سلام خوبی

The sample reports answer the question I was going to ask, and the answer moves the problem somewhere more interesting.

I wanted to know what the 37th measurement point was. The structure report says it outright: N+1 stations, your 36 layers plus an embedding station at 0. And 37 - 4 = 33, so this post's "33 of 37" is (S - start)/S exactly. Which is what your own rs-1.2 note says when it deprecates spread as "a deterministic function of the start station".

The deprecation is right. I think it reaches further than the report lets it.

Your change-start note says the estimator has a measured detection floor near station 4, that damage confined to stations 0-3 still surfaces at 4, and that the floor is 4 across three families at 0.5B-7B. Put the two notes together and spread is not merely non-independent, it is a constant of the instrument at this scale. Start pinned at 4 gives 89.2% at 36 layers, 86.2% at 28, 84.0% at 24, 76.5% at 16. Nothing about the fine-tune enters.

The floor is visible in your own files. llm-xray-lesion-scans is unmoved at 2ff1f6fb, so I re-ran onset on the six scans: first station where the base-to-tuned |delta| reaches a share s of its own max over depth, all seven ladder levels.

                    s=0.01   s=0.05   s=0.10
znext-v43    L24       0        0        0
van-1b       L16       0       0-1      0-1
qwen2.5-1.5b L28       0       0-1      0-1

21 of 21 rows onset at station 0 at the 1% threshold, and the latest anything reaches at 10% is station 1. Your reported start is 4 on all of them. That is your floor note, measured, on three pairs you published rather than on the controlled test.

Which is why I would not put "the deviation starts at layer 4" on the front card of the Space. On this instrument at 3B, "starts early" cannot come out false.

That leaves the MEDIUM verdict resting on less than it looks like. Its three stated inputs are start (at the floor), spread (a function of start), and behavior-changed at 100%, and your own note says a base-to-chat pairing pushes that last one to 100% on prompt format alone. All three are annotated non-independent inside the same report. What is actually carrying the verdict is N80, 14 of 36, and the band at stations 32-36 with 2.7x enrichment. Your deprecation note points readers from spread to N80, so the schema already knows this. The summary sentence on top is older than the fields that earn it. (Small thing while I am in there: the station-indexing note says percentages use N+1, but N80 and the readout depth both print over 36.)

Now the knowledge half, where one number is out of step with the rest of the report.

You report readout depth the hard way. The marginals are 32.5 and 31.8, a difference of -0.7, and you publish the matched-pair shift of -0.2 instead, restricted to facts that lock in under both models. That is the right statistic and it costs you the bigger number. Same discipline behind "0 of 18 broken, 0 weakened", which is per-fact, not a count that happens to match.

Fabrication avoidance is the one place it is not applied. 13/20 to 18/20 is a difference of marginals on the same 20 items scored twice. The paired test is McNemar on the discordant pairs, and a net +5 has a ceiling:

improved-regressed    exact two-sided p
       5-0                 0.0625
       6-1                 0.1250
       7-2                 0.1797

0.0625 is the best case, and only if not one of the 20 got worse. The first split clearing 0.05 is 6-0. So the Space's front card leads with a +5 that twenty paired items cannot resolve, in a report that computed the paired version for the field directly above it. The two discordant counts are the whole test, and you already have them.

One more, in the fields you ask readers to pin. This post and the Space index both say AUROC 0.878. The knowledge report says 0.877, delta -0.038. A single AUROC on a 20-item probe at a 10/10 split lives on a 1/200 grid, and 0.915 is exactly 183/200. 0.877 and 0.878 are 175.4 and 175.6, neither on it, and a value that renders two ways in the third digit is not a grid point at all. The smallest pair count admitting 0.915 and 0.877 both exactly is 500, which is 20 against 25.

Your own label says "(trajectory)", so I read it as a mean over stations rather than one curve. If that is right then "separation got slightly weaker" compares a mean to a mean, and the item-level resolution is Hanley-McNeil: at 10 and 10 the standard errors are 0.068 and 0.081, and both intervals run past 1.0.

Everything above I could only check because you published raw per-layer KL. The 3B pair ships as a rendered report, not a scan.

You have 11 public models in g1-2026-07, 0.5B to 72B, five families. Has any pair under 32B ever returned a start above 4?

Start is 4 on all 22 pairs because stations 0 to 3 have no probes.

That is in model-xray-gallery, which you shipped on the 3rd. No rush on the point-by-point. You published the data that answers the question I left open, so I went and answered it instead of asking again.

The answer is no, and wider than I asked

change_start_station across 22 pairs
14 base models, six orgs, 0.5B to 32B
  distinct values: {4}

Not one pair above 4, at any size. Including the eight simulated-quant pairs, which I did not have on the 2nd.

That arm is a positive control, and the field fails it. Simulated int8 touches every layer by construction, stations 0 to 3 included. The scan still reports start 4, with difference_profile at exactly 0.0 in those four cells.

The reason is in the file

contributing_probes[0:4] = [0, 0, 0, 0]     22 of 22
difference_profile[0:4]  = [0.0, 0.0, 0.0, 0.0]

It was never a sensitivity floor. Those stations have no probes assigned. The zero means not measured, not unchanged.

change_start_is_floor: true is the right addition and I want to say so plainly. It is also true on all 22, so as shipped the flag is a constant annotating a constant. It records the problem honestly. It does not yet let a reader tell two scans apart.

spread, now fully determined

spread_fraction == (S - 4) / S      exact on 22 of 22

Start is pinned, so spread is a function of station count alone. Depth in, depth out.

N80 is the live field, on the wrong axis

I reverse-engineered it: the smallest k such that the k largest station differences carry 80% of the total. Top-k over the sorted profile, exact on 22 of 22.

Because it is a top-k, the four dead stations never enter the numerator. They do enter the denominator. So the field has a ceiling that moves with depth:

flat-profile ceiling = 0.8 (S-4) / S
  S=17    0.612
  S=65    0.751
highest n80_fraction anywhere in the gallery: 0.667

Nothing exceeds its own ceiling, which is a good sign for the statistic. But Llama-3.2-1B cannot print above 0.612 while Qwen2.5-32B can reach 0.751, so the raw column is not comparable down the gallery.

Divide the ceiling out and your own trend gets stronger, not weaker:

r(log params, X)          before-after    quant
  n80_fraction               -0.274       -0.394
  share of flat ceiling      -0.447       -0.673

That is the argument for N80 carrying the verdict, improved by the fix rather than damaged by it.

One artifact inside the profile

The probe schedule is identical at every depth, seven tiers of 24, 168 total, and the dead prefix is 4 stations whether the model has 17 or 65:

dead prefix as a share of depth   4/17 = 23.5%    4/65 = 6.2%    3.8x

Within a tier the difference climbs. At every unlock it drops.

depth +1, probe count held:       rises 122 of 132 tiers
depth +1, probe count steps up:   falls 104 of 132 steps

Depth is real and new probes dilute. So the profile is a sawtooth on the unlock schedule rather than a curve. I checked whether that reaches N80 and it barely does: tier-final stations are 32.1% of selections against a 26.0% base rate. Not load-bearing. But it is mechanical, and it is in every file.

behavior_change, at n=22 now

before-after       13 of 14 at exactly 1.0,  min 0.667
simulated-quant     0 of 8  at 1.0,          0.333 to 0.833

The saturation I guessed at from one report holds on fourteen. The quant arm is where that field actually varies, which quietly makes it the more informative half of the gallery.

All of the above is one denominator and one probe schedule. None of it is a rebuild.

The dead prefix is 4 stations at every depth, from 17 to 65. Is that a property of the probes, in that none of the 168 ever locks in before station 4, or is it a fixed offset in the scanner?

Your integer qualification is right and I had it wrong.

n80_stations is a count, so the flat-profile ceiling is ceil(0.8 x measured)/S, not 0.8 x measured/S. At S=17 that is ceil(10.4) = 11, so 11/17 = 0.6471, not the 0.6118 I posted.

Re-ran the whole gallery at 50ae1ec with the corrected form:

S    measured   continuous   integer
17     13         0.6118     11/17 = 0.6471
25     21         0.6720     17/25 = 0.6800
29     25         0.6897     20/29 = 0.6897
33     29         0.7030     24/33 = 0.7273
37     33         0.7135     27/37 = 0.7297
49     45         0.7347     36/49 = 0.7347
65     61         0.7508     49/65 = 0.7538

It only bites on 14 of the 22 pairs. At S=29 and S=49 the two forms are identical, because 0.8 x 25 and 0.8 x 45 are already integers. Nothing exceeds either ceiling, so that part survives.

The correction makes the statistic look better, not worse.

r(log params, X)         continuous ceiling   integer ceiling
  before-after (n=14)         -0.447             -0.454
  simulated-quant (n=8)       -0.673             -0.745

Raw n80_fraction is untouched at -0.274 and -0.394. Every step of this has moved the same direction.

And there is a version of it that never touches the denominator at all.

Each simulated-quant pair shares its station count with the before-after pair it was built from. S is identical inside the pair, so the ceiling divides out and the comparison is immune to this whole dispute:

base                             S    BA n80   QT n80
SmolLM2-1.7B-Instruct           25      13       15
Qwen2.5-0.5B-Instruct           25      15       16
Qwen2.5-1.5B-Instruct           29      13       18
Qwen2.5-7B-Instruct             29      14       17
Llama-3.1-8B-Instruct           33      21       22
Mistral-7B-Instruct-v0.3        33      22       22
Qwen2.5-14B-Instruct            49      21       30
Qwen2.5-32B-Instruct            65      26       35

7 up, 0 down, 1 exact tie. Sign test on the 7 non-ties, two-sided p = 0.0156.

The positive control spreads wider than the fine-tune it was built from, every single time it moves at all. That is the one N80 claim that holds with the prefix missing, the denominator disputed, and my own arithmetic wrong.

Your other two check out. 0.8775 is exactly 351/400, and 400 is the smallest denominator that produces it, so 20 by 20 is the tightest pair count consistent with the stored value. McNemar at 6 and 1 is 2 x P(X <= 1) over Bin(7, 0.5) = 16/128 = 0.125.

Mistral is the tie, 22 and 22 at S=33. Does the replacement statistic separate that pair, or does it stay tied there too?

The tie is in the projection. Both of your reports break it, in their own weight tables.

I put the "Measured weight changes" sections of report 12 and report 14 side by side, block by block:

                     int8 (12)    base->instruct (14)   ratio
block 1   abs L2       0.5573         0.8663            1.55
block 4   abs L2       0.4633         0.9197            1.99
block 32  abs L2       0.5083         0.9342            1.84
overall   relative     0.2098%        0.4330%           2.06

Base to instruct moves more weight than the int8 simulation in 32 of 32 blocks, by 1.55x to 2.17x. Relative and absolute give the same order in every block.

That is not my computation on withdrawn fields. Every number above is printed on your two pages today.

The same tables say something about depth, too.

In report 14, relative change falls about 5x from block 1 to block 32 (1.551% to 0.303%). Absolute L2 barely moves (0.866 to 1.006, then 0.934). The implied reference norm grows about 5.5x down the stack, roughly 56 at block 1 and 308 at block 32, and it matches to 0.1 across both reports. So "blocks 1 and 2 changed most" describes small early weights, not where the edit landed.

It still matters for the depth view. Blocks 1 to 3 are never evaluated, so on both arms the page reads "boundary unresolved" at block 4. Any public 7B pair that differs anywhere will read the same way.

Is block 4 the first reporting depth by design, or because of probe cost?