Skip to content
BIOPOLYMER Intelligence

Overview / Methods

How we verify

Four stages, each a different evidence level. The site labels which level an item has reached. This site follows development; it does not support purchasing or qualification decisions.

Editorial status

From the topic owner's input sheet (6 Oct 2026). Readers: Material Innovation team; possibly the Raw Material team. Cadence: monthly; default window: previous calendar month; language: English.

  • OPEN G4 — Technical reviewer name
  • OPEN G4 — Issue delivery date
  • OPEN G4 — Distribution channel
Decided scope (7)
  • G2 Global scan; no region excluded.
  • G3 Keep both early research and pilot/industrial items; no cost or LCA required for significant early research.
  • P1 PHA, PLA and PEF are in scope and the core focus, including their blends and copolymers.
  • P2 Prioritise genuinely new chemistry; label other novelty types separately.
  • P3 Modified natural polymers qualify. Drop-ins, mass-balanced products and biofilled fossil-matrix composites go to a labelled scope-review route.
  • P4 No numerical pass/fail limits.
  • P5 Newsletter: two A4 pages all-inclusive (visuals and references), plain English, at least one verified structure.

1 · Discovery

A monthly GitHub Actions job (3rd of each month, for the previous calendar month) queries Crossref, OpenAlex and Europe PMC (filtered by publication date) and date-bounded Google News RSS in English, German, French, Chinese and Japanese, with route-specific queries from data/routes.yaml — including searches for delays, cancellations and closures. Preprints and patents are only ever leads. A keyword gate in code drops records that do not mention polymer terms before any model call. Discovery records are not evidence.

2 · Machine triage (DeepSeek flash)

Records go in batches to DeepSeek flash (deepseek-flash) with a versioned prompt (prompts/triage.md). The model assigns a route, a novelty type, relevance, signal type, region, a one-sentence summary and a supporting quote. The code then decides:

  • the quote must appear verbatim (whitespace-normalised) in the title or abstract, or the summary is discarded;
  • route and novelty type must be defined ids;
  • every number in the summary must appear in the quote as a whole number (so “12” does not match inside “2012”);
  • records dated after the cutoff are rejected;
  • leads below relevance 0.6 are kept in the data but hidden.

Each run first checks that the configured model is listed by the API and stops otherwise — there is no silent fallback to another model. Each run has a call budget (MAX_LLM_CALLS_PER_RUN, default 60). Each record is hashed (DOI or normalised URL + title) and never re-sent once triaged. Records the budget did not reach stay “triage pending”. Token use is logged per run in data/runs/.

Last run 2026-10-20261020T183350Z-d30b0f: 1229 records discovered, 337 dropped by the keyword gate, 887 sent to the model, 60/60 calls (222,553 input + 196,922 output tokens), 168 leads shown, 292 pending. Cost: not estimated (no price configured). Total calls to date: 120.

An optional draft model (DEEPSEEK_DRAFT_MODEL) may draft newsletter text from evidence-record fields only (max 3 calls per issue); code re-checks that every number in a draft exists in the record. It is unset by default, so drafting also uses flash.

3 · Evidence check

An editor shortlists leads and fills a 12-point evidence record from the primary source: polymer and route, event and dates, actors, locations, demonstrated stage, TRL, capacity, performance, cost, climate footprint, residual fossil content and constraints, and open questions. Every value needs a verbatim quote that contains it. The validator rejects values without quotes, capacity without a status (nameplate is never shown as output), climate reductions across different system boundaries, sources after the cutoff, and repeats of the same technology + organisation + facility + event across months unless there is a new milestone, correction or reversal. These attributes are separate fields and are never inferred from each other: bio-based origin · measured bio-based carbon content · mass-balance attribution · biodegradability (and in which environment) · industrial or home compostability · mechanical or chemical recyclability · climate footprint.

4 · Technical review

Only a named reviewer's signature (name and date in the issue YAML) moves an issue to signed. Until then items say “Evidence-checked, not yet signed”. The technical reviewer is still OPEN.

Labels

machine-triaged peer-reviewed supplier announcement preprint/patent lead hold. Tracking labels: Watch (Worth following; no action implied.) Data gap (A key attribute is not disclosed or not yet verified.) Milestone to check (A dated claim (plant start, launch, result) to verify when it falls due.) Novelty types: New chemistry · New route · Formulation / application · Scale or supply.

Gaps and limits

  • Absence of news is not absence of activity. Many developments are never published, or only in sources we cannot reach.
  • Language coverage: discovery queries run in English, German, French, Chinese and Japanese news; scholarly indexes are mostly English. Languages seen in the data so far: de, el, en, fr, id, it, ja, tr, zh. Non-English quotes are kept in the original language.
  • Access: paywalled full texts are not read at discovery; triage uses titles and abstracts only. Patents are not yet searched automatically.
  • References: 59/59 DOIs and 61/61 supporting abstract quotes verified on 20 Oct 2026.

Contribute

Submit a lead, a correction or a structure through GitHub issues. Content lives in plain YAML under data/; every new reference needs a resolvable DOI and every new structure a SMILES and a citable source.