How Borgh reads a dossier.

What Borgh has read

Borgh’s findings are grounded in a fixed body of EFSA material, not in a model’s general training knowledge. The knowledge base holds 11,045 passages (10,975 unique) drawn from four sources:

• 240 of the 241 published EFSA opinions on novel food (2018–2026) — every opinion the NDA Panel has issued under Regulation (EU) 2015/2283, bar one not yet released by the publisher.
• The 2024 NDA Panel scientific guidance, the administrative guidance, and sector-specific guidance for insects, cell culture and microorganisms.
• 16 EFSA working-group minutes across nine panels — the record of what the panel discussed before it adopted an opinion.
• EFSA and OECD methodology documents — genotoxicity, 90-day toxicity, benchmark dose, the OECD test guidelines — and the academic literature those opinions cite.

Borgh does not read the open web during a review. It reads this corpus, and it tells you which passage of it each finding came from.

How Borgh finds the passage that matters

For each section under review, Borgh searches the knowledge base for the passages that bear on it. Search is by meaning, not keyword: each passage and each query is converted to a 1,024-dimension vector (Voyage AI’s voyage-3-large), and Borgh retrieves the closest matches.

A single similarity search tends to return ten variations of the same guidance paragraph. Borgh runs three searches in parallel instead — one per pool — and takes a quota from each:

The quota guarantees a review sees both the rule and the precedent for it, not one at the expense of the other.

How Borgh checks precedent

Beyond retrieval, Borgh queries a structured database of 239 past opinions — searchable by novel food type, by what triggered an additional data request, and by outcome. Before it writes a finding, Borgh looks up comparable cases: what did the panel ask of the last precision-fermentation applicants? Which compositional gaps stopped the clock?

This is an agentic step, not a fixed lookup. Borgh decides which cases to pull and when, runs the queries itself, and reads the results before drafting. A concern is checked against what EFSA has done before, not against a guess about what it might do.

Cite or stay silent

Borgh makes only claims it can source. Every sentence in a finding either cites a specific passage from the knowledge base or is marked as Borgh’s own reasoning from cited material. There is no third category. If Borgh cannot ground a concern in EFSA’s published material, it does not raise it.

This is a deliberate constraint. An earlier version let Borgh label a concern “unsourced” — and it reached for that label constantly, to avoid the work of finding the exact citation. The label was removed. The rule is now plain: no source, no finding. The cost is fewer findings; the gain is that every finding is anchored to a passage you can read.

How Borgh grades a finding

Borgh sorts findings into four tiers, ordered by their consequence inside the EFSA process — not by a generic severity score:

Borgh does not collapse these into a single readiness score. One admissibility blocker makes a dossier inadmissible regardless of how clean the rest is — a weighted percentage would hide that.

How Borgh is tested

Borgh’s retrieval and citation behaviour are measured against fixed test sets, re-run whenever the pipeline changes:

• A retrieval eval — 13 representative queries, three runs each — checks that the right guidance passages surface in the top results.
• A cite-or-flag eval — seven dossier sections across five EFSA sections and four novel food types — checks that no finding contains an unsourced claim. The current result is 30 of 30 findings fully cited.

These are small sets. They catch regressions when the pipeline changes; they do not certify that every finding is correct.

What Borgh cannot do yet

Borgh is a first-pass reader, not a panel. Its current limits, stated plainly:

• It reads tables as text. Borgh does not yet re-compute a reported mean, flag an outlier batch, or check that a value sits below a stated regulatory limit. A numerical inconsistency inside a data table can pass unmarked.
• Findings vary run to run. The model is not deterministic, so a second review of the same section may surface a slightly different set of findings. For a decision that matters, read more than once.
• A citation can land on the wrong passage. Borgh matches each claim back to a source passage, exactly or approximately; the approximate match occasionally points to a neighbouring fragment. The cited text is always shown so you can check it against the claim.
• The corpus has an edge. Borgh has read novel food opinions, not the whole of EFSA. For an organism category with little published precedent, treat the findings as procedural and bring your own scientific judgement.