Prose names are PaperSpec, PaperDesk, PaperClaim and so on; commands, packages and paths stay lowercase. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014zNf4shD4zaCKLucZWPYL9
27 KiB
18. Layers and annotations
Everything people and models add to a document is an annotation, and annotations are grouped in layers.
scienting-layer/1 is the W3C Web Annotation Data Model with the anchor as a selector, so any Web Annotation tool can read it.
The JSON-LD context is context/scienting-1.jsonld, published at https://paperspec.scient.ing/context/scienting-1.jsonld.
18.1 One rule decides what is a claim and what is a layer
If a statement is true independently of this paper, it belongs in a claim store as a claim, where it is checked and leveled. If it is about reading this paper, it belongs in a layer, where it is attributed and signed but not checked.
18.2 The layer
A layer is a folder with layer.json and annotations.jsonl.
layer.json is an AnnotationCollection with the two contexts, format, an id of the form tag:scient.ing,2026:layer:<store>/<record id>, a label, a purpose, an audience (novice, student, expert or mixed), a visibility (private, team or public), a license (CC BY 4.0 by default when public), its creator, created, its project, the docs it annotates and its total.
A reader turns layers on and off, and the audience dial picks the level of glosses.
18.3 The annotation
Each line of annotations.jsonl is a W3C Annotation:
| Property | Holds |
|---|---|
@context |
["http://www.w3.org/ns/anno.jsonld", "https://paperspec.scient.ing/context/scienting-1.jsonld"]. |
id |
tag:scient.ing,2026:annotation:<store>/<record id> (4.6). |
type, layer |
Annotation, and the layer's ID. |
motivation |
A W3C motivation: commenting, describing, linking, questioning, assessing, replying and the others. |
creator, created, modified |
An agent {id, type, name}, with type Person or Software, and times. |
body |
One body (18.4). |
target |
A SpecificResource whose source is tag:scient.ing,2026:doc:<store>/<key> and whose selector lists an AnchorSelector holding a scienting-anchor/1, optionally followed by a W3C TextQuoteSelector for tools that do not know ours; or, for a reply, a SpecificResource whose source is another annotation. |
status |
proposed, confirmed, rejected or withdrawn. A model's annotation is proposed until a person confirms it. |
receipt |
For a confirmed annotation, the receipt that covers it. |
The TextQuoteSelector, when present, repeats the anchor quote's exact, prefix and suffix, with an empty string for a prefix or suffix the quote lacks.
It is present only when the anchor's quote is plain text; an anchor whose quote is hashed or keyed (6.3) has none, so the fallback never shows text the anchor withholds.
An annotation's content hash is SHA-256 of the canonical JSON of the annotation without status, receipt and modified.
A confirmed annotation is covered by its confirmer's receipt, which lists ["annotation", <record id>, <content hash>].
Replies are annotations that target annotations, so discussions are threads anchored to sentences.
18.4 Bodies
kind |
W3C type | Adds | Says |
|---|---|---|---|
gloss |
TextualBody |
audience |
An explanation at novice, student or expert level. |
claim-link |
SpecificResource |
source, pin, role |
That the span states, uses or disputes a claim. |
symbol |
SpecificResource |
source, pin, key, symbol |
That this occurrence of a symbol, the one whose key is key, is the symbol a symbol claim defines. |
comment |
TextualBody |
category |
A remark, opinion or interpretation. |
caveat |
TextualBody |
category |
A warning a reader should see. |
question |
TextualBody |
for, category |
A question for a reader, a group or the authors. |
exercise |
TextualBody |
exercise |
A task with a key and a rule (19). |
citation-context |
TextualBody |
reference, verdict, reason |
Why a sentence cites a reference, and whether it supports it. |
prerequisite |
SpecificResource |
source |
What to read or know first. |
crosslink |
SpecificResource |
source, selector |
That the span corresponds to a span in another document. |
trace |
SpecificResource |
source, selector, reference, note, verdict, proposer |
That the target's sentence cites reference on the strength of a passage of the cited document (18.5). |
structure |
TextualBody |
format, role, identity, written, label, accessions, terms |
That the span names, draws, gives parameters for or uses a chemical structure (18.6). |
sequence |
TextualBody |
alphabet, role, identity, start, label, organism, accessions, regions, variants |
That the span names, shows, uses or reports a variant of a protein, DNA or RNA sequence (18.7). |
data |
TextualBody |
data_type, role, dataset, quantities, subjects, method, label, accessions |
That the span presents data of a kind, what its fields measure and which molecules and sequences it is about (18.8). |
translation |
TextualBody |
language |
The span in another language. |
description |
TextualBody |
role |
A description of a figure, panel or table for readers who cannot see it: alt, a short text equivalent, or long, a full description. |
generated |
TextualBody |
generator |
Text a model wrote about the span; never mistaken for source. generator is {model, run} as a correction's model author is (11.2). |
A textual body has a value, and optionally a format (text/plain or text/markdown) and a language.
The source of a claim-link or symbol body is a claim in a PaperClaim claim store, written as PaperClaim writes a reference: claim: and the claim's tag, eight Crockford base32 capitals (the digits and the letters without I, L, O and U), as in claim:272XTX7C, or claim:<host>/<tag> for a claim in another store, <host> being a domain name.
A tag never changes, while a claim's slug, such as amber/k-bond, may be renamed, so a slug is never a reference.
The optional pin fixes the claim as the annotator read it: the same reference, @, and the first six lowercase hex digits of the claim's statement hash, as in claim:272XTX7C@02f47b.
common.schema.json defines both, as claimRef and claimPin.
A symbol body names its symbol twice, as the target's span and by key, the symbol's key in the math form scienting-math/1 (9.4), which a claim store gives the symbol a symbol claim defines: K_{r} for a span $K_r$.
So a reader finds every occurrence of a symbol across a library by its key, however its source spaced or braced it, and a claim store matches the annotation to its claim without parsing the span.
The span is still required, since the key says which symbol and the span says where; symbol, optional, is the symbol as the span writes it.
A comment, caveat or question body MAY carry a category, which says what it is about, from a fixed list:
category |
Says |
|---|---|
comment |
A general remark. |
context |
Background, related work, or what a reader should know to follow the span. |
critique |
A flaw in the method, the analysis or the argument. |
question |
A question. |
correction |
An error of fact, number, citation or reference in the paper. |
feedback |
A suggestion addressed to the authors. |
resource |
Data, code or a related paper. |
A producer that supports categories SHOULD set one on every comment, caveat and question; the property is optional only so that layers written before it stay valid.
The category is independent of the kind, so a caveat may be a correction and a question a critique.
A correction category says that a reader found an error in the paper; correcting the document's reading of it is a correction of 11.
The store's annotations table keeps the category beside the body's kind (13.4).
Nothing a model writes is ever stored as source text.
A model may propose a description, and like any annotation it stays proposed until a person confirms it; an accessible export uses confirmed descriptions only, and a figure with none has no description.
Other text a model writes about a span, such as a summary, is a generated annotation, which names its generator.
fixtures/records/layer/ holds a layer with one annotation of each common kind, each comment, caveat and question with its category, a claim link and a symbol that reference a PaperClaim claim and pin it, and the receipt that confirms some of them.
18.5 Traced citations
"As previously reported [12]": a reader who finds where the cited paper reports it marks the passage, and the trace body keeps the mark.
A crosslink says only that two spans correspond, in both directions, whereas a trace goes from one sentence's citation to one passage, and says which reference it follows, why, and who or what found it.
The annotation's target is a leaf or span anchor (6.4) on the citing sentence, which MUST be a leaf with runs, a sentence or a cell, and MUST cite the reference the body names: one of its cite runs lists reference (3.5).
reference is the ID of a reference node of the citing document, because a sentence may cite several references at once, as in [3, 12], and the target says only which sentence.
source is the cited document, written as a target's source is, tag:scient.ing,2026:doc:<store>/<key>, and never the citing document itself.
selector selects the passage in it: an AnchorSelector whose anchor holds the passage's pin (its node_hash) and, for part of a leaf, its span, then a TextQuoteSelector when the anchor's quote is plain (18.3).
The passage may be a container, such as a table, whose anchor has no quote.
| Property | Required | Holds |
|---|---|---|
reference |
Yes | The reference node of the citing document that the sentence cites and the trace follows. |
source, selector |
Yes | The cited document and the passage in it. |
note |
No | What the passage reports, in a person's or a tool's words. |
verdict |
No | supports, partly, does-not or cannot-tell, as a citation-context body has it (19.2). |
proposer |
When the creator is Software |
{model, run, score?, rank?}: the model and run that found the passage as a candidate, as a generated body's generator names them, the candidate's score as a decimal string, and its rank, from 1, among the candidates the run proposed for this citation. |
The state of a trace is the annotation's status.
A trace a tool proposes is proposed until a person confirms it, and rejected when a person turns the candidate down; a rejected candidate is kept, so that the tool is not offered it again.
A trace is disputed while a reply to it, an annotation whose target is the trace and whose motivation is assessing, is proposed or confirmed.
The dispute is computed from the replies and never written on the trace, so a reply can be withdrawn.
Whether a trace still leads where it did is also computed, from the current state of both documents. Resolve the target's anchor against the citing document and the body's anchor against the cited document (6.5). The first row that applies gives the outcome:
| Outcome | When |
|---|---|
stale |
An anchor is stale, on the citing side, the cited side or both. |
unlinked |
The sentence no longer cites reference: a relink removed the citation (11.7), or the reference was deleted. |
misresolved |
A resolve made the reference name a DOI or a document that is not the cited work. A resolution of none names nothing and contradicts nothing (10). |
changed |
Both anchors resolve and at least one is not current. |
current |
Both anchors are current, and the reference still resolves to the cited work or to nothing. |
A resolution names the cited work when its document is the cited document, another version of the same work (a key with the same work key, 2.4), or one the caller knows to be the work; and a doi resolution does when the work is known by that DOI (same_as, 2.4).
Besides the outcome, a resolution gives the passage's current address, key#node@pin[start:end], while its anchor resolves, and the citations inside the passage.
A chain of "as previously reported" is then followed one trace at a time: the reader opens the passage, and a citation in it that has a trace of its own leads on.
A trace breaks layer.trace.source when its selector has no AnchorSelector, an anchor is in another document than source names, source is the citing document, or the TextQuoteSelector does not repeat the anchor's plain quote.
It breaks layer.trace.target when the target is not a leaf with runs of the citing document selected by an AnchorSelector, or the sentence does not cite reference (22.4).
vectors/traces.jsonl resolves traces against the fixture packages after corrections, with every outcome and every refusal, and fixtures/records/traces/ holds a layer of four annotations on writer2027fixed-ynr7.v3 that cites hale2026harmonic-damb: a trace a model proposed to the paper's abstract and a person confirmed with a receipt, a candidate a person rejected, a person's trace to Table 1, and a reply that disputes the first.
18.6 Chemical structures
A paper writes about a molecule in words, and nothing in a converted document finds it there.
A structure annotation ties one to the words: its target is the span, as for any annotation, and its body holds the structure.
The words stay the paper's, the structure is one reader's reading of them, attributed and withdrawable, and a model's tie stays proposed until a person confirms it.
What a paper reports about a structure, such as the force-field constants of a bond, is about reading this paper (18.1), so it is an annotation here; a claim store holds a parameter once it is checked.
| Property | Required | Holds |
|---|---|---|
value |
Yes | The structure as a SMILES or a SMARTS, whichever format says. |
format |
Yes | chemical/x-daylight-smiles for a SMILES, chemical/x-daylight-smarts for a SMARTS. |
role |
Yes | What the span does with the structure: names it, depicts it (a figure, scheme or table draws it), parametrizes it (the span gives parameters for it) or uses it (a reagent, solvent or probe). |
identity |
Yes | What finds the same structure however it was written: the structure's InChIKey when it is a SMILES and the producer has one, else smiles: or smarts: and the first twelve hex digits of the SHA-256 of the value as written. |
written |
No | The structure as the paper writes it when that is not SMILES or SMARTS: R-OH, CT-CT. |
label |
No | A short name, such as NMA. |
accessions |
No | Entries of outside databases that the structure is: a list of {db, id}, with db in lower case (pubchem, chebi, cas). |
terms |
No | The bonds, angles, torsions and impropers the paper gives values for. |
A structure that is not one molecule, such as a group, a fragment with open positions or a pattern of atom types, is a SMARTS.
An open position is an atom that matches any atom, [*:1], with its label as the atom map number, so R-C(=O)-R' is [*:1]-C(=O)-[*:2] and R-OH is [*:1]-[OX2H], and the paper's own notation goes in written.
SMILES text that a toolkit would refuse, because a valence rule fails, is kept as a SMARTS rather than dropped, since a paper's drawing often is not a molecule a toolkit can sanitize.
Atom numbers.
Atoms are numbered from 0 in the order the value writes them, a bracket atom as one, so CC(=O)NC has atoms 0 to 4.
A hydrogen that is not written as an atom has no number: a term on a hydrogen needs [H] in the value.
The rule is of the text and not of any toolkit, which is what lets the numbers be checked without one.
A term has a kind (bond, angle, torsion or improper), atoms (2, 3, 4 and 4 atom numbers; an improper lists its central atom first), and optionally types (the force field's atom types, one for each atom), values (the parameters by the name the paper uses, each a decimal string as 1.2 writes numbers that carry meaning), units (a unit for each value) and source (where in the paper).
The atoms of a bond, angle or torsion are a chain, each bonded to the next, and the atoms of an improper are each bonded to the first.
A structure body breaks layer.structure.value when its value is not a SMILES or SMARTS as its format says: an unclosed bracket, branch or ring, a stray character, a bond with no atom, or a query (,, ;, &, ! or $ inside a bracket, or a or A) in a SMILES.
It breaks layer.structure.identity when its identity is not of the form its value gives, or, for a smiles: or smarts: identity, is not the hash of the value.
It breaks layer.structure.term when a term cites an atom the value does not write or cites an atom twice, when its atoms are not joined as its kind needs, when types has another length than atoms, or when units names a value that is not given (22.4).
The reference implementation, paperspec.structures.problems, reads atoms and bonds from the text with the standard library alone.
It does not check valence, aromaticity or the InChIKey, which need a chemistry toolkit; a producer with one SHOULD also check that the value sanitizes as a molecule when format says a SMILES, and that the InChIKey is the molecule's.
vectors/structures.jsonl reads the atoms and bonds of SMILES and SMARTS texts, mints identities, and breaks every rule, and fixtures/records/chemistry/ holds a layer of four structures on hale2026harmonic-damb: ethane, named and confirmed; water, named by a model and still proposed; the CT–CT pattern of Table 1, a SMARTS with the bond constants of both parameter sets as terms; and water again, as a SMILES that writes its hydrogens, with the OW–HW constants.
The two writings of water share one InChIKey, which is what finds them as one.
18.7 Sequences
A sequence annotation ties a protein, DNA or RNA sequence to the words that name it, show it or report a variant of it, as 18.6 does for a structure.
| Property | Required | Holds |
|---|---|---|
value |
Yes | The residues in capital letters, a gap as -. |
alphabet |
Yes | protein, dna or rna, which says what the letters mean. |
role |
Yes | names the sequence, depicts it (the span shows it, as text, a figure or an alignment), uses it (a construct, primer or probe) or reports something about it (a variant, a region). |
identity |
Yes | The alphabet, a colon and the first sixteen hex digits of the SHA-256 of the residues with the gaps taken out, so a sequence is found whatever gaps it was shown with and whatever its numbering. |
start |
No | The number of the first residue of the value in the molecule's own numbering; 1 when omitted. |
label |
No | A short name, such as HP35. |
organism |
No | The organism the sequence is from, as the paper names it. |
accessions |
No | Entries of outside databases that the sequence is, as in 18.6 (uniprot, genbank, refseq, ensembl, pdb). |
regions |
No | Stretches of the sequence the span is about: a list of {start, end, label?}, both ends included. |
variants |
No | Changes the span reports: a list of {position, ref, alt}. |
The alphabets allow the twenty amino acids, selenocysteine (U), pyrrolysine (O), the ambiguity codes B, Z, J and X and a stop (*) for protein; A, C, G, T and the IUPAC ambiguity codes for dna; the same with U for T for rna; and - for a gap in all three.
A residue's number is start plus the number of residues before it, gaps not counted, so a fragment of a protein is numbered as the whole protein is, and regions and variants use those numbers.
A variant says that the residue at position is ref in the value and alt in the variant; alt is a residue, * for a stop or - for a deletion.
A sequence body breaks layer.sequence.value when its letters are not those of its alphabet or are all gaps, layer.sequence.identity when its identity is not the hash of its residues, layer.sequence.region when a region starts after it ends or reaches outside the residues, and layer.sequence.variant when a variant is outside the residues, names a ref the value does not have at that position, changes a residue to itself, or has an alt its alphabet does not allow (22.4).
The reference implementation is paperspec.sequences.problems.
It needs no toolkit, since an alphabet is a set of letters; a producer that knows the molecule SHOULD also check the sequence against the accessions it names.
vectors/sequences.jsonl mints identities and breaks every rule, and fixtures/records/biology/ holds a layer of three annotations on the draft writer2027fixed-ynr7.v3, whose last sentence tests its charges on the villin headpiece subdomain HP35: the words that name it, the printed sequence with its three helices as regions, and a model's proposal of the N68H variant, in the numbering 42 to 76 that the headpiece uses.
18.8 Data
A paper reports data in a table, a figure or a passage, and a dataset (15) holds its values.
A data annotation says what kind of data its span presents, what the dataset's fields measure and which molecules and sequences it is about, so that data can be found by kind and by subject.
Its target is the span, as for any annotation.
The words and the dataset stay as they are, the kind is one reader's reading, attributed and withdrawable, and a model's tie stays proposed until a person confirms it.
Whether the values hold is a claim (18.1); this body says only what they are.
| Property | Required | Holds |
|---|---|---|
value |
Yes | The data in a few words: what it is and what it is about. |
data_type |
Yes | What kind of data it is, one of the kinds below. |
role |
Yes | What the span does with the data: reports it (the span is the table, figure panel or passage that gives it) or uses it (the span relies on data given elsewhere). |
dataset |
No | The ID of the dataset that holds the values, ds- and a key (15). |
quantities |
No | What the dataset's fields measure: a list of {field, name, unit?}, with the field of the dataset, a name in words (harmonic bond force constant) and the unit as the dataset spells it. |
subjects |
No | The molecules and sequences the data is about: a list of {kind, identity, label?}, with kind structure or sequence and the identity of a structure (18.6) or sequence (18.7) body. |
method |
No | How the data was obtained, in the paper's words: isothermal titration calorimetry, DFT, B3LYP/6-31G*. |
label |
No | A short name. |
accessions |
No | Entries of outside databases or vocabularies that the data is, or is of: a list of {db, id}, with db in lower case (zenodo, geo, edam, chmo). |
data_type |
Covers |
|---|---|
abundance |
Counts, expression levels or concentrations measured in samples. |
benchmark |
How well a method or model does against a reference or against other methods: errors, scores, run times. |
binding-affinity |
How tightly molecules bind: dissociation constants, inhibition constants, IC50, binding free energies. |
dose-response |
A response as a function of dose or concentration. |
energy-profile |
Energy along a coordinate: scans, potentials of mean force, reaction paths. |
force-field-parameters |
Constants a force field or model uses for bonded and non-bonded terms. |
geometry |
Lengths, angles, torsions, distances or deviations of structures. |
kinetics |
Rates, rate constants, half-lives and barriers. |
physical-property |
A property of a substance: density, melting point, pKa, logP, solubility. |
sequence-property |
Values for the positions of a sequence: conservation, disorder, fitness scores. |
simulation-observable |
A quantity computed from a simulation: distributions, time series, order parameters. |
spectrum |
A spectrum or its peaks: NMR, infrared, Raman, ultraviolet-visible, mass. |
thermodynamics |
Enthalpy, entropy, free energy, heat capacity, or a melting temperature of a process. |
variant-effect |
What a change to a sequence or a structure does to a property. |
other |
Data none of these describes; the value says what it is. |
The list is fixed by this chapter, and a new kind is a minor change (0.4).
Where a kind has an outside term, accessions carries it, so a producer is not limited to this list when it says what the data is.
A dataset's origin (15) says where its values came from, so the body does not repeat whether they were reported, digitized, computed or measured.
A data body breaks layer.data.type when its data_type or its role is not one the body allows, layer.data.quantity when two quantities are for one field, layer.data.subject when a subject's identity is not of the form its kind gives or a subject repeats another, and layer.data.dataset when dataset is not a dataset ID or, for a checker given the dataset's descriptor, is not that dataset, or a quantity names a field the dataset lacks or gives it a unit other than the dataset's (22.4).
The reference implementation is paperspec.data.problems, which takes the descriptor as an optional second argument.
It compares units as strings and needs no unit library; a producer SHOULD also check that each unit parses and that its dimension fits the quantity it is given for, and that each subject's identity belongs to a structure or sequence annotation of the same layer or another.
vectors/data.jsonl breaks every rule, and fixtures/records/data/ holds a layer of three annotations on hale2026harmonic-damb: the caption of Table 1, tied to its dataset as force-field parameters for ethane, water and the CT–CT pattern with its force constant and equilibrium length; the caption of Figure 1a, tied to the dataset read from the scan as an energy profile of ethane; and the sentence that introduces Table 1, tied by a model and still proposed.
The subjects' identities are those of the annotations in fixtures/records/chemistry/.