.rs)=e905ba8eeb53 != 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).

@gHashTag
gHashTag merged commit bcec69f into master Aug 23, 2026
24 of 26 checks passed
@gHashTag
gHashTag deleted the fix/backprop-control branch August 23, 2026 10:50
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant