Finding: vendor probabilities arrive on a 2-decimal grid, bounding achievable ECE resolution

[TMHSDigital]

This is a recorded finding, not a defect. The vendor is entitled to round
what it puts on the wire, and two decimals is a reasonable place to round for
a decision API whose output is meant to be thresholded. This documents a
bound on what can be measured through that API.

What was measured

On a live run of 40 choice rows against jev-1.13.0 on 2026-09-21, all 165
probability values landed exactly on a multiple of 0.01, as did all 40
confidence values. The distribution is quantized to two decimals on the wire.

What it bounds

METHODOLOGY records this under "Probabilities arrive quantized, which bounds
the resolution of any figure here". The four consequences:

• Binning. The default ECE scheme is ten equal-width bins, which are wider
  than the 0.01 grid, so quantization does not bias the binning. A scheme with
  more than 100 bins would be measuring the grid rather than the model.
• Thresholds. A cascade threshold cannot be tuned more finely than 0.01. A
  sweep reporting a cut to four decimals is reporting three digits of noise.
• Ties. At 0.01 resolution over three to six options, exact ties for the
  maximum are ordinary rather than freak events. See the ties issue.
• Recalibration. A fitted temperature is applied to quantized inputs, so
  the corrected probabilities are smooth but carry no more information than
  what arrived.

What remains open

• Is the achievable ECE floor higher than the sample-size floor already
  computed? The nulls currently model binning noise and sample size. They do
  not model input quantization. If quantization raises the floor, every ECE
  figure through this transport is being read against a floor that is slightly
  too low, which would make models look marginally worse calibrated than they
  are. Probably a small effect at ten bins, but it is unquantified.
• Should the report state the observed grid? It could detect the
  quantization from the returned values and say so, rather than relying on a
  reader having found this in METHODOLOGY. That is a small change and might be
  the whole remedy.
• Does it vary by model or by option count? One run, one model version.

[TMHSDigital]

Two of the three open questions are answered in #98:

• Should the report state the observed grid? It does now, as a Resolution line, with a warning when the bins are narrower than the grid.
• Is the achievable ECE floor higher than the one computed? No. scripts/quantization_floor.py shows a calibrated model reported on a 0.01 or 0.05 grid clears the floor's 95th percentile at the nominal 5 percent, from 40 to 10,000 rows. METHODOLOGY has the table.

Still open: whether the grid varies by model or option count. That needs live runs against more than one model version.