KERYNEXARTIFACT RELAY

Your AI cites its sources.
What entered the retrieval pipeline?

Citation and claim checks answer questions about a source or a generated answer. Kerynex addresses a narrower question: what did a bounded retrieval worker fetch, when did it arrive, and over what route? The architecture records the retrieval event; it does not determine whether a source is true, supports a claim, or was relied on by a model.

Kerynex documents a designed pre-sale guide and sanitized reference implementation for engineers who want inspectable retrieval-event records: bounded, SSRF-safe fetching, fetch-time hashes, transport metadata, and explicit failure/escalation handling. Captured, not certified: a fetch-time record, not a verdict on the source.

The gap between a generated URL and an inspectable retrieval record

When downstream systems rely on retrieved context, a broken citation is not an edge case. It surfaces in court sanctions and in statutory enforcement.

REF
FINDING
REF-01
Citation fabrication rates

In published tests, 11–57% of AI-generated citations are fabricated across tested models: non-existent URLs, hallucinated DOIs, or real URLs pointing at unrelated content.

Cross-model citation-fabrication studies, 2024–2026 aggregates [B]
REF-02
Statutory enforcement tier

Transparency duties under the EU AI Act fall in the tier carrying fines up to €15M / 3% of total worldwide annual turnover, in force since August 2, 2026.

EU AI Act Art. 99(4)(g), via EC AI Act Service Desk [A]
REF-03
Judicial sanctions

Attorneys sanctioned by a federal court for filing ChatGPT-fabricated case law, submitted without verification of what the retrieval actually returned.

Mata v. Avianca, 678 F. Supp. 3d 443 (S.D.N.Y. 2023) [A]
provenance_receipt.jsonillustrative example
{
  "raw_sha256":        "sha256:d69204b4fa3f…",
  "retrieved_at":      "2026-08-24T14:02:11Z",
  "http_status":       200,
  "transport":         "httpx",
  "canonical_url":     "https://example.com/doc.html",
  "redirect_chain":    ["https://example.com/doc"],
  "elapsed_ms":        412,
  "bytes_received":    18243
}
Illustrative retrieval record. Full schema ships in the reference implementation.

What the pre-sale package was designed to include

A pre-sale architectural guide plus a sanitized reference implementation. The validated internal pipeline exists; the buyer-facing package ships only after it is built and passes fresh-clone validation in its supported environment.

01

The Architecture Guide

How to build a bounded retrieval layer that records what it fetched, with explicit limits and failure handling. It is not a compliance or legal guide.

02

Bounded, SSRF-Safe Fetcher

A reference fetch worker with explicit IP egress guards, DNS rebinding mitigations, response-size ceilings, and strict domain policy enforced before any payload reaches memory.

03

Transport Provenance Schema

A structured record for raw and normalized hashes, timestamps, routing identity, HTTP/content metadata, redirect history, and fetch outcome details.

04

Status Classification & Local Records

Success, failure, escalation, and health classifications with local JSON records and aggregated escalation output. The pipeline records escalation events; it does not execute browser, crawler, Firecrawl, or OCR tiers.

Honest related work

Reliable retrieval is an active engineering space, and other projects attack adjacent parts of the problem:

  • PCRAG works at the answer-certificate layer: signed certificates and post-hoc verification of generated answers.
  • verbatim-rag works at passage selection: returning source excerpts with citations instead of free generation.

Kerynex Artifact Relay teaches the retrieval-event boundary: SSRF-safe bounded fetch, SHA-256 artifact hashing, transport metadata, and explicit failure/escalation records. It does not certify a source, support a claim, or establish later record integrity.

Pre-sale closed; research retained

The 30-day public pre-sale validation gate closed 2026-09-23 without reaching its threshold (0 sales) and the pre-sale was withdrawn on 2026-09-27. The specification below is retained as a read-only archive of what the package was designed to be. Nothing on this page is for sale. If the gate had been reached and later folded, purchases would have been refunded in full — none were made.

INCLUDED

  • The Architecture Guide (PDF / HTML), redacted from internal project context before shipping.
  • A sanitized, standalone reference implementation: bounded fetch worker, fetch-time hashing, transport metadata, status classification, and local JSON records.
  • LICENSE, dependency notices, pinned extras, buyer README/quickstart, and fresh-clone validation of the documented flagship workflow.

EXCLUDED

  • No hosted service, SaaS endpoint, vector storage, embedding generation, or reranking plumbing.
  • Not the internal production repository; what ships is the sanitized reference implementation described above.
  • No VERIFIED/DEGRADED/FAILED field set, TLS-cipher, egress-policy, CIDR-validation, signed/anchored, or tamper-evident record promise.
  • No automated citation verification, source-truth or claim-support determination, observability-stack export, audit certification, legal sufficiency, consulting, or custom SLA.
release_formatguide (PDF/HTML) + sanitized reference implementation archive
supported_environmentLinux with Python 3.13 only
support_modelbest-effort asynchronous; no response-time commitment
refund_conditionfull refund if the validation gate folds or the supported flagship workflow cannot be completed
PRE-SALE WITHDRAWN · ARCHIVE ONLY · NOT FOR SALE

Educational product: not legal advice and not a compliance certification.

Pre-sale closed · No purchase path · Educational product · Not legal advice · Not a compliance certification