---
name: examiner-retrieval-test
description: Use this skill when the user wants to know whether a claim-shaped search would find one of their public disclosures, the way a patent examiner or a challenger would search. Inputs include one or more Prior Art Records, the list of approved search surfaces, and claims from any patent the team watches. The skill writes claim-worded queries plus a control in the subject's own words, runs each three times on each approved surface, records the rank of the disclosure or that it was not found in the top 20, scores the worst run, and produces the prior-art retrievability row for the Actionboard. Triggers on requests like "run the examiner retrieval test," "would an examiner find our note," "test our disclosure with claim-worded queries," "how retrievable is our prior art," or "update the prior-art row on the Actionboard."
---

# Examiner Retrieval Test
This skill is Labor as Code™ for the Examiner Retrieval Test in
*Discoverability Engineering*. The know-how is an operator's habit of
searching for a disclosure the way a stranger under time pressure would,
in the words a claim uses, and never in the words the company uses about
itself.

The method comes from "Patents and Prior Art Are Public Evidence That
Needs a Date Anyone Can Prove" and follows the controlled-test rules in
"Controlled Tests and the Actionboard Show Whether Discovery Is Getting
More Correct": the same queries, the same surfaces, three readings, the
worst one scored. Read both. This skill repeats the test on the user's
disclosures. It does not retell Merriden's first run.

The skill measures retrievability. It does not say that a disclosure is
or is not prior art, and it does not change any page, record, or search
engine setting. It never tries to get a disclosure to rank by any means
other than the disclosure's own text and dated record.

## Before running

1. **Which records.** Get the Prior Art Records to test. Each needs a
   `teaches` sentence. Stop and ask if one has none.
2. **Which surfaces.** Read `specs/surfaces/patent-and-literature.yaml`
   for the approved surfaces (for example a general web search engine
   and a scholarly and technical-literature index). The engineer
   approves surfaces. Do not add one.
3. **What was run before.** Read `state/retrieval-tests/` so this run
   keeps the same queries. If the wording has to change, the old query
   is retired and the new one gets its own ID.

## Step 1: Write the queries

For each record, write four to six queries:

1. Take the wording from the record's `teaches` sentence and, where
   they overlap, from the claims of patents in `state/patent-watch/`.
2. Write at least one query as long as a claim's first line, one as a
   shorter technical phrase, and one that names the standard or
   product.
3. Add one control query in the subject's own words. The control shows
   whether the disclosure is findable by people who already know it.
4. Give each query an ID (`PQ-1`, `PQ-2`, and so on) and store its text
   word for word. Do not retype it on a later run.

## Step 2: Run each query three times on each surface

1. Run each query three times per approved surface, on separate
   occasions where the surface allows it. Record the date, the surface,
   and any setting that changes results (signed in, location).
2. Record the rank at which the disclosure's address appears, or NOT
   FOUND in the top 20. Save the result list it came from.
3. Score the worst of the three readings for each query and surface.
4. Do not open a result page on the user's behalf and click through to
   a login. Record what a search returns.

## Step 3: Compute and record

Write one record per record ID and date in
`state/retrieval-tests/<date>-<record-id>.yaml` with the queries, the
surfaces, the ranks, and the count of readings in the top ten out of
the total. The count is the prior-art retrievability figure for the
Actionboard's fourteenth row ("3 of 8" means three readings out of eight
ranked in the top ten).

## Step 4: Conclude

State in plain words:

1. **The result.** Which queries found the disclosure and which did
   not, and whether the control found it when the claim-shaped queries
   did not.
2. **The likely reason,** from the record: a vague title, no class,
   words that differ from a claim's, or a surface that does not index
   that kind of page.
3. **The smallest fix,** with the reason it should help: a dated new
   version worded as a claim, a stated class, a copy where technical
   literature is indexed. The person who owns the record decides
   whether to publish it.
4. **When to run it again,** on the test set's calendar and after any
   new version goes up and has had time to be indexed.

Where the surfaces changed between runs, say so and do not compare the
counts. Where the evidence is too thin to explain a miss, say so and
name what would settle it.

## What this skill does not do

- It does not change a page, a record, or a search setting.
- It does not treat a rank on the company's own site as a result.
- It does not add a surface the engineer has not approved.