How ANARCI results compare to the bench

You have an ANARCI result. Here is what it supports.

• ⊘ Known issue, stated first because it changes what you should do with the result. The platform's anarci_number has returned framework residues labelled as CDRs while the job reported status SUCCESS. The cause was a boundary table named for IMGT that held Chothia/Kabat-style ranges; on a heavy chain it shifted CDR-H3 by 12 positions, and the measured overlap between the delivered CDR-H3 and the correct one was 0 of 15 residues. A second defect in the same tool meant assign_germline was never passed to ANARCI, so every germline call was silently null. Both are fixed in the compute repository and the fixes are present at its remote trunk; neither this page nor the docs repository can establish that the fixed image is the one serving your job. Re-verify any gene call or any numbering you intend to act on. See what the figure doesn't cover.
• This is a classification tool, so the honest comparator is not an experiment. A residue number is not a physical quantity. There is no RMSD-style ground truth here and no assay that returns "position 27": the reference is itself a curated annotation standard, and when two tools disagree the published conclusion is sometimes that nobody can say which is right.
• Numbering and germline calling are two different reliability questions, and they do not have the same answer. On pure IMGT numbering ANARCI agrees with the other tools on roughly 98% of therapeutic sequences. On germline V-gene assignment from amino acids it scored 65.42% against a competitor's e-value oracle. Treat the numbering as solid and the gene call as a hypothesis.
• IMGT is ANARCI's native coordinate system; every other scheme is a translation. Its HMMs are built in the 128-position IMGT alignment and the result is converted into whichever scheme you asked for. On this platform that shows through unevenly, and the result tells you how: cdr_sequences are cut with the boundary table of the scheme you asked for (imgt, chothia, kabat and aho each have one), but framework_sequences are IMGT-bounded under every scheme, because IMGT is the only scheme with a framework table here. wolfguy has neither, so both its region kinds fall back to IMGT. Read cdr_numbering_scheme and framework_numbering_scheme per sequence — they are the exact answer — and region_numbering_scheme only when both agree, which is why it is null on a chothia, kabat or aho result. ⚠︎ If your result carries no cdr_numbering_scheme at all, it predates this and every region in it is IMGT-bounded; see the known issue above.
• Confirm a numbering or a gene call with independent re-annotation — IgBLAST, IMGT/HighV-QUEST, or a second numbering tool — and, for anything consequential, manual expert curation against the conserved anchors (IMGT Cys23, Trp41, Cys104, and the position-118 Trp/Phe).

Jump to agreement with wet lab or against other in silico methods. The general framing for every tool is in reading an in silico result against the bench.

How in silico is used currently

Numbering is infrastructure, not a result. Nobody runs ANARCI to learn something about an antibody; they run it so that every other tool can refer to the same residue by the same name. That changes how the field validates it: numbering tools are graded against each other and against curated reference annotations, never against a bench measurement, because there is no bench measurement of a residue number to grade against.

Three uses, all of them upstream of the question you actually care about:

• Coordinate assignment. Position 27 in one antibody has to mean the same structural place as position 27 in the next, or no cross-antibody comparison is meaningful. This is what the major resources use it for: the Observed Antibody Space database numbers its contents with ANARCI and discards what ANARCI cannot process, and structural pipelines gate on it the same way.
• CDR delineation for downstream tools. The CDR and framework split is the substrate for humanness scoring, developability profiling and paratope prediction. An error here is not visible as an error — it propagates as a confidently wrong answer in the next tool, which is exactly what the known issue above did.
• Germline assignment for mutation counting. Naming the closest V gene is how you measure how far a sequence has drifted from germline. On this platform that is the weakest axis of the tool, and it is a side-product of the numbering HMM rather than the thing ANARCI was built to do.

What numbering is not used for is deciding anything on its own. It is a precondition for the decision, which is why a silent numbering error is expensive.

Against other in silico methods

No independent third-party benchmark of ANARCI exists. Every published ANARCI number comes either from ANARCI's own group or from the authors of a competing tool, and each one has a named circularity or oracle-choice problem. That is the single most important caveat in this section, and it holds for every figure below. Note also that the field's standard benchmark of repertoire-annotation tools — the one that compares IMGT/HighV-QUEST, IgBLAST and MiXCR — does not include ANARCI at all, so its figures describe neither this tool nor this task.

On numbering, ANARCI agrees with everyone — and the scheme you pick changes how much. The 1,274-sequence therapeutics set was numbered with AbNum, AbRSA, ANARCI, AntPack and RIOT, with AbRSA then dropped from the comparison because "it caused most conflicts (CDR numbering does not comply with scheme definitions)." Comparing the rest in IMGT — ANARCI, AntPack and RIOT, since AbNum does not support IMGT — gave "only 23 cases where tools rendered conflicting numberings." Comparing in Kabat, Chothia and Martin — the same three plus AbNum — gave "~200 conflicts in numberings."

⚠︎ Those two counts are not a clean ratio, and this page will not present one. Three things differ between them: the Kabat-family comparison adds a fourth tool, it pools three schemes rather than one, and the paper attributes the majority of its conflicts to "differences in CDR numbering heuristics or AbNum assigning the last residues of light chains to 106A instead of 107" — so part of the increase is the extra tool's known quirk rather than scheme difficulty. What the counts do support, directionally and with the confounds stated, is that IMGT is where annotation tools agree and the Kabat-family schemes are where they stop agreeing — which is the published shape of the risk the platform's own region caveat sits on.

The mechanism named for that asymmetry is ANARCI's, and it is the same mechanism that produces the platform's region caveat: "While renumbering CDRs, ANARCI reorganizes residues from the loop, with two flanking positions on each side. IMGT definitions of CDR boundaries are used for this regardless of the selected numbering scheme." And: "the smoothing logic applied by ANARCI in IMGT space that is translated further to the user-selected schemes can lead to issues in Kabat-like schemes." The same authors also report that where numberings diverge on sequences far from germline, "we were unable to determine which tool is more correct or less wrong" — which is what the absence of a ground truth looks like in practice.

The scheme definitions themselves disagree by design, and the spread is large. From the statistical and structural evaluation of the schemes over 124,317 variable domains:

⚠︎ The first four columns are in Kabat numbering and the IMGT column is in its own, so the IMGT cells cannot be compared arithmetically with the others — only within the first four. On that basis, summing the six loops gives Kabat 56 position labels, AbM 54, Chothia 47 and Consensus Chothia 34 (derived here by counting the published ranges, insertion codes excluded). The same antibody therefore has a "CDR" that is 22 position labels longer under Kabat than under Consensus Chothia.

Two loops carry almost all of it. HCDR2 is 16 positions in Kabat and 5 in Chothia — a factor of 3.2 on the same loop. And HCDR1 barely overlaps at all: Kabat's H31–H35 and Chothia's H26–H32 share only H31 and H32, two positions out of the ten either scheme ever calls CDR. That is why a CDR sequence quoted without its scheme is not a usable annotation.

⚠︎ Two published versions of this table disagree, and this page follows the majority. A 2025 review by the same first author prints Chothia HCDR1 as H26–H35 — identical to AbM — where the 2024 evaluation prints H26–H32. H26–H32 is corroborated independently, so the table above uses it; if you are reconciling against the 2025 review, that cell is the discrepancy. The 2025 review is also where the statement comes from that "there is not always a strong consensus among different numbering schemes when applied to the same amino acid sequence input," along with a re-cited estimate that "approximately 10% of the entries annotated using the Kabat numbering scheme exhibit errors or inconsistencies." The 2024 evaluation's own conclusion is the cleanest summary available: "we found that no one numbering scheme fully encompassed the most variable residues in all CDRs, and the schemes vary in their inclusion of adjacent framework-like residues."

On germline calling, ANARCI is last of three, and IgBLAST is not in the same comparison. On the therapeutics set: V gene RIOT 96.94%, AntPack 80.04%, ANARCI 65.42%; J gene 97.72% / 77.99% / 78.94%; V allele 96.86% / 76.35% / 62.17%. ⚠︎ The wall-clock figures in that same table are not measured on that set. Its authors state that the therapeutics set was too small for reliable timing — "Therefore, to measure the processing efficiency, we constructed another amino acid dataset by translating NGS sequences" — so the 00:00:32 for AntPack, 00:02:40 for RIOT and 00:18:01 for ANARCI are a numbering run over 362,180 translated sequences, not the 1,274 accuracy sequences. Read the two halves of that table as two experiments. ⊘ And do not pair any of it with IgBLAST's 99.88%. IgBLAST was scored in the same paper but on nucleotide input and on its own dataset; amino-acid and nucleotide germline assignment are not the same task, and no like-for-like ANARCI-versus-IgBLAST germline comparison has been published.

On speed, ANARCI is the slow option by a wide margin. Numbering 3,492 non-redundant PDB sequences across five repeats: "AntPack is faster than ANARCI by > 60x … numbering all 3492 sequences in slightly under 0.5 seconds, while ANARCI takes 35—45 seconds and AbRSA 12—13." At scale: "it took ANARCI running four separate jobs > 15 days to number 100 million sequences." ANARCI's own group puts it at ~75,000 sequences per minute on 32 CPUs without gene calling, against ~90,000 for a GPU successor.

On robustness to heavy mutation, ANARCI sits between two competitors. On a simulated set of hypermutated chains (18,720 heavy and 7,020 light in the subsample the comparators were run on), at 50 introduced mutations — roughly half the sequence changed — the share still numbered at all was ANARCI 82.7% heavy and 90.2% light, against AbNum 60.7% / 74.5% and AbRSA 90.6% / 96.3%. ⚠︎ Tested on ANARCI v1.3, which is not the version running here.

On unusual formats, ANARCI fails outright rather than degrading. "ANARCI failed to return a numbering for all 456 VNAR sequences" — shark single-domain receptors — while a competitor failed on 99 of 456. On 4,363 VHH nanobody sequences, by contrast, overall disagreement between methods was low at 143 sequences. And ANARCI will not number a T-cell receptor at all when restricted to antibody germline HMMs.

One axis where ANARCI is the best tool tested, reported by a competitor. Given 7,459 non-antibody sequences selected as the ones most similar to antibody consensus (over 45% identity), "ANARCI identified all the sequences as non-antibody," while AbNum misclassified a viral TCR-beta-like protein as an antibody light chain. A zero false-positive rate on the hardest available decoy set, published by a rival, is the most credible favourable figure ANARCI has.

Against its own successor. A GPU model from the same group replicated ANARCI's numbering exactly on 99.82% of ~9.6M held-out heavy, 99.59% of ~1.5M kappa and 99.54% of ~2.4M lambda OAS sequences. ⚠︎ That is agreement with ANARCI as the reference, so by construction it cannot show ANARCI being wrong. On the deliberately hard subset where two ANARCI builds disagreed, agreement with current ANARCI fell to 52.4% of heavy sequences, with most divergence in CDR-H3 around position 118.

Parkinson J, Wang W. For antibody sequence generative modeling, mixture models may be all you need. Bioinformatics. 2024;40(5):btae278. doi: 10.1093/bioinformatics/btae278

Zhu Z, Olson KS, Magliery TJ. 50 Years of Antibody Numbering Schemes: A Statistical and Structural Evaluation Reveals Key Differences and Limitations. Antibodies (Basel). 2024;13(4):99. doi: 10.3390/antib13040099

Zhu Z, Ashrafian H, Tabrizi NM, et al. Antibody numbering schemes: advances, comparisons and tools for antibody engineering. Protein Engineering, Design and Selection. 2025;38:gzaf005. doi: 10.1093/protein/gzaf005

Dondelinger M, Filée P, Sauvage E, Quinting B, Muyldermans S, Galleni M, Vandevenne MS. Understanding the significance and implications of antibody numbering and antigen-binding surface/residue definition. Frontiers in Immunology. 2018;9:2278. doi: 10.3389/fimmu.2018.02278

In silico vs. wet lab

The standard approach this tool displaces is not an assay — it is sequencing plus a human. You obtain the variable-region sequence by Sanger or NGS repertoire sequencing, then an expert curates the numbering and the V/J gene call against a reference directory, which is how the Kabat and IMGT reference sets were themselves built.

The primary literature states this outright, so it is not an inference of this page's. On the methods ANARCI belongs to:

"All these methods rely on either a representative consensus sequence or an IMGT germline reference set compiled using data available at a specific point in time. Germline references are continually being updated and modified as new species, genes, alleles and sequence types are discovered. This means the ground truth needed to number an antibody is not static, and the numbering returned may differ depending on the timestamp of the reference set."

That makes this page structurally different from a docking or structure page, and the difference is worth being blunt about: sequencing returns the sequence, not the numbering. No experiment returns a residue number. The comparator for the numbering step is a curated annotation standard — IMGT's reference directory and the tools that implement it — and the comparator for the gene call is an alignment against a germline database that is itself a curated, versioned artifact. Agreement is therefore measured as concordance against a curated reference, not as error against a measurement. Published practice reflects this: the ANARCI source paper reports only how many sequences it could number, never how many it numbered correctly.

Read the agreement column by cell, never down the column. A coverage rate (did it return anything), a tool-agreement rate (did two programs match) and a gene-call accuracy against an e-value oracle are three different questions, and none of them is an error rate against an experiment. What the table supports is this: seconds instead of expert-days, at an agreement with other annotation tools that is high for numbering and materially lower for the gene call — with the platform's own delivery history as a separate and larger caveat than anything in the literature.

Two asymmetries matter. Sequencing can be wrong in ways numbering cannot fix: the ANARCI paper attributes its own numbering failures to "very unusual insertions or deletions that may be a result of sequencing errors," so a numbering failure is sometimes a sequencing failure wearing a different coat. And a curated standard can be revised, which means an annotation that was correct when made can become non-current without anything about the antibody changing.

Agreement with wet lab

There is no wet-lab concordance figure for ANARCI, and that is a property of the task rather than a gap in sourcing. No experiment measures a residue number, so nothing here can be reported as agreement with a measurement. What the literature does contain is concordance against curated references and against other tools, and the spread between those figures is the useful part of this page.

Coverage, author-reported. The source paper's only benchmark is a throughput and coverage run: ANARCI numbered a set of 1,936,119 VH sequences in three hours of wall clock on 32 cores, and "All but 9560 sequences were successfully numbered." That is a 0.49% no-output rate, or 99.51% coverage — derived here from the paper's two counts, and it is a coverage rate, not a correctness rate. The paper reports no accuracy figure of any kind, and no comparison against Abnum, DomainGapAlign, PyIgClassify or IgBLAST. Any claim that "ANARCI is X% accurate" sourced to the 2016 paper is unsupported.

Against curated PDB annotations. On a benchmark of 983 heavy and 833 light non-redundant PDB chains, a competing tool's authors report ANARCI producing no output for 8 heavy and 0 light chains, with 2 mismatches against the pre-annotation in each of the Kabat and Chothia schemes. ⚠︎ This figure is close to tautological and should not be read as an independent error rate. The paper states that "The pre-annotated numbering was obtained using AbNum and ANARCI," with 13 conflicting chains resolved by hand from 3D structure alignments — so ANARCI is part of its own ground truth here, and its 4 mismatches can only come from those 13 manually re-annotated cases.

Against an independent tool on real repertoire data. The cleanest available number. On 1,492,031 paired heavy–light OAS sequences, each numbered with ANARCI in IMGT and cross-checked against IgBLAST annotations, any CDR-boundary discrepancy excluded the chain: 9,822 heavy and 6,414 light chains were dropped. That is 0.658% of heavy and 0.430% of light chains on which the two tools disagreed about where a CDR starts or ends — derived here from the published counts. The measurement was taken as a data-cleaning step with no tool being promoted, which is why it is the most trustworthy agreement figure on this page.

Against itself, across build dates — the figure with no circularity problem at all. Because the reference is a curated standard that keeps changing, ANARCI's answer changes with it. Renumbering all of OAS with an ANARCI compiled in December 2023 and comparing against the numbering stored when each dataset was first processed in 2018–2022 produced disagreement on ~3.6M heavy sequences (3.8% of all heavy) and ~0.9M light (1.2% of all light), of which 530,804 heavy and 410,148 light differed at more than four position labels. The authors attribute this to the germline reference drift directly: "the addition of new alleles to the IMGT reference set may result in alternate numbering of certain sequences between versions of ANARCI built with, and without, this updated information." This needs no external oracle — the same software, the same sequence, two build dates, two answers — and it is reported by ANARCI's own group, which strengthens rather than weakens it.

Germline gene calling, the weak axis. On 1,274 therapeutic antibody amino-acid sequences with heavy and light scored separately, ANARCI's calls matched the reference on 65.42% of V genes, 78.94% of J genes, 62.17% of V alleles and 78.60% of J alleles. ⚠︎ Read that with the oracle in view. Ground truth was defined by aligning each sequence to every germline allele with striped Smith–Waterman and taking the lowest e-value, and the authors say plainly why ANARCI scores low against it: "the surprisingly low germline assignment accuracy of ANARCI could be attributed to its germline assignment method: iterating over the alignment to select the gene with the highest sequence identity. We advocate for assigning genes using e-values." So the 65.42% measures disagreement with an e-value oracle chosen by the competing tool, not demonstrated biological error. It is still the only published germline-accuracy figure for ANARCI, and the direction is consistent with the gene call being a by-product of the numbering HMM rather than a designed feature.

Dunbar J, Deane CM. ANARCI: antigen receptor numbering and receptor classification. Bioinformatics. 2016;32(2):298-300. doi: 10.1093/bioinformatics/btv552

Li L, Chen S, Miao Z, Liu Y, Liu X, Xiao ZX, Cao Y. AbRSA: A robust tool for antibody numbering. Protein Science. 2019;28(8):1524-1531. doi: 10.1002/pro.3633

Greenshields-Watson A, Abanades B, Deane CM. Investigating the ability of deep learning-based structure prediction to extrapolate and/or enrich the set of antibody CDR canonical forms. Frontiers in Immunology. 2024;15:1352703. doi: 10.3389/fimmu.2024.1352703

Greenshields-Watson A, Agarwal P, Robinson SA, et al. ANARCII enables alignment-free antigen receptor numbering using a generalised language model. Communications Biology. 2026;9(1):1085. doi: 10.1038/s42003-026-10186-z

Dudzic P, Janusz B, Satława T, et al. RIOT—Rapid Immunoglobulin Overview Tool—annotation of nucleotide and amino acid immunoglobulin sequences using an open germline database. Briefings in Bioinformatics. 2024;26(1):bbae632. doi: 10.1093/bib/bbae632

Smakaj E, Babrak L, Ohlin M, et al. Benchmarking immunoinformatic tools for the analysis of antibody repertoire sequences. Bioinformatics. 2020;36(6):1731-1739. doi: 10.1093/bioinformatics/btz845

What the figure doesn't cover

• ⊘ The platform's own delivery history is the largest caveat on this page, and it is a known issue rather than a literature limitation. anarci_number shipped CDR and framework regions sliced out of IMGT numbering with Chothia/Kabat-shaped boundaries, under status SUCCESS, to every caller. The offsets were CDR-H1 −1, CDR-H2 −5, CDR-H3 −12 and FR-H4 −15 positions; measured against delivered output the overlap was CDR-H1 4 of 8 residues, CDR-H2 1 of 8 and CDR-H3 0 of 15. The shift was a whole region: what was labelled FR-H4 carried the conserved Cys-Ala-Arg motif that IMGT places at position 104, immediately before CDR-H3. Separately, assign_germline was never forwarded to ANARCI, whose own default is off, so germline_vh and germline_vl were always null despite the platform defaulting the parameter to true. Both are fixed in the compute repository and present at its remote trunk; whether the fixed image is serving your job is not established here, and the realized-harm population across past jobs has not been swept. Treat any stored ANARCI result predating this as suspect and re-run it.
• ⊘ Frameworks are IMGT-bounded even when you ask for another scheme — CDRs are not. This is current behaviour, not a bug, and the two region kinds have different answers. cdr_sequences are cut with the requested scheme's own boundary table where one exists (imgt, chothia, kabat, aho). framework_sequences have a boundary table in IMGT only, so a non-IMGT request gets a second IMGT numbering pass for them and they come back IMGT-bounded whatever you asked for. numbered_positions are in your scheme throughout. Per sequence, cdr_numbering_scheme and framework_numbering_scheme state which scheme produced which kind; region_numbering_scheme names a single scheme only when both agree and is null otherwise, so a null there is the signal that the two kinds came from different schemes, not a missing value. wolfguy has no table of either kind and falls back to IMGT for both. ⚠︎ The canvas card does not surface region_caveat. The card shows the scheme, the success count and the failure count; the caveat reaches the model in the payload but is not displayed, so a reader looking only at the card will not see it. If the IMGT pass finds no domain, the regions are omitted entirely and the caveat explains that instead. ⚠︎ A result carrying no cdr_numbering_scheme predates this split, and in it every region — CDRs included — is IMGT-bounded regardless of the scheme named on it; that older shape is the known issue described above, and this change is in the compute repository but not established as deployed, so check the field rather than the date.
• No J gene is returned, and no identity score. ANARCI exposes both a V and a J germline call, plus an identity value for each, in its alignment details. The platform reads only the V allele: there is no J-gene field in the output, and the identity figure is discarded. So the gene call arrives with no confidence attached, which is precisely the number you would need to judge a 65%-accurate call. Other pages in this cluster describe the parameter as assigning "V and J" genes; only V is delivered.
• Only the first domain of each sequence is numbered. An scFv or any multi-domain construct returns one domain's numbering, silently. Submit heavy and light as separate sequences.
• The germline database vintage is fixed at deploy time and never reported. Gene calls and, as the self-inconsistency figure above shows, numbering itself depend on which IMGT germline reference the HMMs were built from — and that reference changed 221 times in under nine years. ANARCI's own authors describe the mechanism and its reach: "each time a user installs ANARCI, they are required to compile the HMM models for all reference germlines in IMGT," and "Small changes in the reference sets, e.g. the addition of new alleles to IMGT data, can lead to alternate predictions which can lead to large differences in numbering, including changes to CDR delineation." They also report that "the addition of new species to ANARCI reference sets, for example, to enable numbering of Rhesus Macaques, can in some cases corrupt the numbering of human sequences" — so a reference update made for an unrelated species can move your answer. The output does not state which vintage produced your call, so two results from different deploys are not guaranteed comparable.
• Published figures test versions that are not this one. The robustness curve tested ANARCI v1.3; the coverage figure is from the 2016 release; the germline-accuracy figure used HMMs the benchmarking authors rebuilt themselves from an OGRDB germline set rather than the platform's. A figure quoted without its version and its germline set is not usable here.
• Every ANARCI accuracy number in print has a circularity or an oracle-choice problem. The PDB benchmark used ANARCI to build its own ground truth. The successor comparison uses ANARCI as the reference. The germline figure uses the oracle the competing tool optimises. The three figures closest to clean are the timing comparison, the zero false-positive rate on non-antibodies, and the 0.658%/0.430% CDR discordance against IgBLAST.
• Agreement is not correctness, and the field says so. Where tools disagree on sequences far from germline, the published conclusion was that the authors could not determine which tool was less wrong. Two tools agreeing means they share a convention, not that the convention is right — and the conventions themselves disagree badly: 11 residues on HCDR2 alone, and on HCDR1 the Kabat and Chothia definitions share only two of the ten positions either one calls CDR. The published conclusion is that no scheme captures all the variable residues, so "which CDR definition" is a question with no correct answer, only a convention you have to declare.
• No figure here was measured on this platform. This section aside, everything above is literature about upstream ANARCI. No published study evaluates the platform's configuration, and the platform's own history on this tool is worse than the literature's worst figure.
• Nothing about numbering validates what comes next. A correct CDR boundary is a precondition for humanness scoring, humanization and developability, not evidence that those outputs are right. Each carries its own translatability question.

Go / no-go

Go — act on the result when:

• You asked for the default imgt scheme and you want the numbering as a coordinate system for downstream tooling. This is the configuration with no scheme-translation caveat, the one the literature's agreement figures describe, and the one whose region boundaries are the canonical IMGT ranges (FR1 1–26, CDR1 27–38, FR2 39–55, CDR2 56–65, FR3 66–104, CDR3 105–117, FR4 118–128).
• You have checked the conserved anchors before using the output. Position 23 and 104 should be cysteines, 41 a tryptophan, and 118 the J-region Trp or Phe. These are scheme invariants rather than sequence facts, so they are a free correctness check that catches exactly the failure mode that shipped — and you should run it on every result until you can confirm which image served it.
• You want to know whether a sequence is an antibody variable domain at all. This is ANARCI's most credible published strength: zero false positives on 7,459 near-antibody decoys, reported by a competitor.
• You are working with conventional VH/VL or VHH sequences, one domain per submitted sequence, in modest numbers. CPU-only, 1 credit, a couple of minutes.

No-go — do not act on the result when:

• You intend to act on a germline gene call without re-verifying it. It is 65% V-gene by the only published measurement, it arrives with no identity score, no J gene, and it was returning null for every call until recently. Confirm with an independent annotator before it reaches a biosimilar, humanization or freedom-to-operate decision.
• You requested a non-IMGT scheme and intend to read the framework sequences, or cannot confirm which build served you. Frameworks are IMGT-bounded under every scheme, the card will not tell you, and the Kabat-family schemes are where annotation tools disagree roughly eight times as often as in IMGT. The CDRs are cut in your scheme on a current build — but a result with no cdr_numbering_scheme is not one, and in that case the CDRs are IMGT-bounded too. Read cdr_numbering_scheme, framework_numbering_scheme and region_caveat in the payload, or re-run in imgt and convert deliberately.
• You are holding a stored result and cannot establish which build produced it. Given the known issue, an old SUCCESS is not evidence of a correct answer. Re-run it.
• Your input is a VNAR, a TCR, an scFv, or any non-canonical format. ANARCI returned nothing for all 456 VNARs tested, will not number a TCR under antibody HMMs, and numbers only the first domain of a multi-domain sequence.
• Your sequence is heavily mutated or truncated. Numbering success falls to about 83% for heavy chains at 50 mutations, and divergence between tools concentrates at truncated ends and in long CDR-H3s — the regions you most likely care about.
• You need the numbering to be reproducible across time. The same software at two build dates disagreed on 3.8% of all heavy sequences in OAS. Pin and record the version alongside the result if reproducibility matters.
• You plan to cross-check only with IgBLAST on this platform. It is a reasonable second opinion — the 0.658% CDR discordance figure is exactly that comparison — but it is a sibling tool with its own history of reporting SUCCESS over an empty result, so for anything consequential verify off-platform against IMGT/HighV-QUEST or by hand.

Caveats on reading this page: every figure above is quoted from the cited source with its original metric and sample, and each is labelled author-reported or independent. Where a percentage was computed from published counts, the page says so. Figures from different studies are not comparable, because reference sets, germline databases, input types, schemes and tool versions differ between them. Where no figure exists — and for the central question of agreement with a wet-lab measurement, none can — this page says so rather than substituting a figure from a different tool or a different task.