Self-improvement

Human labels are the binding constraint on everything in this standard. Not compute, not model capability, not corpus size. A loop that improves a judge is therefore an allocation policy for a scarce resource, and most of what can go wrong is a spending mistake rather than a modelling one.

The loop

1. Score a population under the current rubric version, recording abstentions and their causes.

2. Select candidates for labeling. Not at random, which is the point of the loop, and not by the judge’s stated confidence, for the reason below.

3. Label them, independently, under a fixed rubric version.

4. Improve the instrument. Adjust the rubric, the evidence given to it, or the fold. This is a rubric change with everything that implies.

5. Re-score, and measure against a set that was reserved before any of the above.

Step 5 is the one that fails silently, because by the time anyone reaches it the labels that would have supported it have usually been spent in step 3.

The rule that governs the whole loop

A label spent improving the instrument cannot also measure it (Dwork et al. 2015). Once a label has informed a rubric edit, a threshold, or a model choice, a statistic computed against it answers “how well does this reproduce what shaped it”, which is a question about consistency rather than accuracy.

So the label budget partitions, and the partition is permanent:

$$L = L_{\text{improve}} \;\sqcup\; L_{\text{measure}}, \qquad L_{\text{improve}} \cap L_{\text{measure}} = \varnothing$$

Membership is decided before a label is looked at, and it cannot be revised afterwards. Moving a label from the measurement side to the improvement side is ordinary. Moving one the other way is not a reallocation, it is a claim that nobody remembers seeing it.

The uncomfortable consequence: an active-labeling loop is a machine for consuming the resource that validation needs. Run it without a reservation and it will eat the holdout, one useful case at a time, while every iteration looks like progress.

Sample on disagreement, not on confidence

The obvious selection rule is to label where the judge is least confident. On this standard that rule is unsound, and specifically so.

A single judge’s stated confidence is a word mapped to a constant: high, medium and low become fixed numbers. It is not a posterior. Verbalized confidence can carry some signal about correctness, and has been shown to on tasks a model was trained to report it for (Lin, Hilton and Evans 2022), so the objection is not that it is noise. It is that the relationship is a property of a particular model on a particular task, established by measurement and not by assumption, and none has been measured here. Sampling by an unmeasured confidence selects cases where the model chose a particular adjective, which is an assumption dressed as a signal.

Panel disagreement is different. It is measured: members split, and the split is a fact about the instrument rather than a self-report. Better still, a panel that abstains on ties concentrates its abstentions exactly where runs are ambiguous, so the abstention set is already a candidate pool built by the scoring process rather than by a heuristic.

Selection signal	Basis	Use
Panel disagreement, abstention cause	Measured	Preferred
Proximity to a decision threshold	Measured, where a continuous score exists	Sound
Stated confidence of a single judge	Stipulated constant	Avoid until it is shown to predict correctness
Random	Measured by construction	Required as a control, see below

The loop must carry a random arm

If the judge selects what gets labeled, and those labels reshape the judge, then the judge sets its own curriculum. Two failures follow, and neither is visible from inside the loop.

Selection compounds. Each iteration draws from the region the previous iteration found hard, so coverage of the easy region silently decays. The instrument can improve on the sampled region while degrading elsewhere, and nothing in the loop looks at elsewhere.

Actively-sampled statistics do not estimate population quantities (Beygelzimer, Dasgupta and Langford 2009). An agreement figure computed on a deliberately hard pool is not the agreement rate on the population, and reporting it as one overstates difficulty. It is a conditional number and needs saying so.

The remedy is a second, randomly-sampled pool, labeled at whatever rate the budget allows. It is the only arm that supports a population claim, and it is the first thing cut when labels are short, which is exactly backwards.

Every iteration is a version boundary

Improving the instrument changes what it measures. Under the versioning discipline this standard assumes, a change that moves a decision is a major version, and a major version voids the calibration measured against the previous one.

So each turn of the loop ends the comparability of the turn before it. Scores from iteration $k$ and iteration $k+1$ are not on one scale, and a time series across iterations is not a time series of one quantity. A loop that reports improvement across its own iterations is comparing two instruments and attributing the difference to the agent.

The cheap discipline is a frozen regression set, re-scored at every iteration, whose purpose is not accuracy but detecting that the instrument moved and by how much. It needs no human labels at all, because it compares two versions of the instrument against each other rather than against truth.

When to stop

The loop terminates when the instrument reaches the ceiling imposed by its labels, and that ceiling cannot be observed without independent labels on the same items. Until a human ceiling exists, the loop has no defined stopping condition: agreement can keep rising toward a bound nobody has measured, and there is no way to tell approaching the bound from approaching perfection.

That is why the validation protocol puts the ceiling first. It is not a nicety that improves reporting. Without it, an improvement loop cannot know it is finished, and will keep spending the scarcest resource in the system on gains it cannot size.

Named gaps

No reservation exists, so the partition above is currently notional. Every labeled set in use was either used for tuning or made visible before first measurement. The rule this page states is therefore a rule the system does not yet keep, and stating it does not close it.

The improvement step is unspecified. This page says the instrument is adjusted and does not say what may be adjusted, in what order, or with what evidence a change is kept. Rubric wording, evidence selection, fold policy, and model choice are all levers with different blast radii, and treating them as one step hides that a model swap and a wording edit are not comparable interventions.

Nothing measures the loop’s own effect on coverage. The selection-compounding failure above is argued and not instrumented. Detecting it needs the population’s easy region sampled deliberately, which is the random arm, which is the thing most likely to be skipped.

References

Work	Bears on
Beygelzimer, A., Dasgupta, S. and Langford, J. (2009). Importance weighted active learning. Proceedings of the 26th International Conference on Machine Learning, 49-56. DOI	An actively-sampled statistic does not estimate a population quantity without reweighting
Dwork, C., Feldman, V., Hardt, M., Pitassi, T., Reingold, O. and Roth, A. (2015). The reusable holdout: preserving validity in adaptive data analysis. Science 349(6248), 636-638. DOI	A set that has informed a choice no longer measures what it was reserved to measure
Lin, S., Hilton, J. and Evans, O. (2022). Teaching models to express their uncertainty in words. Transactions on Machine Learning Research. ISSN 2835-8856. OpenReview	Verbalized confidence can track correctness where it has been measured to