The Reading Room / Compare searches
What did both
searches miss?
Say you need measured melting temperatures to choose a material for a hot part. See what each search finds in four reads and which evidence both leave out.
Loading the comparison. Controls and links unlock when ready. Your linked example stays in the address if loading fails.
Walk through the example
More reading, or a better search?
Start by comparing a reported melting temperature with a paragraph that only gives an oven setting.
The labels cover the whole collection, including paragraphs missing from both lists.
Loading the labelled collection.
The same deadline for all three options
How many paragraphs can you read?
Each paragraph uses one read. Every option gets the same maximum budget. A shorter list can leave reads unused.
Useful paragraphs checked / all useful labels
Select a paragraph to open its full text and positions. On narrow screens, scroll the three reading piles sideways.
Alternate rank positions from A and B. Skip a repeated paragraph ID without charging another read. This is one explicit merge policy, not an optimal ranking.
Before reading: what reached either list?
Different finds.
A shared blind spot?
Only paragraphs labelled useful appear in this strip. Their positions show whether A, B, both or neither returned them. This does not change with the reading budget.
Open a paragraph to inspect the evidence
The supplied useful label is the reference for this comparison. Reading correctly is assumed. A returned paragraph is not necessarily a reliable measurement or a usable scientific record.
Inspect the whole reference collection
Missing from a list
is still in the collection.
To count what a search missed, you need labels outside its returned list. Lists alone cannot establish recall. A paragraph can be useful for this question and irrelevant to another.
Open the paragraph table
| Paragraph | Supplied label | A rank | B rank | Text |
|---|
Use your own shared collection
Compare your lists.
Keep the labels explicit.
Supply one question and every paragraph in the reference collection, including useful paragraphs returned by neither list. Include each list’s rank or leave it blank when absent. This page cannot verify that your collection is complete or its labels are right.
Files are read in this browser tab. Nothing is uploaded to a search service, and paragraph text is never placed in a setup link.
1 to 200 unique paragraph IDs; explicit “useful” or “other” labels; ranks 1, 2, 3 without gaps or ties. Exact CSV columns: question, id, text, useful, rank_a, rank_b. Repeat the same question on every row. UTF-8, up to 1,000,000 bytes.
Inspect exact counts, denominators and model limits
One collection, three reading queues.
A and B are ranked IDs over the same fully labelled paragraphs. Combined alternates their rank slots in the chosen order, removes duplicate IDs and drains the longer list when needed. This is not reciprocal-rank fusion, a relevance score, an automatic search or an AI run.
| Option | Returned | Useful returned | Precision | Recall | Checked | Useful checked | Useful unread | Useful absent | Unused reads |
|---|
Precision = useful returned / all returned. Recall = useful returned / all useful labels in the reference collection. The known prefix gives an exact checked count under the supplied labels and perfect-recognition assumption. It is not an estimate from precision × budget.
The original Reading Room’s unchanged unknown-order bounds bracket these exact prefix counts. Bounds are possibilities, not confidence intervals. Paragraphs take equal budget places here; real reading times and recognition errors vary. There is no statistical significance test, calibrated label quality, general performance claim or proof that these fictional results represent any real tool.