# Scoring Claude Opus 5.5 with retrieval on the twenty entries

Jim Clifford gave the same twenty entries to Claude Opus 5.5 in a chat with web search, 7 October 2026, with the same prompt as the Gemini run: "Parse these bios and ground the places, schools and people to Wikidata ids where possible." The model fetched Wikidata item pages (and one Wikisource author page) and reported only identifiers it had seen on a page. It stopped at its tool limit after the twenty subjects, eight other people and seven schools; places were not attempted. Its response is in `opus-response.md`. Every identifier was resolved with `wbgetentities` on 7 October 2026 (`opus-scoring-raw.json`).

## The subjects

| # | Entry | Answer key | Opus | Verdict |
|---|---|---|---|---|
| 1 | Sharpe | Q459640 | Q459640 | correct |
| 2 | Williams | Q7287329 | Q7287329 | correct |
| 3 | Price | Q123279571 | Q123279571 | correct; noted the "Hamblyn" spelling |
| 4 | Archer | Q5534618 | Q5534618 | correct |
| 5 | Tibbits | none found | declined | agrees |
| 6 | Gorges | Q116445705 (pipeline) | Q121494224 | **a Wikidata duplicate**: "Edmond Gorges, South African politician (1872–1924)" and "Edmond Howard Lacam Gorges, colonial administrator in South West Africa (1872–1924)" are two items for one man. Opus also flagged the trap item Q5586263, an 18th-century Irish lawyer called Gorges Edmond Howard |
| 7 | Denton | Q19059378 | Q19059378 | correct |
| 8 | Wilkinson | Q4120852 | Q4120852 | correct |
| 9 | Prince | none found by vector search | Q47891631 Edward Ernest Prince, Canadian fisheries biologist, 1858–1936 | **correct, and new**: birth year and F.R.S.C. match; the MCP search had not surfaced it |
| 10 | Plumer | Q336040 | Q336040 | correct |
| 11 | Vaughan | none found | declined | agrees |
| 12 | Harris | Q5075044 | Q5075044 | correct; noted the entry's birth year (1858) disagrees with the item (1855) and resolved it on Christ's College and Newfoundland |
| 13 | Bland | Q20036709 | Q20036709 | correct; noted a redirect |
| 14 | Dyer | none found | declined | agrees |
| 15 | Wilson | Q20725021 | Q20725021 | correct; from a Wikisource author page, flagged as such |
| 16 | Anderson | none found | declined | agrees |
| 17 | Alldridge | Q16933745 | Q16933745 | correct |
| 18 | Marsh | Q1292947 | Q1292947 | correct |
| 19 | Belcher | Q5077823 | Q5077823 | correct |
| 20 | Clementi | Q839225 | Q839225 | correct; flagged the trap Q5056012, his uncle Cecil Clementi Smith, also a Straits governor |

**Result:** 16 of 16 identifiers resolve to the right person (one via a duplicate item). The four declines are the four the MCP search also finds nothing for. It added one correct identification the pipeline lacked (Prince), caught two near-miss traps, and flagged three OCR errors in the entries ("Cuthbert Peak Grant" for the Cuthbert Peek Award; "hurricane relic offr." for relief; an impossible "May to Apr., 1896").

## Other people and schools

All eight other people (Chamberlain, Churchill, Asquith, Lyttelton, Buxton, Selborne, Devonshire, J. H. Thomas) and all seven schools (Rossall, Queens' Cambridge, Trinity Cambridge, Harrow, Christ's Cambridge, Lincoln's Inn, Gray's Inn) resolve to the right items. Compare Gemini without tools on the same names: Chamberlain a Swiss footballer, Rossall a motocross rider, Lincoln's Inn a Spanish municipality.

## What this shows, set beside the other two columns

| | Gemini 3.8 Flash, no tools | Pipeline: vector search + gates | Opus 5.5 with retrieval |
|---|---|---|---|
| Person identifiers correct | 0 of 16 | 12 of 12 given; 8 left ungrounded, of which 4 have items | 16 of 16 given |
| Declined | 4 | 8 | 4 |
| Found that the others missed | | | Prince; a duplicate item for Gorges |
| Cost per entry | lowest | low, scales to 100,000 | highest; stopped at its tool limit before the places |

- The difference between zero and sixteen is not the model's reading of the biography. Both named the people correctly. It is whether the identifier came from a page or from memory.
- Retrieval with a strong model and a human-style search is the most accurate and the most expensive. The pipeline's gate is conservative by design (ties resolve to no link) and misses real items (Price, Wilson, Alldridge, Prince); those are what a review queue is for.
- Even retrieval needs the historian: Gorges has two items, and someone has to decide which to use and whether to merge them on Wikidata. Harris's birth year disagrees between the source and the item, and someone decided the match on the college and the colony.

## The finished file

Opus then completed the schools, the other named people and the main colonial jurisdictions and packaged everything as `opus-grounding.csv`: 202 mention rows keyed by entry id, with mention text, resolved label, QID, status and a note. Every one of the 75 identifiers in the file was resolved with `wbgetentities` on 7 October 2026 (`opus-grounding-check.json`), and every one has a label or alias consistent with the name it was attached to.

| Entity type | Verified | Probable | Not found | Not searched |
|---|---|---|---|---|
| subject | 16 | 0 | 4 | 0 |
| person | 15 | 1 | 0 | 0 |
| place | 17 | 0 | 0 | 121 |
| school | 26 | 0 | 0 | 2 |

Two details from the file worth saying in the room:

- Its first guess for George V, from memory, was Q269. Q269 is Tashkent. It checked the page, replaced it with Q269412, and left the note "Not Q269" in the file. That is the whole argument in one row.
- It recorded the near-miss items beside the right ones (Cecil Clementi Smith Q5056012; the 18th-century Gorges Edmond Howard Q5586263; a St Paul's School disambiguation page Q7595173; Robert Gordon University Q1788342 carrying the college's name as an alias). Those are the wrong answers a no-tools model is likely to produce, and they belong in the test as distractors.
- Places are where it stopped: 17 grounded, 121 normalised for lookup but not searched, because each search grounded about one entity. That is the cost line. The pipeline's vector search resolves a normalised label in about a second and runs over 3,207 distinct place surfaces in the Colonial Office List graph without anyone watching.