Curator Guide

How to read this guide. Two kinds of statement appear here, and they are marked differently.

• Policy. Enforced by the code, or ratified in the project's design decisions or the extraction pipeline's rulings (as of 2026-09-27). Apply it as written.
• Draft — maintainer to confirm A recommendation for a judgement call that nothing yet settles. Use it as a default and note in the record when you do. Only the maintainer can turn a draft into policy.

Contents:

1. The review workflow
2. What makes a record
3. Conflicting evidence
4. Knockouts, expression hosts and engineered strains
5. Duplicates and renames
6. Unresolved organism names
7. The extraction rulebook: R-16 to R-22
8. Reject reasons
9. Visibility decisions are not curator decisions

• One inbox. The Curator Inbox lists every pending item across every queue: papers to review, papers needing a PDF, running extractions, community flags and suggestions, and change/removal requests. Counts always cover the whole backlog. Each row opens the page where the item is actually worked.
• Assignment is a to-do mark, not a lock. "Assign to me" puts a paper on your list and lasts until you release it. Anyone can still open and approve an assigned paper. Only a superuser can assign a paper to someone else or take it off a colleague's list. The live "open now" badge (the claim) is separate: it lapses minutes after a tab closes and only says someone is typing.
• Who may publish what.
  • Approving or rejecting an AI extraction of a published paper: curator or superuser.
  • Approving or rejecting a user submission (someone else's unpublished data): superuser only. Curators may read, correct and prepare it.
• Reject, withdraw and correct are different outcomes.
  • Reject: a pending extraction is turned away before anything is written.
  • Withdraw: an approved paper's records are deleted from the live database. Use it when the paper should never have been approved.
  • Correct (reopen, then approve a revision): the records were right in kind but wrong in detail. The earlier run becomes superseded. When the extraction producer re-delivers corrections for many papers at once, its revisions are grouped by sweep at Revision batches: read each row's changes there and approve or reject a selection. Each paper is approved separately, exactly as from its review page; one that fails (an unresolvable organism, say) stays waiting and is corrected on its own page.
  All three keep the reason you type on the record, so write a reason a later reader can check.

The unit of record is one actor tested against one plastic, in one paper (an interaction). Its evidence is one or more experiments, each with one or more observations.

2.1 Scope and provenance

• The paper must put an enzyme or a microorganism in contact with a plastic (R-13). Identity, cloning, expression or structure work with no plastic anywhere is out of scope, even for a famous enzyme.
• Only this paper's own Results count (R-15). A degradation claim cited from another study, or stated as background, produces no interaction, not an inconclusive one. Put the cited study forward for harvest instead.
• The actor must meet the plastic, not a compound derived from it (R-16). Oligomeric materials the vocabulary lists as plastics are plastics (R-17).
• An insect larva is a host, never an actor (R-18).
• Only microbial enzymes are actors (ratified 2026-09-27). The test is the source organism: animal and plant enzymes are not recorded; fungal and bacterial lipases and proteases are.
• A community is never an actor. A consortium, sludge or enrichment culture cannot be credited. An enzyme recovered from a metagenome with no cultured source is a protein actor from uncultured bacterium. An unnamed isolate is one organism, not a community: record it at the nearest node NCBI holds (fungal sp., unidentified bacterium) with the strain code in strain_designation.
• No predicted results. Allowed results are positive, negative and inconclusive. A computational prediction with no experiment is omitted, not downgraded.

2.2 Grading the result

• Clear zone or halo → positive (R-2). Do not downgrade an observed halo because plate assays are weak evidence. Put the caveat in notes instead.
• Mass loss below 2% → inconclusive (R-3). This applies to weight_loss and mass_remaining only, strictly below 2.0%, and only when the paper gives no verdict of its own.
• Growth on the plastic is inconclusive (R-14), when growth is the only evidence. Sole carbon source strengthens the claim, not the evidence.
• A value inside its own error bar is not activity (R-22).
• A measured negative is a finding. Record negatives, including wild-type negative controls. Never push a clean negative into inconclusive.
• Evidence strength has no field. There is no confidence grade on an interaction. Write caveats such as "no abiotic control" or "mixed-flask value" in the interaction's notes.

2.3 Evidence minimums

• Each paper keeps its own result. Interactions are per paper, so two papers that disagree about the same actor and plastic both stand. Public pages show the pair as mixed. Never edit one paper's result to agree with another's.
• The merge tool will not reconcile a conflict. A duplicate group whose members disagree positive/negative on the same plastic is refused, as is one with a softer disagreement (section 5).
• A paper that contradicts itself. inconclusive covers results that are ambiguous or contradictory, or that the paper itself calls inconclusive. If you grade on one reading, say which one in notes.

• A gene implicated only by knockout, complementation or overexpression is not a protein actor (R-6). The enzyme must be purified or recombinantly expressed and assayed against the plastic in its own right. Name the gene in the notes and keep the interaction at the organism level.
• A purified-enzyme result is not an organism result. Do not create a microorganism interaction from an enzyme-only result.
• A halo around a knockout-versus-wild-type pair is evidence about the organism, not about the gene.

Strain type (genetic_background). Every whole-organism record carries the kind of strain the paper tested: wild type, expression host, knockout / deletion mutant, engineered strain, or not classified (the default). The last three-but-one are constructs: their results stay visible on the paper and protein pages, labelled, and are never counted towards the species — not in its badges, its paper count, "First Reported", the plastic page, the statistics, the exports, the Taxonomy Tree or Annotate Taxa Table. Set it on the ingestion review form ("Strain type"), or for an existing record on Curation → Evidence checks.

• Expression host (for example E. coli BL21 or K. phaffii carrying a foreign PETase): record the activity on the enzyme (protein actor, enzyme_production_context = heterologous). Record a whole-cell interaction for the host only if the paper assays the host as a degrader, and mark it expression host.
• Engineered strain (for example Acetivibrio thermocellus DSM1313::pHK-LCC, a strain of the species itself carrying a construct): mark it engineered strain.
• Knockout or deletion mutant (for example Δpetase): its loss of activity is a control for the wild type, not a finding about the species. Keep the genotype in strain_designation and mark it knockout; it no longer turns the species' result mixed.