Bioelectrochemical-systems papers report values in inconsistent ways: power density normalised to anode area in one paper and to total volume in another, peak vs steady- state values reported without distinction, voltages quoted without the reference electrode, and microbial communities described as "mixed culture" with no abundance data. A free-text scrape produces noise. The MESSAI extractor reads each paper's methods/results/tables and emits a structured JSON document with a fixed schema, so every value carries the context needed to compare it against another paper.

Below is every field the extractor produces, organised the way the JSON is shaped. Each field has a What line (the value), a Why line (what it lets the platform do downstream), and where applicable, the rule used to fill it.

Before extracting any measurements, the LLM decides whether the paper actually describes a bioelectrochemical system at all. Many corpus PDFs turn out to be off- topic: zinc-air fuel cells, COMSOL simulations of PEM hydrogen cells, VLSI circuits, clinical guidelines, mobile educational apps, etc. The classification gate rejects these cheaply (~$0.001 per call) and quarantines the file from downstream processing.

The LLM only emits the rich extraction schema when is_bes_paper: true. When false, it returns rejection_reason + concept_tags describing what the paper IS about, and the platform soft-hides the paper.

Twelve canonical types covering the full BES landscape. MES specifically means Microbial Electrosynthesis (cathodic CO₂ reduction), not the umbrella term — the umbrella is BES. See the Glossary tab for the full list with definitions.

Multi-tag list — a paper can be many things at once. A "constructed wetland MFC for nitrogen removal" gets variants=[constructed_wetland, single_chamber] + applications=[wastewater_treatment, power_generation]. Variants describe the chamber/biology architecture orthogonal to the primary type.

What the paper is for. Most papers carry 1–3 of these. Lets users discover "what BES research has been done on metal recovery" or "all CO₂-reduction papers" without needing keyword guesses.

Dimensional and topological information needed to render a 3D model and (eventually) run a multiphysics simulation that can be cross-checked against the paper's own measurements. When a paper omits these the field is null — never invented.

Each electrode role gets a structured object with two parallel fields: a canonical id (slug from mess-materials ormess-membranes) and an as_reported field carrying the exact text from the paper for audit. This means even unmatched materials retain their original wording — nothing is silently dropped.

Biological electrodes (biocathode, bioanode, algal_biofilm, photosynthetic_biofilm, enzymatic_electrode) are first-class material slugs starting in v1.2 — when a paper describes a photosynthetic biocathode, that information lands in materials.cathode.id, not just in system_variants.

What organism(s) the paper studied, with provenance about how the identification was made (16S? metagenomics? culture-dependent?). Distinguishes a paper claiming "dominant Geobacter from V1-V3 16S" from one claiming the same identity from full-length PacBio sequencing — the methodological tier matters when training the predictor.

Per-electrode kinetic parameters. When a paper reports them, we capture them; when it doesn't, we leave the slot null and the predictor uses literature defaults. This schema mirrors the parameters needed by the platform's COMSOL-style validation harness.

When a paper varies a condition (temperature, pH, substrate, R_ext, applied voltage, ...) and reports outcomes for each, the LLM emits a separate condition_set per tuple plus an observations entry tied to that set. This preserves within-paper variation, which is the only kind of variation that supports causal-grade inference.

Every observation carries provenance: which section it came from, what value type (peak vs steady-state vs endpoint), what it's normalized to (anode area? cathode area? volume?), the page number, and a 200-character snippet for audit.

When a paper reports a polarization curve — even partially — the LLM extracts every (current density, voltage, power density) point as a row. These power Move 1's within- paper effect modeling and let the multiphysics validation harness fit Tafel slopes per condition set.

Same shape as polarization curves but for electrochemical impedance spectroscopy. Each Nyquist point (frequency, Z_real, Z_imag) is captured per electrode (anode, cathode, or whole cell). Equivalent-circuit fits live separately in ElectrochemicalKinetic.

One row per (organism × substrate × parameter). Captures Monod μ_max, K_s, Y_X/S, decay rate, maintenance coefficient, current-Marcus K_M, and lag phase when reported. Used by the Move 2 substrate-aware Monod prior.

Per-electrode kinetics: exchange current density (i₀), transfer coefficient (α), number of electrons (n), Tafel slope (b), equilibrium potential (E_eq), and overpotential partitions (η_act, η_ohmic, η_conc) at named operating points. Required by the multiphysics validation harness.

What the paper itself acknowledges as gaps and what it suggests as next steps. Surfaced on the AI Insights tab as "opportunities" — they're a more honest signal than the platform inferring research gaps from the corpus alone.

concept_tags are short keywords describing the paper's contribution beyond the system type (e.g. "biofilm_engineered", "novel_architecture", "scalable"). cited_models are theoretical models the paper uses (Butler-Volmer, Monod, Nernst, Tafel, Bruggeman). analysis_software records the tools (Origin, MATLAB, EC-Lab, NOVA, COMSOL).

The full extraction stack runs each paper through six layers:

1. Layer 1 — self-consistency: 3× sampling at T=0.3, fields with 3/3 agreement go forward at confidence=1.0; 0/3 flagged.
2. Layer 2 — cross-model consensus on disputed fields (Gemini 2.5 Flash + Cerebras llama-3.1-70b for triangulation).
3. Layer 3 — adversarial critic prompt: "find errors in this extraction" looks for unit confusion, value-type misattribution, arithmetic inconsistencies, vs-SHE sanity, stoichiometric impossibilities.
4. Layer 4 — rule-based: range bounds (CE>100% impossible), internal arithmetic (P ≈ V·I within 15%), substrate stoichiometric closure.
5. Layer 5 — cross-paper consistency: same-author baseline checks, methodological-clone outlier detection, citation cross-checks, oracle outliers vs Logan/Patil/Rodríguez literature meta-analyses.
6. Layer 6 — Sonnet 4.6 escalation only when Layers 1+2 disagree AND critic flags errors AND rules trigger. Expected to fire on ≤5% of papers.

The current corpus state runs primarily Layer 1 + Layer 4 (free-tier round-robin, automatic critic on first 5). Higher layers activate when budget allows.