Fill 418,271 rescored secondary-metric values

interface_lddt, the CDR RMSDs and epitope_jaccard were empty on the cells labelled by a host whose environment lacked PyYAML and ANARCI. They are recomputed by the same scorers from the same structures and filled in: 418,271 cells go from null to a value in boltz2, esmfold2 and protenix-v2. No populated value changed and no populated value became null. opendde-abag was already complete and is unchanged. epitope_jaccard stays null on the seven targets with no resolvable native antigen chain, so the 2026-08-14 false-zero correction stands.

README.md

@@ -67,6 +67,14 @@ its pLDDT selector is quantised to 4 decimals, which leaves about 181 distinct v
 a top-selector tie on 20 of the 161 scorable targets. Break those ties the other way and
 esmfold2's delivered@512 reads 0.2878 rather than 0.2852.
 
+## Computed on Tenstorrent
+
+All 335,360 folds ran on **[Tenstorrent](https://tenstorrent.com)** hardware, a 32-chip Wormhole Galaxy, using
+**[TT-Bio](https://github.com/moritztng/tt-bio)**, our open-source stack for running structure
+prediction models on Tenstorrent. Four models, 512 samples per target, 164 targets. At this scale
+cost per prediction is what decides whether a study like this is affordable at all, which is most of
+why this dataset exists.
+
 ## The benchmark this builds on
 
 The 164 targets are **2026ARK-AB**, the antibody-antigen benchmark released with OpenDDE: 164 PDB
@@ -135,9 +143,9 @@ per distinct chain sequence).
 | `irmsd` | float32 | Å | interface backbone RMSD |
 | `lrmsd` | float32 | Å | ligand RMSD |
 | `fnat` | float32 | [0,1] | fraction of native contacts recovered |
-| `interface_lddt` | float32 | [0,1] | lDDT over interface atom pairs. Partial coverage, see limitations |
-| `cdr_h1_rmsd`, `cdr_h2_rmsd`, `cdr_h3_rmsd` | float32 | Å | per-CDR-loop RMSD after alignment. Partial coverage |
-| `epitope_jaccard` | float32 | [0,1] | overlap of predicted and native antigen contact residue sets. Partial coverage, and null on the 7 targets with no resolvable native epitope, see limitations |
+| `interface_lddt` | float32 | [0,1] | lDDT over interface atom pairs. Null on 4 targets and a few scattered poses, see limitations |
+| `cdr_h1_rmsd`, `cdr_h2_rmsd`, `cdr_h3_rmsd` | float32 | Å | per-CDR-loop RMSD after alignment. Null on 4 targets for H1, 5 for H2 and H3, see limitations |
+| `epitope_jaccard` | float32 | [0,1] | overlap of predicted and native antigen contact residue sets. Null on the 7 targets with no resolvable native epitope, see limitations |
 | `seed` | int32 | | diffusion seed for the chunk |
 | `mps` | int8 | chips | chips per fold job. Null where the fleet recorded `auto` |
 | `wall_s` | int32 | s | wall time of the 64-sample chunk, not of one sample. All 64 rows of a chunk share it |
@@ -194,22 +202,54 @@ leaving the value empty. Those rows read as "the model missed the epitope entire
 "not computable". They cover 7 targets (9kwy, 9ly2, 9ly3, 9lz2, 9ull, 9ulm, 9ynx) in all four models.
 
 Rows were selected on the scorer's own `native_epitope_size == 0`, never on the value being zero.
-**The 44,433 exact zeros on other targets are real measurements** and are unchanged, as is every
-other value in every other column: 11,776 cells moved across the whole `samples` config and nothing
-else did.
+**The exact zeros on other targets are real measurements** and none of them changed, nor did any
+other value in any other column: 11,776 cells moved across the whole `samples` config and nothing
+else did. There were 44,433 of those real zeros at this commit. The rescore below then filled rows
+that had been empty, some of them with real zeros, and the count is now 52,440.
 
 The dataset was public with the old values from about 10:41 UTC on 2026-08-14 until this commit. To
 tell which copy you hold, read `epitope_jaccard` on target 9kwy: null is the corrected data, 0.0 is
 the old data. Either re-pull, or drop `epitope_jaccard` on those seven targets.
 
+## Rescored secondary metrics, 2026-08-14
+
+**418,271 `interface_lddt`, CDR-RMSD and `epitope_jaccard` values changed from null to a value.**
+Nothing else moved. The gaps were never a property of the folds. The labelling ran across two hosts
+and one of them had no PyYAML and no ANARCI, so every metric that needs them came back empty on the
+cells that host scored, which is why the missing values fell on whole 64-sample jobs rather than on
+individual poses. All 111,616 affected predictions were still on disk, so the repair was to rescore
+them with the same scripts, the same interpreter and the same structures that produced every value
+already in this dataset.
+
+That makes it a fill, and it is asserted as one. Re-running those scorers on 22,368 already-populated
+values reproduced all 22,368 bit-for-bit, and the upload moved nothing: `value_changed` is 0 and
+`value -> null` is 0 on every column of every file. The fills are 41,472 in boltz2, 249,301 in
+esmfold2 and 127,498 in protenix-v2. opendde-abag was already complete and its file is unchanged.
+
+`epitope_jaccard` stays null on the seven targets named above. The rescore does compute it there, as
+`0.0`, which is exactly the artifact the correction above removed, so those 14,336 rows are left
+empty on purpose.
+
+**The stated reason for the three unscorable targets was also wrong, and is corrected.** This card
+said 9ly2, 9ly3 and 9lz2 are 3-way Ab:Ag hetero-hexamers whose antibody-antigen interface the scorer
+cannot resolve. The chain map is in fact correct and explicitly declared. The real mechanism is
+chemical: 71/71, 78/78 and 48/48 of the antigen-side contact atoms on their declared interface sit on
+phosphoserine, which DockQ drops. No value changed with that correction, only the explanation.
+
+Earlier pulls carry the old, ragged columns. To tell which copy you hold, count non-null
+`cdr_h3_rmsd` in `samples/esmfold2.parquet`: 81,408 in this data, 15,360 in the old.
+
 ## Known limitations
 
 * **164 targets.** Enough to separate the four models' oracle-delivered gaps with intervals that do
   not overlap zero. Not enough to support per-epitope-class or per-germline claims, and not a
   general statement about co-folding beyond antibody-antigen complexes.
-* **Three targets carry no DockQ.** 9ly2, 9ly3 and 9lz2 are 3-way Ab:Ag hetero-hexamers in their
-  asymmetric units and the scorer resolves no antibody-antigen interface for them, in any model.
-  Their confidence values ship; `dockq` is null. So 161 targets are scorable, in every model.
+* **Three targets carry no DockQ.** 9ly2, 9ly3 and 9lz2 are anti-phosphoepitope antibodies. Every
+  contact atom on the antigen side of their declared interface sits on a phosphoserine (71/71, 78/78
+  and 48/48), and DockQ scores only standard residues, so the interface has nothing to score. The
+  fold input carries unmodified serine at those positions, so the quantity does not exist on the
+  prediction side either. Their confidence values ship; `dockq` is null. 161 targets are scorable, in
+  every model.
 * **`epitope_jaccard` is null on seven targets.** 9kwy, 9ly2, 9ly3, 9lz2, 9ull, 9ulm and 9ynx have
   no resolvable native antigen chain, so there is no native epitope set to compare a prediction
   against and no overlap exists to measure, in any model. Read those nulls as "not computable", not
@@ -220,11 +260,19 @@ the old data. Either re-pull, or drop `epitope_jaccard` on those seven targets.
   scorer, and a scan over the whole panel found it the only such case. It is a pipeline artifact,
   not model behaviour, and no published number used it. That is why `samples` has 655 x 512 rows,
   not 656 x 512.
-* **The secondary metrics are ragged, badly so for esmfold2.** `dockq` and `irmsd` are populated on
-  every scorable sample in all four models. `interface_lddt`, `epitope_jaccard` and the CDR RMSDs
-  are not: mean per-target depth for `cdr_h3_rmsd` is 419.8 for boltz2, 505.6 for opendde-abag,
-  330.0 for protenix-v2 and **95.4 for esmfold2** out of 512. Quote those metrics at their own
-  depth. No epitope or CDR claim from this dataset is a claim at n=512.
+* **The secondary metrics are near-complete, and null where the quantity does not exist.** `dockq`,
+  `irmsd`, `lrmsd` and `fnat` are populated on every scorable sample in all four models. Mean
+  per-target depth out of 512, for boltz2 / esmfold2 / opendde-abag / protenix-v2: `interface_lddt`
+  499.4 / 498.9 / 499.4 / 499.4, `cdr_h1_rmsd` 496.8 / 496.8 / 496.8 / 496.8, `cdr_h2_rmsd` and
+  `cdr_h3_rmsd` 496.4 / 496.4 / 496.3 / 496.4, `epitope_jaccard` 490.1 / 490.1 / 490.0 / 490.1. The
+  shortfall is named targets, the same ones in every model. `interface_lddt` is null on 9ly2, 9ly3,
+  9lz2 and 9mz8, whose native antigen chain does not resolve. The CDR RMSDs are null on 9l9y, 9mnu,
+  9msc and 9udq, whose native heavy chain cannot be IMGT-numbered, and H2 and H3 additionally on
+  9lwc. `epitope_jaccard` is null on the seven targets above. Beyond those, a thin per-pose residual
+  remains, scattered rather than by target: `interface_lddt` on 13 / 108 / 0 / 16 samples and
+  `cdr_h1_rmsd` on 439 / 450 / 434 / 449. `cdr_h2_rmsd`, `cdr_h3_rmsd` and `epitope_jaccard` have
+  none. Quote these metrics at their own depth, and read the nulls on the named targets as "not
+  computable" rather than as misses.
 * **512 is a decision cap, not a measured knee.** The oracle's gain per doubling is still positive at
   the top rung, so the ceiling has not saturated. Nothing here says 512 is where sampling stops
   paying.

samples/boltz2.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:014c03d6a0b75f39925af118f8878dfc5588d0be01f09c641ee3a98e0926dd01
-size 5038308
+oid sha256:105cf64ee791fd16e7d9b59d01480fceda9841d9ca358fc8b69f3ccc14eb8c71
+size 5254680

samples/esmfold2.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:53a6e9a69eee00940d432ebcb185f2126e9c2009a522a4985dad865004160f90
-size 2490998
+oid sha256:684252e8cdfa4119ef5e6ab7814f6a4f973789eda0a7d1504aab34c833c013a4
+size 3686423

samples/protenix-v2.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9810ee501aee358232eddbb1a776c7a5e27f3be816f88f42b8a158e7a6299904
-size 4747859
+oid sha256:a8e48aa50f836c8acc951a41dd0f105ac15997cf674cfd5d5947eb8c15c07832
+size 5319174