Does the confident floor still earn its place, now that the arms must agree?

[jerome-queck]

Parent

Spec: Syrax v1

The question

Does CONFIDENT_FLOOR still earn its place, now that the two arms have to agree before the unit may
assert — or is agreement the whole of what should gate confident?

This is a decision rather than a task. The measurement below says the floor is not doing the job it
was kept for, and the answer might still be keep it anyway; what it must not be is left implicit.

What the set says

Scored against the one benchmark set as seeded by #164
— 27 entries, of which 7 are answered empty before confident is ever asked, leaving 20.

gate	confident	right	wrong
floor 0.12 + arms agree (today)	3	2	1
arms agree, no floor	5	4	1
floor 0.12, no agreement	6	3	3
floor 0.218 + arms agree	0	0	0

The floor excludes exactly two documents on this set, and both of them are correct answers:

+0.092  right   the book on local analysis for the odd order theorem
+0.045  right   last year's poster on strongly regular graphs

It prevents no wrong assertion at all. The one confident-and-wrong that survives —
MH1101 Calculus II examination May 2026 at +0.215 — is above the floor and above the re-fitted
one, so no threshold reaches it. Precision with the floor is 2 of 3; without it, 4 of 5. Removing it
would produce more correct assertions and a higher share of them correct, on this set.

Meanwhile the agreement condition is doing the work the floor was kept for: ADR-0025 measured it
taking confident-and-wrong from four to one with the floor sitting still, and the table above says
the floor alone gives three right and three wrong where agreement alone gives four and one.

The re-fitted number is a warning, not a recommendation

report.py states the floor would be 0.218 against this set — best wrong answer +0.2146,
worst right answer −0.1169. Applying it takes confident to 0 of 27: it would delete the
verdict, not tighten it, including both cases where it is currently right.

That spread is #34's original finding reproduced
at ten times the corpus, and worse. ADR-0004 recorded correct answers at −0.172 to 0.157 against
wrong ones at 0.002 to 0.118; today it is correct down to −0.117 against wrong up to +0.215. "The
absolute floor cannot separate a right answer from a wrong one" was true of the trial's folder and
is more true now.

What has to be weighed

• Twenty queries is not many, and three of them decide the table. A conclusion this clean off a
  set this small is exactly what The ranking loses a document that matches every term #151 was written to distrust.
• The floor's original job included one this set may not exercise: a query whose fused top is a
  document nothing much matched. Agreement does not obviously catch that, and no entry here is one.
• The floors are not one decision. EMPTY_FLOOR separates cleanly and was re-fitted eight days
  ago (ADR-0025); nothing here touches it. This is about CONFIDENT_FLOOR alone.
• Removing a threshold is not the only outcome. Keeping it and recording why — that it is a
  volume control on assertions rather than a precision filter, and that this is deliberate — is a
  decision too, and a better record than a number nobody has re-read.
• Whatever is decided, ADR-0007's line holds: this issue is the report, and the change is a
  person's act with a pull request behind it.

What it produces

A decision record amending ADR-0004
and ADR-0025, and the
code change if the answer is that the floor goes. The numbers above are the measurement it is
argued against; re-run them first, because the set grows.

Assisted-by: Claude Opus 5
Co-authored-by: Claude Opus 5 noreply@anthropic.com