How an agent uses this · Qualus proposal wiki

Start / How an agent uses this

worked exampleretrieval, not prompt-stuffingreal output from the index

The wiki is not a prompt. Nobody pastes 2,500 pages into a model and hopes. An agent asks a narrow question, gets a handful of sourced passages back, and drafts from those — or is told the corpus has nothing.

1. The agent asks a question

$ python3 scripts/search-sections.py \
    -q "deemed savings verification sampling" --limit 3

That reads a local index of 17,153 extracted passages. It is the retrieval entry point a drafting agent calls, and equally the thing a human runs before claiming the corpus contains something.

2. It gets passages back, each with a receipt

hits      3

[1] 12.983888  sampling_analysis  (prose)
    heading   3.5.1 Prepare Sampling Plans
    cite      F021015@page:17:line:4 .. page:18:line:42
    source    authoritative_submitted / year 2012
    sha256    349a8cec03f74f79
    | As discussed above, we prepare a sampling plan for each
    | program for selecting samples of customers for measuring
    | and analyzing savings. We take into consideration that
    | participants for a program will be accumulating over time…

[2] 12.884958  deliverables_schedule  (prose)
    heading   Statement of Work and Schedule 18
    cite      F022669@page:22:line:33 .. page:24:line:18
    source    authoritative_submitted / year 2012
    sha256    7751fe45131c1924
    | …For the analysis of gross savings for some non-residential
    | customers, on-site monitoring may be performed at some
    | sites. We choose sites for monitoring where there is some
    | uncertainty about the values for important factors…

Three things in that output are the whole point:

A byte span, not a summary. Page and line, both ends. The agent can be handed the exact text and cannot quietly retell it.
A digest of the source file. If the document changes, the receipt stops matching.
An authority tier. authoritative_submitted means we wrote and submitted this, so it is quotable or paraphrasable as our own language. The same field can say third_party, which means read-only.

3. It drafts, and it stays inside the receipts

What the agent does with a hit is deliberately boring: quote the passage, carry the citation into the draft, and fill the gaps with a question rather than a guess. Where the corpus has nothing, the section page says MISSING and the draft says MISSING — it does not become plausible prose.

4. The three ways this goes wrong, and where each is blocked

Failure	Where it is stopped
Invents a statistic to fill a thin section	Every page is generated only from facts that carry a locator; a fact without a receipt never enters
Reuses a competitor's or issuer's words as ours	Authority tier is checked at the point of use; `do not quote` pages are marked on the page and filtered out of the agent's input
Cites a document that has since changed	Source digests are compared on every build, and drifted pages are flagged rather than hidden

5. Why the metadata is the product

A page without status, use_when and a source list is a paragraph. With them it is a decision the agent can make: is this the right page, may I quote it, and where exactly did it come from. That is the difference between a document store with a search box and a reference an agent can be trusted with.