Verification

Nassila verification is layered: L1 anchors each reference to a registry; L2 compares your fields to the canonical record; L3 (Manuscript only) grounds passages against source excerpts via Sanad.

Tier summary

Tier	Name	Where it runs	What it checks
L1	Registry resolution	Bibliography + Manuscript	DOI/PMID/URL anchor to Crossref, PubMed, OpenAlex, or explicit grey/insufficient outcome
L2	Metadata alignment	Bibliography + Manuscript	Title, year, journal, volume, pages, authors vs canonical record when L1 succeeded
L3	Passage grounding	Manuscript loop only	Claims in your draft vs source excerpt (abstract or OA text) via Sanad

L1 — Registry resolution

L1 answers: Can this row be tied to a trusted catalog entry?

Resolution order (simplified):

DOI — Crossref/DataCite → OpenAlex → PubMed where applicable
PMID — PubMed → OpenAlex
No identifier — OpenAlex title/author search or grey-web heuristics

Outcomes include resolved DOI/PMID, OpenAlex work match, or explicit insufficient-data flags.

L2 — Metadata alignment

When L1 finds a canonical record, L2 compares your CSL fields. Crossref/PubMed rows may receive safe auto-patches when the registry clearly disagrees. Remaining differences appear as yellow mismatch cards in Issues — click to jump to the row.

L3 — Passage grounding (Sanad)

L3 runs only when:

You are in Manuscript mode
Passage grounding is enabled (Sanad bar toggle)
A source excerpt exists (abstract or OA text)

Sanad returns structured verdicts: supported, weak, contradicted, not_in_source, insufficient_evidence, with verbatim sourceQuotes when supported. Guardrails block quotes not found in the excerpt.

Batch limits

Limit	Value
Rows per Verify references run	200 (prioritized: DOI/PMID rows first)
Network	Required for L1/L2 verify and Resolve

Mismatch cards

After verify, mismatch cards show:

Reference title / authors (for context)
Field name (e.g. volume, year)
Your value vs registry value

Fix manually in Output or accept patches already applied.

Offline behavior

Action	Offline
Parse, edit, validate, export loaded data	Yes
Verify references (L1/L2)	No
Resolve DOI/PMID/URL	No
Manuscript OA fetch	No
Sanad (local LLM)	Yes, if runner is local