T2 — the BAN fragment shard: the licence hypothesis holds

Pre-registered gate (config header + baselines.json profile fragment-fr-v264, registered against the shipped model before the run): targets move, guards hold.

Verdict: PASS on both boards. Every target moved, every guard held, and the two contextful guards improved. The house-number-licence diagnosis (T1c) is confirmed by the fix working.

This is not a promote gate — see §5.


Run	`ap-onoBImKJriMFy7Pz3i5t0D`, config `v3.1.0-fr-fragment`
Change	ONE variable vs shipped v264: the `synth-fr-fragment` shard @ 12.0
Shard	144,865 rows, BAN (Licence Ouverte), 0 eval surfaces
Training	8000/8000 steps, 0 errors, no NaN, `cross_pollution 0.00%`
Precision	int8 candidate vs int8 baselines (asserted, see §4)

1. Board 2 — the fragment board (the read)

class	v264	95% CI	v310	95% CI	Δ
bare-street	0.215	[0.178, 0.258]	0.715	[0.669, 0.757]	+50.0pp
street-particle	0.273	[0.231, 0.318]	0.855	[0.817, 0.886]	+58.2pp
admin-street-homonym	0.087	[0.064, 0.119]	0.517	[0.469, 0.566]	+43.0pp
date-name	0.055	[0.037, 0.082]	0.158	[0.125, 0.196]	+10.3pp
street-housenumber (guard)	0.925	[0.895, 0.947]	0.948	[0.921, 0.965]	+2.3pp
alnum-housenumber (guard)	0.925	[0.895, 0.947]	0.960	[0.936, 0.975]	+3.5pp
bare-locality (guard)	0.980	[0.961, 0.990]	0.980	[0.961, 0.990]	±0.0
OVERALL	0.494	[0.476, 0.513]	0.733	[0.717, 0.749]	+23.9pp

No target's interval overlaps its baseline. These are not readings anyone has to squint at.

The archetype:

"Rue Montmartre"   v264 -> locality="Rue Montmartre"     v310 -> street ✓

2. The prediction that mattered: `bare-locality` held

T1c's standing prediction was that this cell might collapse. It read 0.980 for the wrong reason — the model called everything without a house number a locality, and on bare localities that was accidentally right. Teaching bare streets gave it every incentive to flip that default and trade one broken prior for another.

It held exactly: 0.980 → 0.980. Same interval. The model learned the distinction rather than a new default, and the discriminating evidence is the designator, which is the only thing that actually separates the two classes.

That is what the counter-distribution was for. 20% of the shard is bare communes carrying no street label; without them, the cheapest way to satisfy every other row is to start calling bare toponyms streets. The guard is why we can say the model learned rather than swapped.

The contextful guards did not merely hold — they improved (+2.3pp, +3.5pp). The shard did not buy fragments with full addresses.

3. Board 1 — the global parity floor (the guard)

label	v264	v310	Δ
street	154/267 = 0.577	162/267 = 0.6067	+3.0pp
house_number	117/146 = 0.8014	117/146 = 0.8014	identical
postcode	71/72 = 0.9861	71/72 = 0.9861	identical

Nothing regressed; street gained +8 fixtures on the broad corpus. The FAIL verdicts printed by eval parity are the v7 campaign floors (street 0.90 / hn 0.97) — v264 fails them too. They are a target, not a regression gate.

4. What the in-training eval said, and why it is not the verdict

	v264	@2000	@4000	@6000	@8000
street	0.8105	0.798	0.811	0.810	0.805
region	0.8402	0.833	0.830	0.830	0.831
locality	0.7969	0.799	0.794	0.794	0.795
val_loss	1.1819	1.1992	1.1990	1.1995	1.2009

The val split contains no bare streets. It is base-distribution, so it could only ever price the cost — never the win. Read alone it says "regression"; the fragment board says +50pp. Both numbers are correct about different questions.

Two things survive from it:

street dipped and re-converged (0.798 → 0.811) while carrying the fragment mass — the same story the boards tell.
region found a new lower equilibrium at ~0.830 (−1.0pp) and plateaued across four evals. It looked like the open item — #1102 is titled "fragment/twin training mass erodes US region+locality recall (~2.5pp) — the promote blocker", the same failure shape at a larger magnitude, and it blocked a promote before. It did not reproduce on the canonical gate (§4.1).

The precision confound, caught in the act. The first read graded an fp32 export against int8 baselines — the shipped weights package ships int8 (39.4 MB), a fresh export_onnx is fp32 (157 MB). The gate specs cap int8-vs-fp32 at 1.5pp, larger than several cells above. The read script's md5 check passed it happily, because "differs from the baseline model" is trivially true of a precision change: an md5 difference proves the file changed, not that it is comparable. Fixed (5e5f9c2a) — the pipeline is export → quantize → grade, and the script now asserts the precision class matches. Every number on this page is int8-vs-int8.

4.1 The golden per-tag battery — the 2pp pre-publish gate

The canonical instrument the promote rule names (eval error-analysis, strict createScorer, full ship config, int8 both sides):

tag	v264	v310	Δ
region	78.8%	78.6%	−0.2pp
locality	48.8%	48.7%	−0.1pp
postcode	97.3%	97.2%	−0.1pp
house_number	97.1%	97.0%	−0.1pp
venue	37.1%	37.1%	±0
street	15.8%	16.3%	+0.5pp
country	89.0%	90.2%	+1.2pp
exact match	25.5%	25.7%	+0.2pp

Every tag is inside the 2pp bar. Two improve. The pre-publish gate PASSES.

So the val split's −1.0pp region was a false positive, not a preview. Worth stating plainly because it inverts the arc's usual failure: the other instruments here missed problems that were real; this one reported a problem that was not. Same root — a metric on a distribution nobody asked about — and the same rule: read the instrument the gate names.

That does not make region uninteresting. It is the tag the fragment mass pulls on, #1102 is real, and a heavier weight or a second locale's shard could push −0.2pp into −2pp. It is a thing to watch, not a thing to ignore, and it is cheap to watch: this table is one command.

4.2 The pre-registered "ALSO" — which did NOT come out as predicted

The gate carried a third clause: "hallucination_rate@v301-span 0.352 should FALL if T1c's diagnosis holds." Measured on the 54 street-free parity rows — the MESSY population (venues, all-caps junk, French zone names), as distinct from the fragment board's clean BAN communes:

	rate	95% Wilson
v264 (shipped)	12/54 = 0.222	[0.132, 0.349]
v310 (fr-fragment)	17/54 = 0.315	[0.207, 0.447]

It rose. 6 introduced, 1 removed. The intervals overlap heavily and McNemar on the paired discordants (6 vs 1) gives p ≈ 0.125, so the rate change is not established — but the direction is against the prediction, and two of the six have a mechanism worth naming:

✗ "ZAC sous la Combe Lavancia-Epercy" -> street="Combe Lavancia-Epercy"
✗ "ZA Entraigues Embrun"              -> street="ZA"

ZAC / ZA are French zone designators (zone d'aménagement concerté), and sous la is exactly the particle pattern the shard teaches. This is on-thesis over-generalization: the grammar the shard installs — "French designator + particle + name = street" — is slightly too wide, and French zone names sit just inside its edge. The other four (BOOM → Boom, new south wales aus, philadelphia museum of art) are pre-existing non-French confusions that shifted.

The two populations do not contradict each other; they partition:

Clean French communes (board, n=400): bare-locality held at 0.980. The counter-distribution worked exactly where it was aimed.
French zone names: a new, narrow class the counter-distribution does not cover, because it mints only communes.
Non-French junk: noise.

The fix is a sibling of what is already there — extend the counter-distribution to French zone designators (ZAC, ZA, Lotissement, Résidence as NOT-street). Not a new idea, a wider one. It is a shard change, so it belongs to the next run, not to this verdict.

This is what pre-registration is for. Two of the three gate clauses passed decisively; the third went the other way and is on the page at the same size as the wins.

5. What this does and does not establish

Establishes: the failure was the training distribution, exactly as T1c diagnosed. A model that read Rue Montmartre as a locality now reads it as a street, and does so without giving up the contextful case or starting to hallucinate streets on bare communes. The BAN registry as a soft prior + synthetic coverage source — the doctrine's whole claim — works on its first real test.

Does not establish:

Promotion. The 2pp pre-publish gate passes (§4.1), which clears the largest single hurdle — but that is one leg, not the battery. A promote still wants mailwoman eval gate --gate <spec> (the per-locale floors, the int8↔fp32 delta cap, the cascade smoke, the mask-regression lock) and an operator GO. Nothing ships on this page.
The v7 floors. street 0.6067 vs a 0.90 floor. +3.0pp is real and it is not 32pp.
date-name. 0.158 is +10.3pp and still 84% wrong. BAN holds only ~1,418 date-name streets and the extractor already takes every one; the shard has 655 rows of it. The lever there is shard weight, not more data — and, per T1c, the underlying confusion (a digit inside a street name vs a house number) is the same one the span decode shows from the other side.
Other locales. FR only. Generalization goes by tier (BAN → BAG → TIGER-with-scope), never folded into one step.
Hallucination. §4.2: it did not fall, and on the messy population it rose (n.s.). The FR-zone class is real, narrow, and uncovered. Anyone reading the +50pp should read that too.

Reproduce: bash scratchpad/read-v310-gate.sh 008000 (export → quantize → both boards, with the precision assertion).

1. Board 2 — the fragment board (the read)​

2. The prediction that mattered: bare-locality held​

3. Board 1 — the global parity floor (the guard)​

4. What the in-training eval said, and why it is not the verdict​

4.1 The golden per-tag battery — the 2pp pre-publish gate​

4.2 The pre-registered "ALSO" — which did NOT come out as predicted​

5. What this does and does not establish​