PSRF v1.5: automatic, scalable validation and sharper accuracy across the board

Pawssier Structured Record Format (PSRF) now validates itself across dates, vaccine identity, and provenance, with clear errors corrected automatically and ambiguous cases routed to a human reviewer.

What's new

PSRF v1.5 introduces an automatic, scalable validation and correction system. Rather than hand-coding a fix for each new type of error as it's discovered, the system is built so that new checks can be added and proven safe before they ever reach production, and so that fields added to the record in the future automatically inherit the same protection.

Consistent dates within every record

A single document, like a multi-page rabies certificate or a visit summary listing several vaccines, often contains multiple dates that need to agree with each other. PSRF v1.5 introduces a validation system that checks every date in a parsed record against every other date in that same record, catching genuinely impossible relationships automatically and routing anything ambiguous to a human reviewer instead of a guess. Before turning this on, we replayed every rule against nearly 3,000 historical jobs and confirmed it would never null a legitimate date. New fields added to the record going forward inherit this same protection automatically.

Accurate vaccine manufacturer attribution

Different clinics and source documents often record a vaccine's manufacturer inconsistently, or not at all. PSRF v1.5 cross-references the brand name against a verified manufacturer source and fills in or corrects the field automatically when there's a clear, known answer, so the record stays accurate even when the underlying documents aren't consistent with each other.

A second-pass review system, building toward automatic correction

Every parse now runs through a second AI pass that checks the first extraction for subtler issues, like a low-confidence field or a value that doesn't quite line up with the document. In early testing, this pass surfaced 214 reviewer-confirmed defects across 366 flagged findings, a strong enough signal that we're now building toward turning specific finding types into automatic corrections, the same way our date and manufacturer checks did before they went live.

Persona-based extraction for the highest-volume formats

We rolled out a new extraction layer built around dedicated personas for the document layouts that move the most volume through our network. Instead of one general-purpose extraction path, each persona is tuned to a specific layout's structure, meaningfully increasing precision on the formats our partners see most often.

More reliable vaccine and treatment history

• Recognizes full booster series accurately, even across multiple visits and providers
• Reads combination vaccines, like bivalent influenza, as a single canonical product even when the label lists each component separately
• Builds a clean treatment history, free of stray or duplicate entries

Cleaner identifiers

• Extracts lot and batch numbers embedded inline in vaccine records, across multiple languages

Better clinic and veterinarian recovery

• Identifies the clinic on a document from contextual details, even without a letterhead or logo
• Recognizes veterinarian names and signatures across more languages and regions
• Enhanced provenance grounding for clinic and provider fields, drawing on more specific evidence from the document, like a header, signature line, or letterhead

Rabies certificates

• When a NASPHV rabies certificate uses duration checkboxes, derives the expiration date directly from the selected box

Behind the scenes

We also shipped a set of reliability improvements that make the overall system more resilient and consistent under the hood. The short version: fewer edge cases, fewer surprises.

By the numbers

• Around 80 targeted accuracy improvements since v1.4
• 27 languages and regional formats supported
• 214 reviewer-confirmed extraction defects surfaced in early grader testing
• 0% blast radius confirmed across nearly 3,000 historical jobs before the new date validation went live

The short version

PSRF v1.5 checks every date within a record for consistency and cross-references vaccine manufacturer against a verified source, correcting clear errors automatically and routing ambiguous ones to a human reviewer. A second-pass review system, already validated against hundreds of real findings, is laying the groundwork for further automatic corrections. The release also introduces persona-based extraction for the highest-volume document formats, and gets meaningfully better at the fields customers care about most: vaccines, identifiers, clinics, and rabies certificates.

A note on backward compatibility

v1.5 is non-breaking. No fields, parameters, or schemas changed from v1.4. What can change is the value returned for a given field: dates that fail a sanity check may now come back null or have their row dropped rather than emitting an implausible value, and clinic or provider fields without traceable evidence are withheld rather than inferred. If you need the previous behavior, you can pin to psrfVersion: "1.4.0".