Cultural Heritage Benchmark — TMK Research Lab

Project Overview

Frontier AI browsing systems can answer many questions by locating obvious keywords or matching information from highly ranked webpages. This project explores how to construct a research question that cannot be solved through a single search. The answer must remain short, stable and objectively verifiable, while the clues require the researcher to connect different people, dates, places, organisations, publications, events, records and quantities.

Research Goal

Design one difficult but fair browsing challenge.

Required Answer

A short, stable and objectively verifiable fact.

Evidence Standard

Every clue must be independently supported by reliable evidence.

Final Deliverable

A question, answer, clue set, source trail, validation log and JSON record.

AIBRC-007

The Research Challenge

Challenge ID: AIBRC-007
Research domain: Art History and Museum Records
Difficulty level: Advanced
Date created: 4 September 2026
Last verified: 4 September 2026
Expected answer type: Short text
Research question: An Italian glazier who had worked for a Paris museum walked out of it early one Monday morning wearing a staff smock, taking a painting that was only recovered after his arrest in December 1913. On what date did he take it?
Correct answer: 21 August 1911
Answer-normalisation rules: Case-insensitive; punctuation ignored. Accepted variations: 21 August 1911; August 21, 1911; 21/08/1911; 1911-08-21.
Confidence level: Medium-high — day-level date from two corroborating records, year confirmed by the museum
Challenge status: Complete and verified

Portfolio integrity note: Every claim on this page was checked against the cited source before publication. No source, quotation, locator or search result is invented; where a detail could not be verified directly it is described rather than quoted.

Research Method

Step 1 — Establish the Answer
Begin with a stable fact that can be confirmed through a reliable primary or authoritative source. Record the exact wording, document title, issuing organisation and publication date.
Step 2 — Build Independent Clues
Develop clues involving different fact types, including dates, people, places, organisations, works, events, records and quantities. Each clue must be independently checkable.
Step 3 — Test Discoverability
Run obvious, broad and targeted search queries. Record the results to identify accidental shortcuts, direct answer leakage and misleading pages.
Step 4 — Construct the Evidence Chain
Explain how each source supports a specific clue and how the clues logically combine to identify one answer.
Step 5 — Validate and Package
Recheck every source, confirm answer uniqueness, document limitations and export the completed benchmark record in a consistent JSON format.

Clue Matrix

Each clue constrains the answer from a different direction. No single row is sufficient on its own, and every row can be checked against a source another researcher can reach.

Verified clue set

Statuses are editable so the matrix can be worked through row by row. Scroll sideways on small screens.

Clue matrix listing clue identifier, fact type, statement, constraint, supporting source, exact locator and verification status.
Clue ID	Fact type	Clue statement	Constraint	Supporting source	Exact locator	Verification status
C01	Person	The man responsible had worked at the museum as a glazier and was Italian.	One named individual: Vincenzo Peruggia.	Louvre collections record for the painting	Historical note on the work	VerifiedVerification status for clue C01
C02	Organisation	The institution is a Paris museum whose own catalogue records the theft.	Confirms the venue from the holding institution rather than from press retellings.	Louvre collections record for the painting	Same note	VerifiedVerification status for clue C02
C03	Method	The thief entered early in the morning wearing the white smock worn by museum staff.	Detail attributed to his own later interrogation, not to legend.	Wikipedia — Vincenzo Peruggia	Section on the theft, citing the Florence interrogation	VerifiedVerification status for clue C03
C04	Date	He was arrested more than two years later, on 12 December 1913.	Fixes the recovery date, which is a common wrong answer.	Wikipedia — Vincenzo Peruggia	Same section	VerifiedVerification status for clue C04
C05	Day of week	The theft took place on a Monday.	Combined with the month and year this leaves a single candidate date.	Wikipedia — Vincenzo Peruggia	Interrogation account: “Monday, 21 August 1911, at around 7 am”	VerifiedVerification status for clue C05

Evidence Records

Each card records a source that was opened and read before it was cited. Excerpts are quoted only where the wording was read directly; otherwise the record describes what the source shows.

E01

Verified

Claim supported: Clues C01 and C02 — the thief and the institution.
Source title: Portrait de Lisa Gherardini, épouse de Francesco del Giocondo
Author / issuing institution: Musée du Louvre (collections catalogue)
Source type: Museum catalogue record
Publication date: Standing catalogue entry
Direct URL: https://collections.louvre.fr/en/ark:/53355/cl010062370
Page number: n/a (catalogue record)
Table, figure or section: Historical note
Relevant excerpt: “Le vol de l’œuvre par Vincenzo Peruggia, ouvrier vitrier au musée du Louvre, en 1911, lui a donné une célébrité mondiale…”
Researcher's interpretation: The holding institution names the thief, his job and the year — but not the day, so it corroborates without leaking the answer.
Access date: 4 September 2026
Archive URL: [Not archived yet]
Classification: Primary
Reliability assessment: High — the institution that holds the work

E02

Verified

Claim supported: The answer, and clues C03, C04 and C05.
Source title: Vincenzo Peruggia
Author / issuing institution: Wikipedia
Source type: Reference database
Publication date: Continuously edited article
Direct URL: https://en.wikipedia.org/wiki/Vincenzo_Peruggia
Page number: n/a (web page)
Table, figure or section: Theft and arrest
Relevant excerpt: “…best known for stealing the Mona Lisa from the Louvre… on 21 August 1911” and “…he entered the Louvre on Monday, 21 August 1911, at around 7 am, through the door where the other Louvre workers were entering, wearing one of the white smocks…”
Researcher's interpretation: Gives the day, the weekday and the arrest date, all attributed to the interrogation record.
Access date: 4 September 2026
Archive URL: [Not archived yet]
Classification: Secondary
Reliability assessment: Medium-high — corroborated by the article on the painting

E03

Verified

Claim supported: Independent corroboration of the date.
Source title: Mona Lisa
Author / issuing institution: Wikipedia
Source type: Reference database
Publication date: Continuously edited article
Direct URL: https://en.wikipedia.org/wiki/Mona_Lisa
Page number: n/a (web page)
Table, figure or section: Theft and vandalism
Relevant excerpt: “On 21 August 1911, the painting was stolen from the Louvre.”
Researcher's interpretation: A second, separately maintained article giving the same date, checked directly rather than assumed.
Access date: 4 September 2026
Archive URL: [Not archived yet]
Classification: Secondary
Reliability assessment: Medium

Search Validation Record

A strong benchmark documents the searches a capable researcher or browsing agent would naturally attempt, including the ones that exposed the answer. Leakage is recorded rather than hidden, and the question is rewritten around it.

Direct wording searchKeyword combinationExact-phrase searchPerson and date searchOrganisation and location searchDomain-restricted searchPDF-only searchGovernment-site searchInstitutional-database searchAlternative spelling search

4 searches run and recorded

Search validation record.
Search number	Exact search query	Search engine or database	Date searched	Top relevant results	Answer revealed?	Problem discovered	Action taken
S01	when was the Mona Lisa stolen	Google	4 September 2026	Exposes 21 August 1911 in the snippet.	Yes	The plain question is answered without opening anything.	The question was rewritten to avoid naming the painting, the museum and the word 'stolen'.
S02	Louvre glazier Italian arrested Florence 1913 painting returned	Google	4 September 2026	Returns material on the recovery; the 1911 date appears only after reading a record.	No	None — this is the intended path.	Retained as the reference path.
S03	musée du Louvre catalogue vol 1911 Peruggia	Google	4 September 2026	Reaches the Louvre catalogue record used as E01, which gives the year but not the day.	No	The institutional record alone is not sufficient to answer.	Documented; it demonstrates that the chain needs a second source to reach a day-level answer.
S04	"12 December 1913" Mona Lisa arrest date	Google	4 September 2026	Returns the recovery date, the most common wrong answer.	No	Confirms the decoy is live in the search results.	Clue C04 states the arrest date explicitly so it cannot be mistaken for the answer.

Source Quality Assessment

Every candidate source is scored from 1 to 5 against five criteria before it is allowed into the evidence chain.

1Authority
3 / 5
Who created or issued the information?
2Proximity
3 / 5
Is the source directly connected to the event or record?
3Specificity
3 / 5
Does it support the precise claim being made?
4Stability
3 / 5
Is the information unlikely to change?
5Traceability
3 / 5
Can another researcher locate the same evidence?

Evidence-Chain Audit

Export a report that maps every clue to the source it rests on, records the verification status and access date of each source, and lists the searches that exposed the answer.

Claims audited: 5
Sources in the chain: 3
Disclosed shortcuts: 1

Structured Benchmark Record

The complete challenge is delivered as a single validated JSON document.

aibrc-007-benchmark-record.json

{
  "challenge_id": "AIBRC-007",
  "title": "Cultural Heritage Benchmark",
  "project_type": "independent_demonstration",
  "research_domain": "Art History and Museum Records",
  "difficulty": "advanced",
  "question": "An Italian glazier who had worked for a Paris museum walked out of it early one Monday morning wearing a staff smock, taking a painting that was only recovered after his arrest in December 1913. On what date did he take it?",
  "accepted_answers": [
    "21 August 1911"
  ],
  "answer_type": "short_text",
  "answer_normalisation": {
    "case_sensitive": false,
    "ignore_punctuation": true,
    "accepted_variations": [
      "21 August 1911",
      "August 21, 1911",
      "21/08/1911",
      "1911-08-21"
    ]
  },
  "clues": [
    {
      "clue_id": "C01",
      "fact_type": "person",
      "statement": "The man responsible had worked at the museum as a glazier and was Italian.",
      "constraint": "One named individual: Vincenzo Peruggia.",
      "source": "Louvre collections record for the painting",
      "locator": "Historical note on the work",
      "status": "verified"
    },
    {
      "clue_id": "C02",
      "fact_type": "organisation",
      "statement": "The institution is a Paris museum whose own catalogue records the theft.",
      "constraint": "Confirms the venue from the holding institution rather than from press retellings.",
      "source": "Louvre collections record for the painting",
      "locator": "Same note",
      "status": "verified"
    },
    {
      "clue_id": "C03",
      "fact_type": "method",
      "statement": "The thief entered early in the morning wearing the white smock worn by museum staff.",
      "constraint": "Detail attributed to his own later interrogation, not to legend.",
      "source": "Wikipedia — Vincenzo Peruggia",
      "locator": "Section on the theft, citing the Florence interrogation",
      "status": "verified"
    },
    {
      "clue_id": "C04",
      "fact_type": "date",
      "statement": "He was arrested more than two years later, on 12 December 1913.",
      "constraint": "Fixes the recovery date, which is a common wrong answer.",
      "source": "Wikipedia — Vincenzo Peruggia",
      "locator": "Same section",
      "status": "verified"
    },
    {
      "clue_id": "C05",
      "fact_type": "day of week",
      "statement": "The theft took place on a Monday.",
      "constraint": "Combined with the month and year this leaves a single candidate date.",
      "source": "Wikipedia — Vincenzo Peruggia",
      "locator": "Interrogation account: “Monday, 21 August 1911, at around 7 am”",
      "status": "verified"
    }
  ],
  "evidence": [
    {
      "evidence_id": "E01",
      "claim_supported": "Clues C01 and C02 — the thief and the institution.",
      "source_title": "Portrait de Lisa Gherardini, épouse de Francesco del Giocondo",
      "issuing_organisation": "Musée du Louvre (collections catalogue)",
      "source_type": "Museum catalogue record",
      "url": "https://collections.louvre.fr/en/ark:/53355/cl010062370",
      "publication_date": "Standing catalogue entry",
      "page_number": "n/a (catalogue record)",
      "section": "Historical note",
      "access_date": "4 September 2026",
      "reliability": "High — the institution that holds the work",
      "verification_status": "verified"
    },
    {
      "evidence_id": "E02",
      "claim_supported": "The answer, and clues C03, C04 and C05.",
      "source_title": "Vincenzo Peruggia",
      "issuing_organisation": "Wikipedia",
      "source_type": "Reference database",
      "url": "https://en.wikipedia.org/wiki/Vincenzo_Peruggia",
      "publication_date": "Continuously edited article",
      "page_number": "n/a (web page)",
      "section": "Theft and arrest",
      "access_date": "4 September 2026",
      "reliability": "Medium-high — corroborated by the article on the painting",
      "verification_status": "verified"
    },
    {
      "evidence_id": "E03",
      "claim_supported": "Independent corroboration of the date.",
      "source_title": "Mona Lisa",
      "issuing_organisation": "Wikipedia",
      "source_type": "Reference database",
      "url": "https://en.wikipedia.org/wiki/Mona_Lisa",
      "publication_date": "Continuously edited article",
      "page_number": "n/a (web page)",
      "section": "Theft and vandalism",
      "access_date": "4 September 2026",
      "reliability": "Medium",
      "verification_status": "verified"
    }
  ],
  "search_validation": [
    {
      "search_id": "S01",
      "query": "when was the Mona Lisa stolen",
      "platform": "Google",
      "search_date": "4 September 2026",
      "relevant_results": "Exposes 21 August 1911 in the snippet.",
      "answer_exposed": true,
      "notes": "The plain question is answered without opening anything. The question was rewritten to avoid naming the painting, the museum and the word 'stolen'."
    },
    {
      "search_id": "S02",
      "query": "Louvre glazier Italian arrested Florence 1913 painting returned",
      "platform": "Google",
      "search_date": "4 September 2026",
      "relevant_results": "Returns material on the recovery; the 1911 date appears only after reading a record.",
      "answer_exposed": false,
      "notes": "None — this is the intended path. Retained as the reference path."
    },
    {
      "search_id": "S03",
      "query": "musée du Louvre catalogue vol 1911 Peruggia",
      "platform": "Google",
      "search_date": "4 September 2026",
      "relevant_results": "Reaches the Louvre catalogue record used as E01, which gives the year but not the day.",
      "answer_exposed": false,
      "notes": "The institutional record alone is not sufficient to answer. Documented; it demonstrates that the chain needs a second source to reach a day-level answer."
    },
    {
      "search_id": "S04",
      "query": "\"12 December 1913\" Mona Lisa arrest date",
      "platform": "Google",
      "search_date": "4 September 2026",
      "relevant_results": "Returns the recovery date, the most common wrong answer.",
      "answer_exposed": false,
      "notes": "Confirms the decoy is live in the search results. Clue C04 states the arrest date explicitly so it cannot be mistaken for the answer."
    }
  ],
  "quality_checks": {
    "The answer is short and objectively verifiable.": true,
    "The answer is supported by an authoritative source.": true,
    "The question has only one defensible answer.": true,
    "No single clue directly exposes the answer.": true,
    "Each clue is independently checkable.": true,
    "The clue set spans multiple fact types.": true,
    "The task requires multi-step reasoning.": true,
    "Obvious searches have been documented.": true,
    "Search-result snippets do not reveal the answer.": false,
    "Sources are publicly accessible.": true,
    "Exact pages, tables or sections are recorded.": true,
    "The evidence chain can be reproduced.": true,
    "The question avoids subjective interpretation.": true,
    "The answer is stable over time.": true,
    "No fabricated evidence is included.": true
  },
  "researcher": {
    "name": "TMK Research Lab",
    "portfolio": "https://theophilusmorkporkpor.lovable.app",
    "email": "tkudjordjie@gmail.com"
  }
}

Skills Demonstrated

Open-Web Research
Investigative Research
Fact-Checking
Source Verification
Primary-Source Research
Institutional Databases
Government Records
Archive Research
PDF Investigation
Search Strategy
Evidence Documentation
Research Problem Design
Benchmark Construction
AI Evaluation
Adversarial Testing
Data Validation
JSON
Structured Data
Technical Writing
Quality Assurance

Research Tools

Google Search and advanced search operators
Google Scholar
Google Books
Google Maps
Government websites
Institutional databases
Digital archives
Public registries
PDF search
Internet Archive
Spreadsheets
Microsoft Word
Google Docs
JSON validators
Browser developer tools
AI systems used for testing, not as final evidence

AI-generated statements are never treated as evidence. Every factual claim must be verified against an accessible external source.

Project Outcome

This benchmark is complete: the answer is fixed, 3 evidence records are attached to specific clues, and 4 validation searches are documented, including those that exposed the answer. It demonstrates the ability to approach an unfamiliar subject systematically, separate evidence from assumption and document research so that another person can audit it.

Reflection and Limitations

What made the question difficult

Two dates circulate — the 1911 theft and the December 1913 arrest — and the museum's own catalogue gives only the year.

Which searches failed

The Louvre catalogue search reached an authoritative record that could not answer the question on its own.

Which source proved decisive

The account of the Florence interrogation, because it fixes the weekday and the hour as well as the date.

How ambiguity was removed

Naming the arrest date inside the question removes the most common wrong answer before the solver starts.

What an AI browsing agent might get wrong

Returning December 1913, or giving 1911 without a day because it stopped at the museum record.

Remaining limitations

The day-level evidence is encyclopaedic rather than archival; a digitised French police or press record from August 1911 would strengthen it.

Recommended future improvements

Add a contemporary Paris newspaper page from the days after the theft, retrieved from a national library archive.

Need a careful researcher for complex web investigations?

I am available for web research, fact-checking, AI evaluation, benchmark construction, source verification and structured research documentation.

Start a research enquiry View My Resume Email Me

TMK Research Labtkudjordjie@gmail.com https://theophilusmorkporkpor.lovable.app https://www.linkedin.com/in/theophilus-kudzordzi

Project Overview

Research Goal

Required Answer

Evidence Standard

Final Deliverable

The Research Challenge

Research Method

Step 1 — Establish the Answer

Step 2 — Build Independent Clues

Step 3 — Test Discoverability

Step 4 — Construct the Evidence Chain

Step 5 — Validate and Package

Clue Matrix

Evidence Records

Search Validation Record

Source Quality Assessment

1Authority

2Proximity

3Specificity

4Stability

5Traceability

Evidence-Chain Audit

Benchmark Quality Checks

Structured Benchmark Record

Skills Demonstrated

Research Tools

Project Outcome

Reflection and Limitations

What made the question difficult

Which searches failed

Which source proved decisive

How ambiguity was removed

What an AI browsing agent might get wrong

Remaining limitations

Recommended future improvements

Need a careful researcher for complex web investigations?