Citation checks and retrieval records answer different questions.

A bounded retrieval worker can record what it fetched, when, and over what route. That record is useful technical context; it is not a verdict on a source or a substitute for human review.

Every AI citation checker on the market answers one question: does this case, article, or URL exist? That's the easy half of the problem — and the past eighteen months have shown it's not sufficient.

A Connecticut court sanctioned a lawyer this year not because a citation was fake. In Barteca Holdings (D. Conn., Aug 2026), the cases were real, and he had verified they existed. The sanction came because nobody had compared the AI-drafted quotations, pinpoint cites, and legal propositions against the real case text. Courts have repeatedly held that the duty to verify can't be delegated to a tool — AI or otherwise (Barteca; Fletcher v. Experian). As the Barteca court put it, quoting Benjamin v. Costco Wholesale Corp., Rule 11 requires attorneys, at a minimum, to "read the cases they cite."

The fabricated-citation file keeps growing. Damien Charlotin's independently maintained AI Hallucination Cases Database — cited in Norton Rose Fulbright's 2026 update at 1,148+ cases — has since grown past 1,700 rulings worldwide. Princeton's Polaris Lab counts 900+ filings containing fabricated or misrepresented citations, using a separate methodology. The sanctions are real and escalating: fines from four figures into the tens of thousands, two-year pro hac vice bars, grievance and bar referrals.

One technical question remains separate from citation and claim review: can a system show what its retrieval worker fetched, when it arrived, and over what route? A fetch record can capture that event. It does not establish whether a model used the material or whether the material supports a later claim.

That's not a citation problem. It's a provenance problem — and the checkers everyone is buying don't solve it.

What the citation checkers actually do

They take a citation string and compare it against a corpus. Two gaps:

1. They check the string, not the fetch. A cite-checker validates that "Smith v. Jones, 123 F.3d 456" resolves to a real case. It cannot tell you what your retrieval pipeline actually pulled from the network — whether it got the real case, a cached stale version, an SSRF'd internal endpoint, or a 404 it papered over.
2. Their output isn't the verification courts want. Courts have repeatedly held that the duty to verify can't be delegated to a tool. A checker can tell you the citation is real; it cannot prove your system fetched the real thing.

A bounded retrieval-event record

A retrieval worker can retain an inspectable record of its own fetch event. The validated Tiered Retrieval Pipeline records bounded fetch behavior, fetch-time hashes, route and transport metadata, and explicit failure or escalation details.

Kerynex is a pre-sale package derived from that internal pipeline. Its future sanitized reference implementation is scoped to:

1. Bounded fetch behavior: destination policy, redirect re-validation, response-size limits, and bounded timeouts.
2. Fetch-time hashing: raw SHA-256 before parsing and normalized SHA-256 on successful normalization.
3. Transport metadata: UTC retrieval time, canonical URL, HTTP/content metadata, redirect chain, byte count, elapsed time, and transport/connector identity.
4. Status classification: success, failure, escalation, and health information recorded locally. Escalation events are recorded; browser, crawler, Firecrawl, and OCR tiers are not executed by this package.

What Kerynex is—and is not

Source-review systems, legal research tools, and AI governance tools address different jobs. Kerynex does not claim to replace them. It is an educational package about bounded retrieval-event records for engineers who need to understand the technical boundary before downstream review happens.

What early access will ship

Kerynex is a pre-sale guide plus a sanitized reference implementation derived from the validated internal retrieval architecture: bounded fetching, fetch-time hashing, transport metadata, status classification, and local JSON record output. The buyer-facing bundle is not yet built. It will ship only after fresh-clone validation in its supported Linux / Python 3.13 environment. The offer excludes hosted service, automated citation verification, source-truth or claim-support determination, observability-stack export, signed/anchored records, audit certification, and legal advice.