T1c — the FR fragment board: the house number is a licence, not a hint

Deliverable (2026-07-16 review, Tier 1c): replace the n=63 Paris anecdote with a BAN-sampled board carrying confidence intervals per phenomenon class, and make it the grading set for the T2 shard.

Built, and its first read reframes the arc. With a house number the shipped model scores 92.5%; without one, 21.5% on the same streets. The gap is not a difficulty gradient. The model has learned that a house number licenses a street reading at all — remove it and a designator-led phrase parses as a locality, designator included.


Board	`mailwoman eval fragment-board` (`eval-harness/fragment-board.ts`)
Fixture	`fixtures/ban-fragments-fr.jsonl` — 2,800 rows, 7 classes, BAN (Tier A)
Builder	`scratchpad/build-ban-fragment-board.py` (seeded, deterministic)
Model	v264 (v6.3.0, shipped), production config

1. The board

class	n	rate	95% Wilson
street-housenumber	400	0.925	[0.895, 0.947]
alnum-housenumber	400	0.925	[0.895, 0.947]
bare-street	400	0.215	[0.178, 0.258]
street-particle	400	0.273	[0.231, 0.318]
admin-street-homonym	400	0.087	[0.064, 0.119]
date-name	400	0.055	[0.037, 0.082]
bare-locality (emits no street)	400	0.980	[0.961, 0.990]
OVERALL	2800	0.494	[0.476, 0.513]

Nothing here needs squinting at: the numbered classes and the bare classes are separated by ~70pp with intervals nowhere near touching. Contrast the Paris board's 3/15 cell, whose interval is [0.07, 0.45] — this is the difference between a measurement and an anecdote with a decimal point.

2. What the model actually says

The rates undersell it. The parses:

"Rue Montmartre"          -> locality="Rue Montmartre"
"Allee Poque"             -> locality="Allee Poque"
"Allee Bienville"         -> locality="Allee Bienville"
"Allee Capitaine Moret"   -> locality="Allee Capitaine"   region="Moret"

The designator is swallowed into the locality. Rue is the most deterministic street signal in French — it can only mean street — and without a house number it is absorbed into a locality span. The arc's §2 framing ("the prior on bare toponyms is wrong") is confirmed and is understated: the model is not mislabelling an ambiguous toponym, it is mislabelling Rue.

And the date-name class, which is the tell:

"Allee du 11 Novembre"       -> street="Allee du"  house_number="11"  street="Novembre"
"Allee du 11 Septembre 1944" -> street="Allee du Septembre"           (11 and 1944 deleted)
"Allee du 8 Mai"             -> locality="Allee"   street="du"        house_number="8 Mai"
"Allee de l'An 2000"         -> "l"

Here the model does emit street — because a digit is present. It then steals the digit as a house number and fragments the street around the hole. 5.5% correct, the worst class on the board.

3. One mechanism explains every cell

The house number is not a hint. It is the licence. The model has learned "digits ⇒ this is a street address"; the street reading is conditioned on that licence, not on the designator.

class	licence?	consequence	rate
street-housenumber	yes, clean	parses correctly	0.925
alnum-housenumber	yes, clean	parses correctly	0.925
bare-street	none	designator-led phrase → locality	0.215
street-particle	none	same	0.273
admin-street-homonym	none	same, and the toponym is a place → stronger	0.087
date-name	yes, but spurious	licence granted, digits stolen, street split	0.055
bare-locality	none	→ locality — correct	0.980

The negative class passes for the wrong reason. bare-locality scores 0.980 not because the model recognises a commune, but because it calls everything without a house number a locality, and on bare localities that is accidentally right. That is a standing prediction for T2: teaching bare streets may regress bare-locality, because the class currently free-rides on the same broken default. This is precisely why the negative class is a guard and not a trophy.

It also unifies the digit findings that looked like two bugs. T1a: the span decode pulls a digit into the street ("korunní" → "korunní 8"). T1c: the BIO decode pushes digits out of the street ("Allee du 11 Novembre" → "Allee du Novembre"). One confusion — the model cannot tell a digit that belongs to a street name from a digit that is a house number — surfacing from both sides.

4. Pre-registered targets for T2

The board is the grading set. Registered as baselines; the shard is graded against these exact cells.

Must move (the thesis): bare-street 0.215 · street-particle 0.273 · admin-street-homonym 0.087 · date-name 0.055.

Must not regress (the guards):

street-housenumber 0.925 and alnum-housenumber 0.925 — the contextful guard. A shard that fixes fragments by degrading normal addresses has moved the failure, not fixed it.
bare-locality 0.980 — the one to watch, per §3. If it falls while the bare classes rise, the shard traded one default for another rather than teaching the distinction.
The global parity floor (eval parity) must hold. Board 2 moving is not a verdict alone.

Split discipline: the board reserves its 2,400 street surfaces in ban-fragments-fr.surfaces.txt. The T2 shard MUST exclude every one — source-disjoint by normalized surface, never by record row. Row-disjoint leaks the surface across the boundary and measures memorization of Rue de Rivoli while claiming generalization to unseen streets.

5. Known limitation

BAN's sharded DBs retain only locality_norm / locality_base, so the negative class is accent-stripped (Amelie-les-Bains-Palalda). French title-casing is reconstructed; the accents cannot be. This makes bare-locality mildly out-of-distribution and may inflate its absolute difficulty — it does not affect the comparison the board exists for, since every model sees the identical input. Source commune surfaces from a register that keeps them if the absolute number ever matters.

Note the two negative-class rates are not comparable across fixtures: the shipped model hallucinates on 2% of these clean BAN communes but 22% of parity's 54 street-free rows, which include venues, BOOM, and New York, NY. Different populations, both real.

Reproduce: mailwoman eval fragment-board --weights-cache <dir> (add --klass bare-street for a fast loop). Rebuild the fixture with python3 scratchpad/build-ban-fragment-board.py — seeded, so it reproduces byte-identically.

1. The board​

2. What the model actually says​

3. One mechanism explains every cell​

4. Pre-registered targets for T2​

5. Known limitation​

1. The board

2. What the model actually says

3. One mechanism explains every cell

4. Pre-registered targets for T2

5. Known limitation