What Goes In, What Never Does

What retrieval looks like, before you govern it

Before governing a corpus, see the artifact you are governing for. Here is a real retrieval against this training’s practice corpus — a question, and what came back:

Q: How often should transformer oil be sampled?

RETRIEVED:
[1] FIST 3-30 "Transformer Maintenance", p.41   (FIST — the Bureau of
    Reclamation's public power-equipment maintenance manuals) - "...oil samples should
    be taken annually as part of routine..."
[2] FIST 4-1B "Maintenance Scheduling", p.12 - "...intervals in this
    volume supersede equipment-specific manuals..."
[3] Trnsfrmr.pdf "Transformers: Basics", p.118 - "...sampling
    frequency depends on transformer criticality..."

Three passages, three documents, and they do not agree — one of them explicitly claims to override another. Everything in this module exists because of what you are looking at: retrieval returns whatever is in the corpus, agreement not included. The system did not create that conflict. Your document collection contains it, and always did.

Select by class, not by availability

The reflex is “index everything we have.” Resist it with the same argument each time: every document in the corpus is a document the system may quote to a consumer. The corpus is not storage — it is the set of things you are willing to have cited as answers.

Selection runs per question class from Lesson 1.3: failure-history needs the field notes and work orders; standard-requirements needs the standards and manuals; contract-obligations needs the filings. A document that serves no class has no reason to be quotable, and “it might be useful” is how a decade-old draft procedure ends up cited as current practice.

Authoritative, superseded, reference — and the rule

Real collections contain the same fact at different ages. The practice corpus has both honest cases:

• Editions over time. The electric construction standards exist in editions spanning decades. The old editions are not garbage — “what did the 2010 standard require?” is a legitimate question — but a current-requirements question answered from 2010 is wrong with a real citation.
• Documents that override each other. The maintenance-scheduling volume states outright that its intervals supersede the equipment manuals’. Both stay in the corpus; one wins on schedule questions.

So every document gets a status — authoritative (current, wins), superseded (kept, answers historical questions only), reference (context, never the sole support for a requirement claim) — and the corpus gets a written resolution rule:

RESOLUTION RULE - Meridian document corpus
  1. On any conflict, the document marked authoritative for that
     class wins.
  2. Scheduling intervals: FIST 4-1B is authoritative over
     equipment-specific manuals (per 4-1B's own supersession note).
  3. Standards questions answer from the newest edition unless the
     question names a year.
  4. A conflict not covered above is SURFACED, both sources cited,
     and routed to the corpus owner - the system does not pick.
  OWNER: field engineering manager (documents), reviewed annually

Rule 4 is the one that keeps you honest. The resolution rule cannot anticipate every conflict, and the system’s behavior on an unanticipated one must be surface, don’t choose — the eval set’s conflict questions from Lesson 1.4 test exactly this.

What never goes in

Exclusion is as much a governance act as inclusion. Standing exclusions, each for a stated reason:

• Drafts and working copies — quotable-as-current is the failure; nothing unapproved enters.
• Documents whose classification exceeds the corpus’s clearance — Lesson 2.3 makes this precise; the short version is the corpus inherits the sensitivity of its most sensitive document.
• Anything you would not want read back verbatim — the system quotes; a candid internal memo becomes a candid citation.
• Personal notes that were never records — a text file on someone’s desktop is not a field note until the records owner says it is.

Prompt first came later, deliberately

This lesson opened with a retrieval, not a prompt, because the governing artifact had to exist first. Here is the working prompt:

Here is my question-class list and a file inventory of the candidate
document collection [paste both].

Propose a corpus manifest: for each document - which class(es) it
serves, a proposed status (authoritative / superseded / reference),
and any supersession relationships you can see from titles, edition
dates, or explicit supersession notes in the text.

Mark every status you inferred rather than found stated - those are
DECISIONS for the corpus owner, not facts. Do not resolve any
conflict; list conflicts for the resolution rule.

The agent reads faster than you and finds the explicit supersession notes. The statuses it infers are exactly the ones a human must confirm — which the prompt makes it label rather than bury.

Stop and escalate when two candidate documents conflict and no one holds authority over both — the interconnect filings and the field procedures may have different owners, and a resolution rule spanning them needs both owners’ signatures, not the builder’s best guess.

Key takeaway

The corpus is the set of things you are willing to have quoted to a consumer, selected per question class, never by availability. Every document carries a status — authoritative, superseded, reference — assigned by someone with domain authority, and a written resolution rule decides conflicts, with “surface both, don’t choose” as the floor for conflicts the rule never anticipated. Exclusion is governance too: drafts, over-classified documents, and never-records stay out. Lesson 2.2 meets the step that decides whether any of this is retrievable at all: extraction.