Request allocation Contact 简体中文

ScienceTechnical reports

JM-TR-2025-007 Technical report Library v3.0

Library v3.0: commit-hash provenance for cell-line constructions, and a retrospective audit of the cross-clade build-filter defect

Dr. Anders Kjeldsen 1,* Dr. Elena Popescu 1 Dr. Mei Ling Tan 1 Dr. Rahul Venkataraman 1

1 Jurassic Meat Shenzhen Research Center, Shenzhen, China
* Corresponding author: Dr. Anders Kjeldsen · regulatory@jurassicmeat.com

Contents
  1. Abstract
  2. 1. Summary
  3. 2. Introduction
  4. 3. Materials and Methods
  5. 4. Results
  6. 5. Discussion
  7. 6. Limitations and known issues
  8. 7. Changes in Library v3.0
  9. 8. Data availability
  10. 9. Acknowledgements
  11. 10. References
  12. Supplementary Table S1

Abstract

Library v3.0 requires every active cell-line construction to reference an immutable library commit hash, replacing semantic version labels as the unit of provenance. This is a breaking change: 61 active construction records were backfilled, of which 52 mapped to a unique historical library state, 7 were resolved by manifest checksum, and 2 could not be resolved and were retired. During preparatory reproducibility work we identified a defect in the build-filter logic, present from Library v2.0 release candidate (March 2019) until its correction, under which theropod-targeted builds could admit non-theropod fossil-derived peptides if the clade filter was omitted, misspelled, or inherited from a reference-import template. A retrospective audit of all 2,961 theropod-targeted builds executed between 25 March 2019 and 13 December 2024 identified 214 builds with a mis-set filter, of which 131 admitted at least one non-theropod entry. Five of these were referenced by cell-line constructions, including the production line DN-2019-04. Re-running the affected builds with corrected filters changed 146 reconstructed positions in total, 93.8% of which carried a posterior probability below 0.6 in either state. For DN-2019-04, three positions out of 11,842 differed. The line was not reconstructed. We report the scope, root cause, corrective and preventive actions, and the assessment that the defect had no impact on released material.

Keywordsprovenancereproducibilityphylogenetic placementbuild configurationcell-line construction recordsretrospective audit

JM-TR-2025-007 · Library v3.0 · 21 January 2025 Dr. Anders Kjeldsen, Dr. Elena Popescu, Dr. Mei Ling Tan, Dr. Rahul Venkataraman Computational Genomics and Cell Line Development, Jurassic Meat Shenzhen Research Center

Public summary of internal report (41 pp.). Library contents are confidential.

1. Summary

  • Breaking change. From Library v3.0, every active cell-line construction must reference a library commit hash. Semantic version labels (e.g. "v2.8") are retained as human-readable aliases only and are no longer accepted by the Cell Line Registry as a provenance reference.
  • Backfill. All 61 active construction records were mapped to a commit hash: 52 directly, 7 by manifest checksum, and 2 not at all. The two unresolvable research constructions were retired.
  • Defect. Between Library v2.0 and v3.0, a theropod-targeted build with a mis-set clade filter could admit non-theropod fossil-derived peptides without warning. 214 of 2,961 theropod-targeted builds (7.2%) had a mis-set filter; 131 admitted one or more non-theropod entries.
  • Impact. Five affected builds were referenced by cell-line constructions. For the production line DN-2019-04, corrected re-running changed 3 of 11,842 reconstructed positions, all below a posterior probability of 0.5 in both builds. No impact on released material.
  • Prevention. Filters now fail closed. Build manifests record per-entry clade assignment and are checked against the build profile before a build can be referenced.

2. Introduction

The Jurassic Meat peptide library is the versioned reference set from which reconstructed targets are built: fossil-derived peptide fragments, recovered and authenticated by the palaeoproteomics group, held in a multiple-sequence alignment together with reference sequences from living birds and crocodilians. Since Library v2.0 the alignment has been versioned, and reconstruction builds draw a subset of it according to a build profile: a configuration that specifies the target clade, the reference taxa, the substitution model, and which fossil-derived entries may serve as anchors (JM-TR-2019-004).

A cell-line construction is the documented act of expressing a reconstructed target in a myogenic background. The construction record identifies the reconstruction build from which the target was taken and, until this release, the library version label against which that build was run. Since the cell line is the unit that is banked, characterized and, in one case, transferred to manufacturing, the construction record is the point at which computational provenance becomes product provenance (Tan et al., internal SOP CLD-014).

Since v2.0, every build has also been identified by a manifest listing the content hash of each alignment partition, and two builds with identical manifests are byte-identical (JM-TR-2019-004). Two gaps remained. The partition hashes covered the alignment but not the metadata layers held alongside it (entry metadata, clade assignments, chain-of-custody fields, modification annotations, override flags and exclusion lists), which were stored in a mutable database keyed by release label. And construction records cited the release label, not the build manifest.

Two observations motivated this release. The first was that semantic version labels did not uniquely identify library content. Between v2.0 and v2.8, corrections to metadata, chain-of-custody fields (JM-TR-2021-005) and diagenetic-modification annotations (JM-TR-2023-006) were on occasion applied to an already-tagged release without incrementing the label. The label "v2.8" therefore referred, at different times, to seven distinct content states. This was known in general terms and considered acceptable at the time, because each state was archived and the differences were documented in the change log.

The second observation arose from the first. In July 2024, while attempting to reproduce historical builds byte-for-byte from their recorded library labels, one of us (E.P.) found that the manifest of a 2022 theropod-targeted build contained 291 entries whose clade assignment was not Theropoda. The build had been configured from a reference-import template whose clade filter was set to Archosauria. The build software accepted this configuration, executed it, and labelled the output as theropod-targeted, because the target clade and the filter clade were held in separate fields and were never compared.

This report describes the commit-hash provenance requirement introduced in v3.0 (Section 3.1, Section 4.1), the defect, its root cause and its scope (Sections 3.2–3.4, 4.2–4.4), the impact assessment for cell-line constructions (Section 4.5), and the corrective and preventive actions (Section 4.6).

Contamination of dinosaur-protein datasets by exogenous and modern material is a long-standing and legitimate criticism of the field (Buckley et al., 2008; Saitta et al., 2019), and our own contamination controls are described elsewhere (JM-TR-2013-001, JM-TR-2021-005). The defect reported here is a different and narrower problem. No exogenous material was involved. The non-theropod entries were authentic fossil-derived peptides with documented provenance. They were simply the wrong clade for the build in which they appeared.

3. Materials and Methods

3.1 Commit-hash provenance

From v3.0 the library is held in a content-addressed repository. Each library state is identified by the SHA-256 digest of its canonical serialisation (alignment, entry metadata, exclusion lists and annotation tables), displayed as a 7-character prefix in human-facing documents (the build IDs listed in the release notes) and stored in full in all machine records. A semantic version label is now an alias that points to exactly one commit; re-tagging an existing label is prohibited by repository policy.

A conforming construction record carries four digests: library_commit, build_profile_commit, toolchain_commit and build_manifest_sha256. The build manifest lists every alignment row that entered the build, with its clade assignment and assignment posterior at the time of the build. The Cell Line Registry rejects any construction record lacking any of the four.

3.2 Historical library-state recovery

To backfill existing records, we reconstructed every historical content state between v2.0-rc1 (25 March 2019) and the v3.0 freeze (13 December 2024) from the archive snapshots and the repository audit log. For each semantic label we enumerated the distinct states it had referred to and the date range of each. Each state was then serialised and hashed under the v3.0 scheme, so that historical states acquired v3.0 commit identifiers retrospectively.

3.3 Build-log audit

All reconstruction builds executed between 25 March 2019 (the first v2.0 release candidate, from which the production build toolchain dates) and 13 December 2024 were enumerated from the build scheduler (n = 3,418). Builds were classified by target clade as recorded in the build profile; 2,961 were theropod-targeted. For each theropod-targeted build we recovered the effective peptide filter (the filter actually applied at run time, which is not necessarily the filter written in the profile) and the set of fossil-derived entries admitted. Builds were classed as mis-set if the effective filter admitted any clade other than Theropoda, and as affected if, in addition, one or more non-theropod entries were actually admitted. Builds before March 2020 were logged at reduced verbosity; for nine of these the effective filter was inferred from the profile and toolchain version rather than read directly.

Builds were further classed by use: exploratory (research notebooks only), reconstruction candidate (submitted to the internal reconstruction review chaired by R.V.), and construction-referenced (cited by a cell-line construction record).

3.4 Corrected re-runs

Every affected reconstruction-candidate and construction-referenced build (n = 22) was re-run using the same library state, toolchain and profile, with the clade filter corrected to Theropoda. Exploratory builds were not re-run (Section 6). For each pair of original and corrected builds we compared the maximum a posteriori reconstructed sequence across the full target set and recorded, at each differing position, the marginal posterior probability of both states. For the production target set (19 reconstruction targets, 11,842 reconstructed residues), positions were additionally annotated by gene and by posterior class.

3.5 Cell-line assessment

For each construction-referenced build, the Cell Line Development group (M.L.T.) assessed the corresponding cell line against the construction acceptance criteria in force at the time of construction and against those of v3.0. Assessments considered the number and posterior of differing positions, the line's status (production, banked research, retired), and whether any material from the line had been transferred to Manufacturing & Process Development. The Food Safety & Quality function assessed released material separately.

4. Results

4.1 Label ambiguity and backfill

The nine semantic labels from v2.0 to v2.8 referred to 37 distinct content states (Table 1). Most re-tagging events corrected metadata only. Eleven altered at least one alignment row, typically to apply a revised diagenetic annotation or to restore an entry that had been temporarily withdrawn pending chain-of-custody review.

Table 1. Distinct library content states referenced by each semantic version label, v2.0–v2.8.
Label Release date Distinct content states States altering alignment rows Longest-lived state (days)
v2.0 2019-06 3 1 211
v2.1 2019-12 2 0 139
v2.2 2020-07 4 1 98
v2.3 2021-02 5 2 176
v2.4 2021-09 3 1 120
v2.5 2022-02 4 1 141
v2.6 2022-08 6 2 77
v2.7 2023-02 3 1 105
v2.8 2023-09 7 2 192
Total 37 11

Reproducing historical builds from their semantic label alone recovered a byte-identical build manifest in 71% of v2.0 builds, rising to 93% for v2.8 (Fig. 1). Using the recovered content state in place of the label, reproducibility exceeded 98% for every release; the residual failures were attributable to toolchain differences, which v3.0 now captures in toolchain_commit.

Figure 1. Reproducibility of historical builds by library labelProportion of theropod-targeted builds whose manifest could be reproduced byte-for-byte when the library state was taken from the semantic label alone (with 95% binomial confidence band) versus from the recovered content state. n = 2,961 builds.Proportion reproduced0.600.650.700.750.800.850.900.951.00v2.0v2.1v2.2v2.3v2.4v2.5v2.6v2.7v2.8Library label at time of buildChain-of-custody fields added(JM-TR-2021-005)Library label at time of build: v2.0 From semantic label: 0.71 (0.62–0.79) From recovered content state: 0.98Library label at time of build: v2.1 From semantic label: 0.74 (0.67–0.80) From recovered content state: 0.98Library label at time of build: v2.2 From semantic label: 0.78 (0.73–0.83) From recovered content state: 0.99Library label at time of build: v2.3 From semantic label: 0.80 (0.76–0.84) From recovered content state: 0.99Library label at time of build: v2.4 From semantic label: 0.83 (0.79–0.87) From recovered content state: 0.99Library label at time of build: v2.5 From semantic label: 0.85 (0.81–0.88) From recovered content state: 0.99Library label at time of build: v2.6 From semantic label: 0.88 (0.85–0.91) From recovered content state: 0.99Library label at time of build: v2.7 From semantic label: 0.90 (0.87–0.93) From recovered content state: 0.99Library label at time of build: v2.8 From semantic label: 0.93 (0.91–0.95) From recovered content state: 1.00Figure 1. Reproducibility of historical builds by library labelProportion of theropod-targeted builds whose manifest could be reproduced byte-for-byte when the library state was taken from the semantic label alone (with 95% binomial confidence band) versus from the recovered content state. n = 2,961 builds.Proportion reproduced0.600.700.800.901.00v2.0v2.2v2.4v2.6v2.8Library label at time of buildChain-of-custodyfields…Library label at time of build: v2.0 From semantic label: 0.71 (0.62–0.79) From recovered content state: 0.98Library label at time of build: v2.1 From semantic label: 0.74 (0.67–0.80) From recovered content state: 0.98Library label at time of build: v2.2 From semantic label: 0.78 (0.73–0.83) From recovered content state: 0.99Library label at time of build: v2.3 From semantic label: 0.80 (0.76–0.84) From recovered content state: 0.99Library label at time of build: v2.4 From semantic label: 0.83 (0.79–0.87) From recovered content state: 0.99Library label at time of build: v2.5 From semantic label: 0.85 (0.81–0.88) From recovered content state: 0.99Library label at time of build: v2.6 From semantic label: 0.88 (0.85–0.91) From recovered content state: 0.99Library label at time of build: v2.7 From semantic label: 0.90 (0.87–0.93) From recovered content state: 0.99Library label at time of build: v2.8 From semantic label: 0.93 (0.91–0.95) From recovered content state: 1.00
Figure 1.Reproducibility of historical builds by library label. Proportion of theropod-targeted builds whose manifest could be reproduced byte-for-byte when the library state was taken from the semantic label alone (with 95% binomial confidence band) versus from the recovered content state. n = 2,961 builds.Source: JM Shenzhen Research Center, build scheduler audit PL-2024-11.
Source data
Library label at time of buildFrom semantic labelFrom semantic label intervalFrom recovered content state
v2.00.710.62–0.790.981
v2.10.740.67–0.80.984
v2.20.780.73–0.830.986
v2.30.80.76–0.840.988
v2.40.830.79–0.870.99
v2.50.850.81–0.880.991
v2.60.880.85–0.910.993
v2.70.90.87–0.930.994
v2.80.930.91–0.950.996

Of the 61 construction records active at the v3.0 freeze, 52 mapped to a single historical state on the basis of build date. Nine fell within a period in which their label referred to more than one state; for seven of these, re-running the build against each candidate state yielded exactly one manifest matching the checksum retained in the build log. The remaining two research constructions (DN-2020-05 and DN-2020-06) had no retained checksum and could not be distinguished between two candidate states. Both were retired from the bank on 9 December 2024. Neither had been used outside the research organization.

Table 2. Backfill of commit-hash references for active construction records.
Construction class Active records Mapped directly Mapped by manifest checksum Unresolvable (retired)
Deinonychus, production 1 0 1 0
Deinonychus, research (banked) 49 42 5 2
Exploratory, non-theropod target 3 3 0 0
Avian reference control lines 8 7 1 0
Total 61 52 7 2

Construction of the production line DN-2019-04 was initiated on 15 April 2019 against release candidate v2.0-rc2 (JM-TR-2019-004), recorded in the construction record under the label v2.0. The rc2 state is the first of the three content states that label came to refer to. It was resolved by manifest checksum. Its construction record now references the v2.0 state it was built from and, by attestation (Section 4.5), the v3.0 commit 5d3b8f1.

4.2 The defect

A build profile specifies the target clade (target_clade) and, separately, a peptide filter (peptide_filter) that determines which fossil-derived entries may enter the build as anchors. The two fields were introduced at different times and served, in principle, different purposes: the target clade governs where the reconstruction is read from on the tree, while the filter governs which data may constrain it. In practice they were expected always to agree for species-targeted builds, and nothing checked that they did.

Three distinct configurations produced a mis-set filter (Table 3).

Table 3. Failure modes among mis-set theropod-targeted builds, 2019-03 to 2024-12.
Mode Description Mis-set builds Of which admitted non-theropod entries
A peptide_filter.mode left at its default value inherit, which inherited the library-level default (Archosauria, intended for reference-import builds) 121 74
B Clade string not matched against the taxonomy table (case, trailing whitespace, or a vernacular synonym such as "theropods"), resolving to an empty include set, which the filter treated as "no restriction" 58 39
C Profile copied from a reference-import template in which the filter was explicitly set to Archosauria 35 18
Total 214 131

All three modes fail open: an ambiguous or missing filter admitted everything rather than nothing. Mode B is the least defensible. The behaviour of an empty include set was documented in the toolchain's developer notes from 2018 as "intentional, for convenience during development" and was not revisited when the toolchain was promoted to production use at v2.0.

Mis-set builds that admitted no non-theropod entries (83) did so because non-theropod entries were excluded by other criteria in the same profile (most often a minimum fragment length, or an assignment- posterior floor), not because the filter operated as intended.

4.3 Scope over time

The incidence of mis-set filters was broadly stable from 2020 to 2023 at between 7% and 9% of theropod-targeted builds (Fig. 2; Supplementary Table S1). It fell in the second half of 2024 following interim containment on 2 August 2024, which required a second reviewer to confirm the effective filter of every theropod-targeted build before submission. No mis-set build was submitted after that date. The 31 mis-set builds recorded in 2024 all predate it.

Figure 2. Theropod-targeted builds by filter state and yearAll theropod-targeted reconstruction builds executed 25 March 2019 to 13 December 2024 (n = 2,961), by effective clade filter. 2019 covers March to December only. Interim containment (second-reviewer confirmation of effective filter) from 2 August 2024.Builds010020030040050060070020192019 Filter correct: 179 Total: 1882019 Mis-set, no non-theropod entries admitted: 6 Total: 1882019 Mis-set, non-theropod entries admitted: 3 Total: 18818820202020 Filter correct: 371 Total: 4022020 Mis-set, no non-theropod entries admitted: 14 Total: 4022020 Mis-set, non-theropod entries admitted: 17 Total: 40240220212021 Filter correct: 473 Total: 5172021 Mis-set, no non-theropod entries admitted: 18 Total: 5172021 Mis-set, non-theropod entries admitted: 26 Total: 51751720222022 Filter correct: 534 Total: 5862022 Mis-set, no non-theropod entries admitted: 19 Total: 5862022 Mis-set, non-theropod entries admitted: 33 Total: 58658620232023 Filter correct: 594 Total: 6412023 Mis-set, no non-theropod entries admitted: 16 Total: 6412023 Mis-set, non-theropod entries admitted: 31 Total: 64164120242024 Filter correct: 596 Total: 6272024 Mis-set, no non-theropod entries admitted: 10 Total: 6272024 Mis-set, non-theropod entries admitted: 21 Total: 627627YearFigure 2. Theropod-targeted builds by filter state and yearAll theropod-targeted reconstruction builds executed 25 March 2019 to 13 December 2024 (n = 2,961), by effective clade filter. 2019 covers March to December only. Interim containment (second-reviewer confirmation of effective filter) from 2 August 2024.Builds020040060080020192019 Filter correct: 179 Total: 1882019 Mis-set, no non-theropod entries admitted: 6 Total: 1882019 Mis-set, non-theropod entries admitted: 3 Total: 18818820202020 Filter correct: 371 Total: 4022020 Mis-set, no non-theropod entries admitted: 14 Total: 4022020 Mis-set, non-theropod entries admitted: 17 Total: 40240220212021 Filter correct: 473 Total: 5172021 Mis-set, no non-theropod entries admitted: 18 Total: 5172021 Mis-set, non-theropod entries admitted: 26 Total: 51751720222022 Filter correct: 534 Total: 5862022 Mis-set, no non-theropod entries admitted: 19 Total: 5862022 Mis-set, non-theropod entries admitted: 33 Total: 58658620232023 Filter correct: 594 Total: 6412023 Mis-set, no non-theropod entries admitted: 16 Total: 6412023 Mis-set, non-theropod entries admitted: 31 Total: 64164120242024 Filter correct: 596 Total: 6272024 Mis-set, no non-theropod entries admitted: 10 Total: 6272024 Mis-set, non-theropod entries admitted: 21 Total: 627627Year
Figure 2.Theropod-targeted builds by filter state and year. All theropod-targeted reconstruction builds executed 25 March 2019 to 13 December 2024 (n = 2,961), by effective clade filter. 2019 covers March to December only. Interim containment (second-reviewer confirmation of effective filter) from 2 August 2024.Source: JM Shenzhen Research Center, build scheduler audit PL-2024-11.
Source data
YearFilter correctMis-set, no non-theropod entries admittedMis-set, non-theropod entries admitted
201917963
20203711417
20214731826
20225341933
20235941631
20245961021

At the v3.0 freeze the library held 347 fossil-derived entries not assigned to Theropoda: 196 assigned to Crocodyliformes, 38 to Pterosauria, and 113 assigned no lower than Archosauria (incertae sedis). No entry is assigned to Sauropodomorpha or Ornithischia. Like the theropod entries, these are dominated by type I collagen, with enamel-protein fragments from tooth specimens (Fig. 3). Their presence in the library reflects the composition of the specimen collection. Isolated Kem Kem teeth are acquired through commercial channels in which attribution at sale is not taxonomic, and a proportion of teeth supplied as theropod are reassigned on peptide and morphological evidence after intake (JM-TR-2016-003).

Figure 3. Non-theropod fossil-derived entries by protein and assigned cladeCount of fossil-derived library entries not assigned to Theropoda at the v3.0 freeze (n = 347), by source protein and clade assignment. These are the entries that could be admitted to a theropod-targeted build under a mis-set filter.Assigned cladeCrocodyliformesPterosauriaArchosauria incertae sedisCOL1A1COL1A1 × Crocodyliformes: 7171COL1A1 × Pterosauria: 1717COL1A1 × Archosauria incertae sedis: 4444COL1A2COL1A2 × Crocodyliformes: 4848COL1A2 × Pterosauria: 1212COL1A2 × Archosauria incertae sedis: 3131AmelogeninAmelogenin × Crocodyliformes: 3939Amelogenin × Pterosauria: 44Amelogenin × Archosauria incertae sedis: 1717EnamelinEnamelin × Crocodyliformes: 2121Enamelin × Pterosauria: 00Enamelin × Archosauria incertae sedis: 99AmeloblastinAmeloblastin × Crocodyliformes: 1111Ameloblastin × Pterosauria: 00Ameloblastin × Archosauria incertae sedis: 44OtherOther × Crocodyliformes: 66Other × Pterosauria: 55Other × Archosauria incertae sedis: 88Figure 3. Non-theropod fossil-derived entries by protein and assigned cladeCount of fossil-derived library entries not assigned to Theropoda at the v3.0 freeze (n = 347), by source protein and clade assignment. These are the entries that could be admitted to a theropod-targeted build under a mis-set filter.Assigned cladeCrocodyliformesPterosauriaArchosauria…COL1A1COL1A1 × Crocodyliformes: 7171COL1A1 × Pterosauria: 1717COL1A1 × Archosauria incertae sedis: 4444COL1A2COL1A2 × Crocodyliformes: 4848COL1A2 × Pterosauria: 1212COL1A2 × Archosauria incertae sedis: 3131AmelogeninAmelogenin × Crocodyliformes: 3939Amelogenin × Pterosauria: 44Amelogenin × Archosauria incertae sedis: 1717EnamelinEnamelin × Crocodyliformes: 2121Enamelin × Pterosauria: 00Enamelin × Archosauria incertae sedis: 99AmeloblastinAmeloblastin × Crocodyliformes: 1111Ameloblastin × Pterosauria: 00Ameloblastin × Archosauria incertae sedis: 44OtherOther × Crocodyliformes: 66Other × Pterosauria: 55Other × Archosauria incertae sedis: 88
Figure 3.Non-theropod fossil-derived entries by protein and assigned clade. Count of fossil-derived library entries not assigned to Theropoda at the v3.0 freeze (n = 347), by source protein and clade assignment. These are the entries that could be admitted to a theropod-targeted build under a mis-set filter.Source: Library v3.0 freeze, entry metadata.
Source data
Source proteinCrocodyliformesPterosauriaArchosauria incertae sedis
COL1A1711744
COL1A2481231
Amelogenin39417
Enamelin2109
Ameloblastin1104
Other658

Among the 131 affected builds, the median number of non-theropod entries admitted was 22 (interquartile range 9–61; range 1–347). By use class, 109 were exploratory, 17 were reconstruction candidates and 5 were construction-referenced.

4.4 Effect on reconstructed sequence

Corrected re-runs of the 22 affected reconstruction-candidate and construction-referenced builds changed 146 reconstructed positions in total. The number of differing positions scaled approximately linearly with the number of non-theropod entries admitted (Fig. 4; slope 0.078 positions per admitted entry, R² = 0.97). One reconstruction candidate admitted three non-theropod entries and showed no differing position.

Figure 4. Reconstructed positions changed by filter correctionNumber of reconstructed positions differing between the original build and the corrected re-run, against the number of non-theropod entries admitted by the original build. Each point is one affected reconstruction-candidate or construction-referenced build (n = 22). Linear fit across all points: slope 0.078, R² = 0.97.Positions changed on corrected re-run051015202530050100150200250300350Non-theropod entries admittedReconstruction candidates linear fit: y = 0.0841x + 0.796, R² = 0.99Construction-referenced linear fit: y = 0.0729x + 2.23, R² = 0.97Reconstruction candidates Non-theropod entries admitted: 3 Positions changed on corrected re-run: 0Reconstruction candidates Non-theropod entries admitted: 5 Positions changed on corrected re-run: 1Reconstruction candidates Non-theropod entries admitted: 9 Positions changed on corrected re-run: 1Reconstruction candidates Non-theropod entries admitted: 11 Positions changed on corrected re-run: 2Reconstruction candidates Non-theropod entries admitted: 17 Positions changed on corrected re-run: 2Reconstruction candidates Non-theropod entries admitted: 19 Positions changed on corrected re-run: 3Reconstruction candidates Non-theropod entries admitted: 24 Positions changed on corrected re-run: 2Reconstruction candidates Non-theropod entries admitted: 28 Positions changed on corrected re-run: 4Reconstruction candidates Non-theropod entries admitted: 33 Positions changed on corrected re-run: 3Reconstruction candidates Non-theropod entries admitted: 38 Positions changed on corrected re-run: 5Reconstruction candidates Non-theropod entries admitted: 46 Positions changed on corrected re-run: 4Reconstruction candidates Non-theropod entries admitted: 52 Positions changed on corrected re-run: 6Reconstruction candidates Non-theropod entries admitted: 70 Positions changed on corrected re-run: 7Reconstruction candidates Non-theropod entries admitted: 88 Positions changed on corrected re-run: 8Reconstruction candidates Non-theropod entries admitted: 113 Positions changed on corrected re-run: 10Reconstruction candidates Non-theropod entries admitted: 196 Positions changed on corrected re-run: 19Reconstruction candidates Non-theropod entries admitted: 347 Positions changed on corrected re-run: 29DN-2019-04 Construction-referenced Non-theropod entries admitted: 14 Positions changed on corrected re-run: 3DN-2020-02 Construction-referenced Non-theropod entries admitted: 61 Positions changed on corrected re-run: 9DN-2020-02DN-2021-07 Construction-referenced Non-theropod entries admitted: 22 Positions changed on corrected re-run: 4DN-2022-03 Construction-referenced Non-theropod entries admitted: 291 Positions changed on corrected re-run: 23DN-2022-03DN-2023-01 Construction-referenced Non-theropod entries admitted: 8 Positions changed on corrected re-run: 1DN-2023-01Figure 4. Reconstructed positions changed by filter correctionNumber of reconstructed positions differing between the original build and the corrected re-run, against the number of non-theropod entries admitted by the original build. Each point is one affected reconstruction-candidate or construction-referenced build (n = 22). Linear fit across all points: slope 0.078, R² = 0.97.Positions changed on corrected re-run0510152025300100200300400Non-theropod entries admittedReconstruction candidates linear fit: y = 0.0841x + 0.796, R² = 0.99Construction-referenced linear fit: y = 0.0729x + 2.23, R² = 0.97Reconstruction candidates Non-theropod entries admitted: 3 Positions changed on corrected re-run: 0Reconstruction candidates Non-theropod entries admitted: 5 Positions changed on corrected re-run: 1Reconstruction candidates Non-theropod entries admitted: 9 Positions changed on corrected re-run: 1Reconstruction candidates Non-theropod entries admitted: 11 Positions changed on corrected re-run: 2Reconstruction candidates Non-theropod entries admitted: 17 Positions changed on corrected re-run: 2Reconstruction candidates Non-theropod entries admitted: 19 Positions changed on corrected re-run: 3Reconstruction candidates Non-theropod entries admitted: 24 Positions changed on corrected re-run: 2Reconstruction candidates Non-theropod entries admitted: 28 Positions changed on corrected re-run: 4Reconstruction candidates Non-theropod entries admitted: 33 Positions changed on corrected re-run: 3Reconstruction candidates Non-theropod entries admitted: 38 Positions changed on corrected re-run: 5Reconstruction candidates Non-theropod entries admitted: 46 Positions changed on corrected re-run: 4Reconstruction candidates Non-theropod entries admitted: 52 Positions changed on corrected re-run: 6Reconstruction candidates Non-theropod entries admitted: 70 Positions changed on corrected re-run: 7Reconstruction candidates Non-theropod entries admitted: 88 Positions changed on corrected re-run: 8Reconstruction candidates Non-theropod entries admitted: 113 Positions changed on corrected re-run: 10Reconstruction candidates Non-theropod entries admitted: 196 Positions changed on corrected re-run: 19Reconstruction candidates Non-theropod entries admitted: 347 Positions changed on corrected re-run: 29DN-2019-04 Construction-referenced Non-theropod entries admitted: 14 Positions changed on corrected re-run: 3DN-2020-02 Construction-referenced Non-theropod entries admitted: 61 Positions changed on corrected re-run: 9DN-2021-07 Construction-referenced Non-theropod entries admitted: 22 Positions changed on corrected re-run: 4DN-2022-03 Construction-referenced Non-theropod entries admitted: 291 Positions changed on corrected re-run: 23DN-2022-03DN-2023-01 Construction-referenced Non-theropod entries admitted: 8 Positions changed on corrected re-run: 1
Figure 4.Reconstructed positions changed by filter correction. Number of reconstructed positions differing between the original build and the corrected re-run, against the number of non-theropod entries admitted by the original build. Each point is one affected reconstruction-candidate or construction-referenced build (n = 22). Linear fit across all points: slope 0.078, R² = 0.97.Source: JM Shenzhen Research Center, corrected re-run set PL-2024-12.
Source data
SeriesLabelNon-theropod entries admittedPositions changed on corrected re-run
Reconstruction candidates30
Reconstruction candidates51
Reconstruction candidates91
Reconstruction candidates112
Reconstruction candidates172
Reconstruction candidates193
Reconstruction candidates242
Reconstruction candidates284
Reconstruction candidates333
Reconstruction candidates385
Reconstruction candidates464
Reconstruction candidates526
Reconstruction candidates707
Reconstruction candidates888
Reconstruction candidates11310
Reconstruction candidates19619
Reconstruction candidates34729
Construction-referencedDN-2019-04143
Construction-referencedDN-2020-02619
Construction-referencedDN-2021-07224
Construction-referencedDN-2022-0329123
Construction-referencedDN-2023-0181

The differing positions were concentrated where the reconstruction was least certain. At 110 of the 146 positions (75.3%), the state selected by the corrected build carried a marginal posterior probability below 0.5; at 137 (93.8%) it was below 0.6 (Fig. 5). No differing position had a corrected-build posterior above 0.8. This is the expected behaviour of a bracketing reconstruction: a small number of additional anchors of the wrong clade shifts the substitution-model fit only slightly, and only positions already close to a decision boundary move across it.

Figure 5. Posterior probability at positions changed by filter correctionMarginal posterior probability of the state selected by the corrected build, at each of the 146 positions that differed between original and corrected builds across all 22 re-run builds.Positions0102030405060700.20.30.40.50.60.70.80.90.2–0.3: 110.3–0.4: 380.4–0.5: 610.5–0.6: 270.6–0.7: 70.7–0.8: 2Posterior probability of corrected state0.5Figure 5. Posterior probability at positions changed by filter correctionMarginal posterior probability of the state selected by the corrected build, at each of the 146 positions that differed between original and corrected builds across all 22 re-run builds.Positions0204060800.20.40.60.80.2–0.3: 110.3–0.4: 380.4–0.5: 610.5–0.6: 270.6–0.7: 70.7–0.8: 2Posterior probability of corrected state0.5
Figure 5.Posterior probability at positions changed by filter correction. Marginal posterior probability of the state selected by the corrected build, at each of the 146 positions that differed between original and corrected builds across all 22 re-run builds.Source: JM Shenzhen Research Center, corrected re-run set PL-2024-12.
Source data
Posterior probability of corrected statePositions
0.2–0.311
0.3–0.438
0.4–0.561
0.5–0.627
0.6–0.77
0.7–0.82
0.8–0.90

None of the 146 differing positions lies in a region directly overlapped by a fossil-derived peptide. Fossil-derived peptides map to collagen and enamel proteins, and no reconstruction target is one of those proteins; their contribution to the muscle and flavor targets is indirect, through calibration of the substitution model, branch lengths and placement (JM-TR-2019-004). The defect propagated by the same indirect route, which is the only route by which fossil evidence reaches a reconstruction target.

4.5 Cell-line constructions

Five affected builds were referenced by cell-line construction records (Table 4).

Table 4. Construction-referenced builds affected by the filter defect, and disposition of the corresponding cell lines.
Cell line Status at audit Failure mode Non-theropod entries admitted Positions changed (of 11,842) Highest posterior at a changed position Disposition
DN-2019-04 Production A 14 (all Crocodyliformes) 3 0.48 Re-attested to v3.0; deviation DV-2024-019; no change to line
DN-2020-02 Retired 2022 (unrelated: passage stability) B 61 9 0.57 No action
DN-2021-07 Research, banked A 22 4 0.52 Re-attested to v3.0
DN-2022-03 Research, banked C 291 23 0.71 Quarantined from research use
DN-2023-01 Research, banked B 8 1 0.39 Re-attested to v3.0

For DN-2019-04, the three changed positions lie in two genes: two in a fast-twitch myosin heavy-chain paralog, one in a gene involved in intramuscular lipid handling. In both the original and the corrected build, neither candidate state at any of the three positions exceeded a posterior of 0.5; the positions were already listed among the low-confidence sites in the line's 2021 characterization package. The line's acceptance criteria at construction, and under v3.0, permit up to 0.5% of reconstructed positions to differ from the current reference build provided all such positions are below a posterior of 0.6. Three positions is 0.025%.

Re-attestation is the procedure by which a construction record is brought into conformity with v3.0 without reconstructing the line. The target is re-derived from the current library commit using the line's original build profile with any defects corrected; the re-derived target is compared with the sequence the line actually expresses; and the comparison is recorded against the acceptance criteria. For DN-2019-04 the re-derived v3.0 target differs from the expressed sequence at 14 positions: the three attributable to the filter defect, and eleven attributable to library growth and scoring changes between v2.0-rc2 and v3.0, including the three identified under diagenetic scoring at v2.8 (JM-TR-2023-006). All fourteen are below a posterior of 0.6 in the v3.0 build. The line therefore references v3.0 and continues to express the sequence constructed in 2019. This is the basis on which the release notes state that production cell lines reference v3.0 or later.

DN-2022-03 was constructed from a build copied from a reference-import template (mode C) and admitted every non-theropod entry in the library at the time. Two positions differed at a posterior above 0.6. The line has been quarantined from further research use pending reconstruction from a corrected v3.0 build. No material from DN-2022-03 was transferred to Manufacturing & Process Development.

Food Safety & Quality reviewed the audit with respect to released material. Every released lot derives from DN-2019-04, whose expressed sequence is unchanged by this report, and every released lot passed food-safety testing and sensory release on the basis of the product itself. No impact on released material.

4.6 Corrective and preventive actions

Table 5. Corrective and preventive actions.
ID Action Type Owner Status at publication
CAPA-2024-031 Peptide filter fails closed: an absent, empty or unmatched filter aborts the build Corrective Computational Genomics Closed (2024-10-14)
CAPA-2024-032 target_clade and peptide_filter compared at build start; mismatch aborts the build unless an exception is recorded Corrective Computational Genomics Closed (2024-10-14)
CAPA-2024-033 inherit mode removed; library-level default filter removed Corrective Computational Genomics Closed (2024-10-28)
CAPA-2024-034 Clade strings validated against the taxonomy table; synonyms rejected Corrective Computational Genomics Closed (2024-10-28)
CAPA-2024-035 Build manifest records clade assignment of every admitted entry; Registry checks manifest against profile before accepting a construction reference Preventive Computational Genomics / Cell Line Development Closed (2025-01-20, with v3.0)
CAPA-2024-036 Commit-hash provenance required for all construction records Preventive Cell Line Development Closed (2025-01-20, with v3.0)
CAPA-2024-037 Reference-import templates moved to a separate repository, not visible from the species-build template directory Preventive Research Computing Closed (2024-11-04)
CAPA-2024-038 Documentation of build-filter configuration revised Preventive Computational Genomics Open (target: next minor release)
CAPA-2024-039 DN-2022-03 reconstruction from corrected v3.0 build Corrective Cell Line Development Open

Since CAPA-2024-031 to -034 took effect, 312 theropod-targeted builds have been executed. Eleven aborted at build start on filter validation; all eleven were corrected and re-submitted. None admitted a non-theropod entry.

The audit and its proposed actions were presented to the Ethics & Bioresponsibility Council on 19 November 2024: reviewed; no concerns raised.

5. Discussion

The defect was a configuration problem with an unusually long life. It was introduced, in effect, when the build toolchain was promoted to production at v2.0, and it went undetected for five years and four months. It was not found by any of the validation steps applied to reconstruction outputs, for a straightforward reason: those steps assess whether a reconstruction is internally consistent and phylogenetically plausible, and a theropod reconstruction lightly constrained by crocodyliform collagen is both. The difference between a correct and an affected build was small, concentrated in low-confidence positions, and of the same size and character as the differences between successive library releases. It was visible only in the manifest, and the manifest was not routinely read.

That it was found at all is attributable to the reproducibility work that preceded commit-hash provenance. Reproducing a historical build requires knowing exactly what entered it, and knowing exactly what entered it exposed what should not have. We consider this the principal argument for the breaking change. Semantic version labels described the library as its maintainers understood it; commit hashes describe it as it was.

The impact on reconstructed targets was limited, and the reasons are worth stating precisely rather than generously. First, non-theropod entries are a small fraction of the library (347 against 2,903 theropod fragments at v3.0), so most affected builds admitted few of them. Second, they share the protein composition of the theropod set, and type I collagen is conserved across archosaurs to a degree that limits how far any single anchor can move the reconstruction. Third, the bracketing taxa, which do not change with the filter, dominate the posterior at most positions. A defect of the same kind in a library with more divergent non-target material, or fewer reference taxa, would not have been so benign. We do not regard the outcome as evidence that the filter was unimportant.

The case of DN-2019-04 illustrates what commit-hash provenance does and does not provide. The line's record now references v3.0, and the v3.0 reference is exact: the commit, the profile, the toolchain and the manifest are all fixed and retrievable. What the reference records is that the line's expressed sequence was compared with the v3.0 target and found to be within acceptance criteria. It does not record, and cannot record, that the line was built from v3.0. It was built in 2019. Re-attestation is the correct procedure for a production line whose characterization, banking and two years of manufacturing history are tied to a specific expressed sequence, and reconstructing the line to accommodate fourteen low-confidence positions would introduce substantially more uncertainty than it removed. The distinction is nevertheless recorded in the construction record, and readers of this report should hold it in mind when encountering the statement that production cell lines reference v3.0 or later.

Builds targeting clades other than Theropoda were outside the scope of this audit. At the time of writing these consist of three exploratory constructions supporting early work on a sauropodomorph target. Their build profiles are under revision and will be addressed in the release that introduces a dedicated profile for that program.

6. Limitations and known issues

  • Exploratory builds were not re-run. The 109 affected exploratory builds were identified and their manifests flagged, but they were not re-run with corrected filters. Their outputs remain in research notebooks. Any result derived from them and later relied upon would have entered a reconstruction candidate, all of which were re-run. We have not attempted to establish whether any exploratory result informed a decision that did not pass through the candidate review.
  • Clade assignments are treated as ground truth. The audit defines "non-theropod" by the clade assignment of each entry in the library. Those assignments are themselves placement results, with a median theropod placement posterior of 0.74 at v3.0. Some entries assigned to Theropoda are, on the same evidence, plausibly crocodyliform, and vice versa. Correcting the filter enforces the assignment; it does not make the assignment correct.
  • Early logs are incomplete. For nine builds executed before March 2020, the effective filter was inferred rather than read from the log. Three of the nine are counted as mis-set. One reconstruction candidate among them was re-run and is included in Fig. 4.
  • Two constructions could not be resolved. DN-2020-05 and DN-2020-06 were retired because their library state could not be determined. They may have been built correctly.
  • DN-2019-04 has not been reconstructed. The production line expresses a sequence reconstructed under a v2.0-rc2 state that admitted 14 crocodyliform-assigned entries as theropod anchors. It references v3.0 by re-attestation. This is the intended outcome of the procedure, and is recorded under deviation DV-2024-019.
  • Display hashes are truncated. Human-facing documents display a 7-character prefix of a 64-character digest. The probability of a prefix collision within the expected lifetime of the library is negligible. It is not zero, and machine records use the full digest.
  • The failure mode was documented. The fail-open behaviour of an empty include set (mode B) was described in the toolchain's developer notes from 2018 onward. The absence of any check of the filter against the target clade was listed as a known limitation in JM-TR-2019-004 and again in JM-TR-2023-006.

7. Changes in Library v3.0

As published in the Peptide Library Release Notes:

  • Breaking change: all active cell-line constructions are now required to reference a library commit hash.
  • Fixed an issue where non-theropod peptides could be included in theropod-targeted builds if filters were mis-set.

Also listed in the release notes for v3.0:

  • Clade-assertion check at build time (CAPA-2024-032, CAPA-2024-035).
  • Signed checksum manifest for each release.
  • Fixed a timezone handling error that recorded some build timestamps as the following day. Archived build logs were annotated, not altered. Build dates in this report are as corrected; the audit window (Section 3.3) was drawn on corrected timestamps.
  • Write access to the release branch reduced to two named maintainers. A release now requires approval from a maintainer who did not prepare it.
  • Build hosts no longer have outbound network access during builds.
  • Free-text origin_note (deprecated in v2.3) removed; content preserved as origin_note_legacy.
  • Version-only library references in cell-line records deprecated; earliest removal 2027-01.

Library v3.0 contains 2,903 theropod-assigned fossil-derived fragments covering 15,960 unique residues, with a median theropod placement posterior of 0.74, and reference sequences from 63 extant taxa. Relative to v2.8 this represents 667 additional theropod fragments and 3,180 additional unique residues. The first change is described in Sections 3.1 and 4.1, and the second in Sections 3.3 to 4.6.

8. Data availability

Library contents are confidential. Build manifests, commit identifiers and the audit dataset (PL-2024-11, PL-2024-12) are retained at the Shenzhen Research Center and are available to authorized research staff and, on request through Regulatory Affairs, to regulators with jurisdiction over released material. Library access is not available to third parties.

9. Acknowledgements

We thank Ruoxi Deng for re-examining the analytical records of the 14 crocodyliform-assigned entries admitted to the DN-2019-04 build, and the Research Computing team for restoring the 2019 build cache from cold storage. We thank Dr. Hao Lin for discussion of specimen attribution at intake, and the Food Safety & Quality function for its independent assessment of released material.

This work was funded internally. All authors are employees of Jurassic Meat and may hold equity. The audit was reviewed by the Ethics & Bioresponsibility Council (19 November 2024): reviewed; no concerns raised.

10. References

  1. Buckley M, et al. (2008). Comment on "Protein sequences from mastodon and Tyrannosaurus rex revealed by mass spectrometry." Science 319:33.
  2. Saitta ET, et al. (2019). Cretaceous dinosaur bone contains recent organic material and provides an environment conducive to microbial communities. eLife 8:e46205.
  3. Witmer LM (1995). The extant phylogenetic bracket and the importance of reconstructing soft tissues in fossils. In: Thomason JJ (ed.), Functional Morphology in Vertebrate Paleontology. Cambridge University Press, 19–33.
  4. Demarchi B, et al. (2016). Protein sequences bound to mineral surfaces persist into deep time. eLife 5:e17092.
  5. JM-TR-2013-001. Endogenous peptide fragments recovered from a Kem Kem Group theropod tooth. Jurassic Meat internal report.
  6. JM-TR-2016-003. Fragments from three additional specimens with documented provenance. Jurassic Meat internal report.
  7. JM-TR-2019-004. Library v2.0: a versioned multiple-sequence-alignment architecture for fragmentary ancient peptides. Jurassic Meat internal report.
  8. JM-TR-2021-005. Library v2.3: keratin-associated peptides and the chain-of-custody metadata schema. Jurassic Meat internal report.
  9. JM-TR-2023-006. Library v2.8: diagenetic modification in placement scoring. Jurassic Meat internal report.

Supplementary Table S1

Table S1. Theropod-targeted builds by year, filter state and use class, 25 March 2019 to 13 December 2024.
Year Theropod-targeted builds Mis-set Mis-set (%) Affected Affected: exploratory Affected: candidate Affected: construction-referenced
2019 (Mar–Dec) 188 9 4.8 3 2 0 1
2020 402 31 7.7 17 14 2 1
2021 517 44 8.5 26 22 3 1
2022 586 52 8.9 33 27 5 1
2023 641 47 7.3 31 25 5 1
2024 627 31 4.9 21 19 2 0
Total 2,961 214 7.2 131 109 17 5