Lower min_n to 1 and move the caution into the apply step (capitals get a running form)

[MarkusNeusinger]

Decision (2026-08-03, author)

min_n goes to 1 — better once than never. A key observed a single time should be visible as a statistic; the plates are what they are, and withholding the only evidence there is helps nobody.

The concern the original issue raised stays valid, but it is answerable now, because it was really two questions wearing one number:

• Seeing a median is measurement — nothing renders, so there is no reason to hide an n=1 row.
• Writing a median is rendering — and since Give apply-laufform a surface: see the difference, then confirm #277 that is its own deliberate, previewed, confirmed step (LaufformApplyDialog), which already lists n_instances per glyph before anything is written.

So the gate splits: min_n = 1 at the aggregate layer, and the caution moves to the apply, where a human is already looking at the numbers and pressing a button. That is the same doctrine the rest of the pipeline follows — measuring layers show everything, the rendering step is deliberate.

What to implement

1. min_n default 1 for POST /hands/{id}/aggregates/rebuild (api/routers/aggregates.py), matching the pair rebuild which already defaults to 1 for the same "pairs are sparse" reason. The gate itself stays in core/aggregate.py::aggregate_instances and keeps counting skipped["below_min_n"] for whatever threshold is passed.
2. Low-n is flagged where it can do harm — in the apply dialog's preview table, not in the statistics. A row with n below a small threshold (2? 4?) gets a visible marker, so "this running form is one occurrence's idiosyncrasy" is stated at the moment of the decision rather than buried in a column.
3. MAD is undefined at n=1 — the letter statistics block and the sketch must not print a spread they do not have (the codebase already has this rule: "an absent MAD prints no ± clause"). Check LensStats handles n=1 rows without drawing a zero-radius circle that reads as "no spread".

Still open: how far a small-n median should travel

Lowering the gate makes the keys visible; it does not answer how much a 1–3 occurrence median should be trusted when it is applied. Option (b) from the original proposal remains the interesting one and is now independent of the gate:

shrinkage toward the chart form — blend median and chart anchors by n/(n+k), so small-n keys move partially and never wildly; fits the H1 doctrine that variant 100 is a derivation, and touches only core/aggregate.py plus the apply step.

With min_n = 1 plus shrinkage, a capital seen once moves a little toward its observed form instead of either not existing (today) or jumping fully onto a single sample (naive n=1). Measure with the wordbench guard; metric and fixtures stay frozen, headline same-hand.

Options (a) per-key override and (c) accept-and-document are superseded: (a) is the special-casing H3 wants to shrink, and (c) is no longer needed once the keys exist at all.

---

Original context

POST /hands/{id}/aggregates/rebuild defaults to min_n=4 (api/routers/aggregates.py:93); the gate itself lives in core/aggregate.py::aggregate_instances (line 106, counter skipped["below_min_n"]). Only a key that reaches four clean occurrences gets an aggregate — and only an aggregate can feed apply-laufform.

The Abb. 19/20 plates give n = 4–39 per letter, most ≤ 20, and nearly every capital stays below 4 (handmodell-stufenplan.md §4, H5). Those keys never get an aggregate, hence never a derived Laufform, hence render chart-true while their neighbours in the same fluent run use the widened running form (core/compose.py ~line 1421, run ≥ ASCENDER_LEAN_MIN_RUN). That is a systematic rhythm break in exactly the words that carry a capital — the word start, where it is most visible.

Size & dependencies

Small now that the decision is made: a default change plus a chip in the apply dialog. The shrinkage question above is a separate, medium follow-up. Related: #270 (the apply surface this leans on, shipped in #277), #278 (a word-level chain fit would raise the clean-fit yield and push more keys over any threshold).

[MarkusNeusinger]

Evidence from a read-only simulation over the stored 218 instances (no write, no rebuild — core/aggregate.py::aggregate_instances run locally on the public instances read; sanity check first: min_n=4 reproduces the 13 stored aggregates with identical n and bit-identical medians).

min_n=3 would add exactly six keys — including the two the pair plates can never supply

key	n	distinct specimens	fit RMSE ⌀/max px	MAD median	disp (n-fair)	positions
b	3	3	1.064 / 1.153	0.0079	in band	medial ×3
longs	3	3	0.545 / 0.771	0.0020	cleanest	initial ×3
o	3	3	1.128 / 1.211	0.0080	borderline (one anchor MAD-max 0.0856)	medial ×3
S	3	3	1.011 / 1.351	0.0058	in band	initial ×3
z	3	3	1.051 / 1.205	0.0095	in band	initial ×3
Z	3	3	1.201 / 1.270	0.0083	in band	initial ×3

• All six: three different specimens (no repeated word), anchor count matches the chart row (120), so the apply step would see clean creations (created: true) — no skip condition fires.
• n-fair dispersion control (because MAD at n=3 is constructionally optimistic): mean radial distance from the median 0.0070–0.0157 and max pairwise 0.0175–0.0376 for the six, vs. 0.0143–0.0258 / 0.0337–0.0901 for the established 13, and vs. 0.0058–0.0113 for all 3-subsamples of the established keys. The six sit in the lower third of the established band. Tightest: longs; borderline: o.
• Notably, S and Z do not appear on the Abb.-20 pair plates at all — for them the threshold (or better word-plate yield) is the only path to statistics.

Recommendation

min_n=3 as the rebuild default for this hand is statistically defensible for these six. Two caveats belong in the decision: the MAD column at n=3 is not comparable with the n≥8 rows (lean on disp/maxpair instead), and each of the six groups is position-pure (b/o only medial, the rest only initial) — it describes one writing position. Since all six would be new Laufform rows (six letters switching from chart form to a measured running form = a rendering change), the apply should stay a per-letter decision through the preview dialog regardless of the threshold.

Full tables and scripts: session scratchpad min_n3_simulation.md (Stage-B step 0 of #278).

---

Generated by Claude Code