Retrieval-augmented generation over 14 CFR Part 25 and EASA CS-25 needs four things: an authority tier on every chunk, point-in-time versioning at section granularity, a paragraph-level citation on every sentence returned, and refusal when no citation exists. Rule text and its means of compliance ship as separate documents on both sides of the Atlantic, so chunking by page severs them.
Ross Jones — Founder, The Hopium Lab. Last modified 21 July 2026.
One retrieval backbone for design-approval and completion engineering work: ingest and snapshotting, authority tiering, point-in-time resolution, a refusal path. Engineering account, not legal advice.
Why does naive chunking destroy regulatory text?
Because the requirement and its means of compliance are published as separate documents, in separate systems, on separate dates. 14 CFR 25.1309 sits in the eCFR. AC 25.1309-1B sits in the FAA Advisory Circular library under its own document ID, issued 30 August 2024. The rule was published on 27 August 2024 and took effect on 26 September — the AC landed three days after publication, in a second, unlinked system, joined by a shared "25.1309" stem no resolver understands. EASA splits rule from means of compliance by design: CS-25 Book 1 carries the specifications, Book 2 the acceptable means.
The fair objection: EASA's Easy Access Rules for CS-25 already interleave specification, AMC and guidance by topic — the co-location I am arguing for, published free by the authority. But the Easy Access Rules are a non-legal consolidation, and they lag the issued instrument: still, at the time of writing, a revision from January 2023 against CS-25 Amendment 28 of 19 December 2023. One code, one revision, no FAA equivalent, no legacy basis.
An authority tier is a label on every chunk recording what the source document is in law: issued rule, issued guidance, policy, programme-specific paper, or accepted draft.
| Source | What it is | What the index must carry |
|---|---|---|
| 14 CFR Part 25, eCFR | Issued rule | Dated snapshot, section-level amendment |
| CS-25 Book 1 | Issued specification | Amendment number and date, never "current" |
| CS-25 Book 2, AMC | One acceptable means | Edge to its CS — beside it, never merged in |
| AC 25.1309-1B | Issued guidance | Its own issue date, not the rule's |
| FAA Order 8110.37F (Revision F) | Direction to the agency and its designees | Flagged as direction, not means of compliance |
| 2002 Arsenal Draft (proposed Rev B, never issued) | Never issued; accepted programme by programme until 2024 | Labelled never issued, or withheld |
AC 25.1309-1A issued on 21 June 1988 and stood unrevised for thirty-six years. From 2002 until Revision B issued on 30 August 2024, serious programmes worked substantially to the Arsenal Draft — a document the FAA never released, accepted programme by programme.
Which version of the regulation applies?
The certification basis applies — the applicable airworthiness and environmental requirements established for a type design or a change to it, plus the special conditions, exemptions and equivalent level of safety findings layered on top. "Whichever amendment was in force on the date of application" is a default in 14 CFR 21.17(a), not the answer.
Read the rest and the default dissolves. A transport category type certificate application is effective for five years; any other type certificate application, three. On an extension under 21.17(d)(2) the applicant selects the date. Elect a later amendment under 21.17(e) and the FAA can pull in any other it finds directly related. For a cabin modification or an STC the governing rule is the changed product rule at 21.101, which works from the certification basis already incorporated in the type certificate, not from a fresh date of application.
A retrieval system over certification regulations that cannot state which amendment level it is quoting is not a research tool. It is an efficient way to cite the wrong rule.
How many versions of a section do you have to serve?
Far fewer versions than the part-level amendment number implies: six for 25.1309, not 152. Part 25 sits at Amendment 25-152 as of July 2026. Section 25.1309 has five distinct texts since 1970 — Amdt 25-23 (1970), 25-38 (1976), 25-41 (1977), 25-123 (2007), and the 2024 replacement under Amendment 25-152, docket FAA-2022-1544, "System Safety Assessments", 89 FR 68706, published 27 August 2024, effective 26 September 2024 — plus the original text adopted with Part 25 at Doc. No. 5066, 29 FR 18291, on 24 December 1964, still live in pre-1970 legacy bases. Six texts, not one hundred and fifty-two. Version at section granularity or the index is fiction.
The 2024 rewrite of 25.1309 also deleted the section's own history: the current citation line carries the 2024 docket alone, while the 2018 snapshot enumerates the lineage back to 1970. Ingest keeps every dated snapshot.
Two limits, disclosed before somebody finds them. The eCFR versioner does point-in-time lookup by date, with a floor: 25.1309 at 2017-01-01 returns text, at 2015-06-01 returns HTTP 404. Bases from the 1990s live in the bound annual CFR editions on govinfo. Second, the 2024 rule begins at 89 FR 68706 while the eCFR citation line points at 68735, the amendatory page — both correct, not interchangeable.
Why should the system refuse rather than guess?
A certification retrieval system should refuse because the reader is an auditor, and an uncited answer costs more to check than it saves. Citation-or-refuse means the system returns text with a paragraph-level citation the reader can open, or nothing.
Grounding alone will not get you there. In the Stanford RegLab study published in 2024, three products from two vendors — all retrieval-backed, all sold on eliminating hallucination — produced hallucinated output between 17% and 33% of the time. Certification work carries the sharper edge, because paragraph identifiers are compositional: a fabricated cite to 25.1309(d) parses, reads correctly, and points at a paragraph marked [Reserved] in the current text.
Refusal does not eliminate fabrication. Refusal converts an invisible failure into a visible one, and a citation the reader can open moves the cost of checking from the author to the reader.
How do you evaluate retrieval when the ground truth is a legal document?
Score retrieval separately from generation, against exact text spans annotated by a practitioner, and report Precision@k and Recall@k per authority tier. LegalBench-RAG is the reference design: 6,858 query–answer pairs over a corpus of more than 79 million characters, traced to exact source locations and annotated by domain experts.
One number from an adjacent domain is worth knowing. A 2025 pipeline over ITU radio regulations reports roughly 97% retriever accuracy under its own domain metric, and nearly 12% relative improvement for GPT-4o against under 1% from stuffing whole documents into context — on that corpus, long-context stuffing underperformed the pipeline by an order of magnitude. Every question in my eval set carries the date its answer is asked "as of", because an undated question has no correct answer: the replay test applied to a corpus rather than a model.
Who decides what the regulation means?
The DER or the ODA unit member holding the delegation, and ultimately the authority. 14 CFR 183.29 conditions a DER's approvals on being made within limits prescribed by, and under the general supervision of, the Administrator; an ODA exercises its functions under a procedures manual the FAA approves. Delegation moves the work, not the authority. The software returns the text and shows where it came from; a person decides what it means, and their name goes on the finding.
EASA's AI Concept Paper Issue 2, published 6 March 2024, supplies the vocabulary: Level 1A is human augmentation — automation support to information acquisition and analysis, full authority retained by the end user. Borrow the words and nothing else; no authority has assessed anything here.
Aviation certification retrieval is not untouched territory. AIAA SciTech 2024-1528 describes a certification digital assistant built over regulatory requirements and means-of-compliance documents. The aerospace literature covered the assistant; it did not cover the corpus engineering underneath — the evidence layer every regulated retrieval system needs.
How do these systems fail in production?
Regulatory RAG systems fail predictably, in ways that survive a demo and surface in review. Name them, because unnamed risks never get budget.
- Amendment blindness. Current text returned against a legacy certification basis. Fluent, sourced, wrong.
- Orphaned guidance. AMC or AC text returned without its rule, or the rule without any acceptable means.
- Tier collapse. A policy statement or accepted draft carrying the on-screen weight of issued rule.
- Synthetic paragraph numbers. A composed identifier that parses and does not exist.
- Silent partial retrieval. Three of five relevant paragraphs, confidently assembled, two missing with no signal.
- Cross-authority conflation. EASA guidance answering an FAA question. Two regimes, one answer, no flag.
- Ingest amnesia. No record of which snapshot produced an answer, so March's query cannot be rebuilt in July.
What does a chunk have to carry?
A chunk has to carry everything a reader needs to check the citation without asking a second question. Copy the contract, fork it, argue with it.
CERTIFICATION CORPUS — CHUNK CONTRACT v1.0 (21 July 2026)
The Hopium Lab. Assistive retrieval. Supports a compliance case,
never confers one. Not legal advice, not a finding of compliance.
REQUIRED ON EVERY CHUNK
[ ] authority — an authority tier is a label on every chunk recording
what the source document is in law: issued rule, issued
guidance, policy, programme-specific paper, or accepted
draft
[ ] instrument_id — 14 CFR 25.1309 | CS 25.1309 | AMC 25.1309 | AC 25.1309-1B
[ ] amendment — SECTION level, not part level (Amdt 25-152, CS-25 Amdt 28)
[ ] in_force_from / in_force_to — nullable end; "current" is not a value
[ ] paragraph_path— 25.1309(b)(1), verbatim from source, addressable
[ ] source_uri — a deep link the reader opens without credentials
[ ] snapshot_hash + retrieved_at — the bytes ingested, committed on ingest
[ ] jurisdiction — FAA | EASA, never blended inside one answer
[ ] companion_ids — the AMC/AC paired with this rule, and the reverse edge
REFUSAL RULES — all must hold or the answer is withheld
[ ] No paragraph_path -> nothing returned
[ ] Query date outside every in-force window -> nothing returned
[ ] Cross-jurisdiction merge attempted -> nothing returned
[ ] Only "accepted draft" or "policy" available -> labelled, absence of
issued text stated first
[ ] Corpus behind the latest published amendment -> banner every result
EVALUATION
[ ] Lookup scored independently of generation, per authority tier
[ ] Ground truth = exact spans annotated by a certification practitioner
[ ] Held-out set includes bases predating the point-in-time API floor
[ ] Every question dated: the "as of" its answer is asked against
None of the chunk contract above is model work. Corpus engineering, versioning discipline, a refusal path. Retrieval is the easy half; knowing which of six texts was in force on a date in 1997, and showing it, is the product.
Engineering account, not legal advice. Assistive retrieval supports a compliance case; it never confers one.