[us-commits] [ehb54/ultrascan3] e2a880: perceiver: make the CI regression gate slack, not ...

emre brookes noreply at github.com
Tue Aug 18 10:50:23 MDT 2026


  Branch: refs/heads/ehb54-issue-1016-somo-merge-2
  Home:   https://github.com/ehb54/ultrascan3
  Commit: e2a880b8512de6620852466ae6a617ba629f8102
      https://github.com/ehb54/ultrascan3/commit/e2a880b8512de6620852466ae6a617ba629f8102
  Author: ehb54 <brookes at uthscsa.edu>
  Date:   2026-08-18 (Tue, 18 Aug 2026)

  Changed paths:
    M .github/workflows/somo-module-tests.yml
    M us_somo/develop/perceiver/Makefile
    M us_somo/develop/perceiver/tests/regression.cpp

  Log Message:
  -----------
  perceiver: make the CI regression gate slack, not pinned

The gate added earlier in this branch would have broken CI for the perceiver
work happening in parallel. It required --min-atoms 41219, which is EXACTLY
the current count -- one atom's difference fails -- and --min-geometric 99.8
against an actual 99.833, leaving 0.033 of room. Any legitimate change to the
module trips both.

The atom floor was there to catch a fixture going missing. That is worth
keeping, but it is the wrong instrument: the atom count moves with parsing
and with which atoms are scorable, i.e. with exactly the work in flight.

Replaced with --require-all-inputs, which asserts that every fixture handed
in produced something to score. It catches the failure that matters -- a
fixture missing, unreadable or truncated, which would otherwise leave the
percentages looking healthy while measuring less -- and it is orthogonal to
perception quality, so tuning the perceiver cannot trip it. The failing
inputs are named.

--min-geometric drops to 95, a collapse detector rather than a pin, leaving
about five points of room below the current 99.833%.

What this keeps is the thing that was actually missing: the regression RUNS
in CI and its numbers are in the log on every build. What it gives up is
catching a small drift, which is not something to enforce on a module still
being developed. Tighten once it settles.

Verified: passes with the fixture set; fails, naming the file, when an input
is missing and when one is empty; the vacuous-pass guard still fails on no
input at all.

Co-Authored-By: Claude Opus 5 <noreply at anthropic.com>



To unsubscribe from these emails, change your notification settings at https://github.com/ehb54/ultrascan3/settings/notifications


More information about the us-commits mailing list