rewirebio.iobenchmarks
Result

Covariance agreement (residue covariance agreement against a 300-sequence natural homolog ceiling) of Evotuned ESM-2 650M, unforced infilling on HER2-VH reference holdout, 200 homologs

0.93 Covariance agreement (residue covariance agreement against a 300-sequence natural homolog ceiling)

Methods

Tested configuration
Evotuned ESM-2 650M, unforced infilling
Protocol
EditJumps section 4.2 frame, 200 references: cov. agr.
Dataset subset
HER2-VH reference holdout, 200 homologs
Procedure
One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.
Evaluation
Evotune on HER2-VH, cov. agr., 200 references
Coverage
Not reported scored / Not reported eligible
Uncertainty
Not reported by the source: The table prints a single value with no interval.
Evidence
Author-reported evaluation · source checkedWhen Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "her2vh", row "Evotune", column "cov. agr."

A source-checked result verifies the numerical transcription, not every model or protocol detail. Evaluation metadata: source checked. Source checked does not mean independently reproduced.

Reproduction

Split
20 templates and 20 variants per template, 400 generated sequences, scored against 200 held-out HER2-VH homologs; agreement ceiling 300 real HER2-VH homologs
Adaptation
Per-family retraining of a 650M trunk for 2,000 steps, then iterative infilling
Scoring implementation
residue covariance agreement against a 300-sequence natural homolog ceiling

No execution recipe has been verified for this exact configuration and evaluation. A benchmark's general instructions may use different inputs, splits or model settings.

Reproducing this published result requires matching its model configuration, data, split and scorer. Source checking or a successful smoke test does not establish score reproduction.

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.

Evidence table

Inspect claims, sources and review details

Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.

One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.

1 evidence row matching the loaded filters

Claims, original sources and review scope · Release 2026-11-21-c8bea25e5abb
Property and statementOriginal source and locationReview and provenance
Reported result
0.93
Individual claims
When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1)

Original source ↗

Table 2, family block "her2vh", row "Evotune", column "cov. agr."

Version: arXiv:2609.18745v1, 16 September 2026
Retrieved: 2026-11-16T00:03:00Z

source checked

["independent-cell-check","deterministic-table-parse","pdf-text-parse","source-hash-verification"] · 2026-11-16T01:30:00Z

author reported

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Re-read from the pinned PDF's text layer and from the arXiv HTML rendering. The value sits in the column and the family block its locator names under both column models, which were built independently of the extraction scripts. The printed dash in the four prof. LL cells of rand. mut. and rand. pair. is a printed dash in both renderings, and no result was created for it.

Field: attributes.printed_value

Source artifact SHA-256: 4211d8ef46c55efe8d89ba4ee6c0b4e57cca1e78d0856e9cdfbde37046a7f4e1

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

Extraction artifact SHA-256: 4211d8ef46c55efe8d89ba4ee6c0b4e57cca1e78d0856e9cdfbde37046a7f4e1

Extraction artifact

Sources and history

Release 2026-11-21-c8bea25e5abb · Record review: source checked

1 source record and release historyDownload this release (gzip)
Technical metadata and extraction receipts

Stable ID: editjumps-20261115-result-ours-200-covariance-agreement-evotuning-her2vh

areas
protein-function; proteins-complexes
contexts
research
limitations
Computational distribution comparison only. The value measures how similar generated sequences are to a reference set of natural sequences; it is not a measured binding affinity, specificity, stability or retained function.; Scores depend on the size of the reference set. Expanding the reference alignment from 200 to 800 homologs alters the normalised spectrum MMD ratio by 50% on Ty1 and 26% on HER2-VH, which the paper states is large enough to invert method rankings, so the ordering of methods is not stable across reference-set sizes.; 113 of Ty1's and 103 of HER2-VH's 200 reference sequences were present in the training set. Re-evaluating with strictly disjoint reference sets shifted spectrum MMD by -36.7% on Ty1 and -12.6% on HER2-VH and increased covariance agreement by 12 to 17 points.; The two held-out targets are single-chain constructs, the camelid VHH Ty1 and the isolated heavy chain HER2-VH, while the training pairs are concatenated two-chain VH.VL sequences. The paper names this as a domain shift.; Aggregate similarity is substantially driven by conserved framework scaffolds. 76% of residues are framework and 24% are complementarity-determining regions, and within those regions the model closes under 25% of the gap between random mutation and natural homologs, so a high aggregate score is not evidence of a better antigen-binding loop.; Seed values are not printed. Table 1's caption says it reports individual runs and intervals across three seeds; Table 2 prints one value per cell with no interval.; The paper states that its computed values and the published EvoFlows values are not comparable and that it could not recover the original experimental conditions from the EvoFlows paper. Table 1's caption adds that the two blocks are not comparable and that each row compares only against its own block's floor and ceiling.; The EvoFlows paper specifies the covariance matrix but omits the scalar reduction, so the reduction scored here is the EditJumps authors' own choice.; The paper reports that this score moves with the reference partition alone, from 83.6% to 85.1%, and with the reference set size, from 0.790 at 20 references to 0.985 at 800.
metric
covariance-agreement
metric direction
higher
metric qualifier
residue covariance agreement against a 300-sequence natural homolog ceiling
missing metadata
uncertainty: note: The table prints a single value with no interval.; reason: unreported
numeric value
0.93
printed value
0.93
review
artifact sha256: 4211d8ef46c55efe8d89ba4ee6c0b4e57cca1e78d0856e9cdfbde37046a7f4e1; method: independent-cell-check; deterministic-table-parse; pdf-text-parse; source-hash-verification; method note: Independent review pass. Every pinned artifact was re-downloaded from the URL in its source record and its SHA-256 recomputed: all nine match. The arXiv PDF's text layer was regenerated with pdftotext -layout from poppler 26.08.0 and matches the hash notes.md records. npm run batch -- check-values over the whole batch reproduced the extractor's figures exactly (8 in_cell, 164 in_row, 3 locator_not_parsed, no wrong_column and no overclaimed_in_row). Because 164 results sit at in_row only, column assignment was then checked separately by review/column_check.py, written for this pass and using neither extract/tables.py nor extract/check_tables.py. It builds two independent column models: the PDF text layer, where Tables 2 and 4 are assigned by clustering printed field right edges and Table 1 by ordinal position within the row, and the arXiv HTML rendering parsed as a cell grid from the table markup. The two models agree on every cell, and all 172 table-located results match their locator's column and block in both. Four negative controls, two columns exchanged within a row, a row shifted one column left, a Ty1 and HER2 sub-column exchanged, and the Table 1 upper and lower blocks exchanged for one row label, are rejected by both models.; note: Re-read from the pinned PDF's text layer and from the arXiv HTML rendering. The value sits in the column and the family block its locator names under both column models, which were built independently of the extraction scripts. The printed dash in the four prof. LL cells of rand. mut. and rand. pair. is a printed dash in both renderings, and no result was created for it.; retrieval url: https://arxiv.org/pdf/2609.18745v1; reviewed at: 2026-11-16T01:30:00Z; reviewer: claude; reviewer note: Claude (Opus 5) review agent for issue #244, independent of the extraction agent; no human review claimed. All nine pinned artifacts were re-retrieved and their SHA-256 re-computed; no value was reproduced and no model was run.
source locator
Table 2, family block "her2vh", row "Evotune", column "cov. agr."
unit
unitless
Related records

Suggest a correction