caiotheodoro/titer-expertise-claims

titer · attested expertise claims

99,984 expertise claims attested by publication record, for 20,000 researchers with an ORCID iD. Built to measure whether a people-search provider can tell a real expert from a claimed one.

Sources: OpenAlex (disambiguated authors, works, topics), ORCID (the identity spine), Crossref DOIs (the attestation chain). All open. A researcher cannot self-assert a DOI into existence, which is why authorship is treated as attested where a self-reported affiliation is not.

Column	Meaning
author_id	OpenAlex author identifier
orcid	ORCID iD
name_sha256	Salted hash. No names are published.
works_count	Total works, banded 20–500 (see below)
topic_id, topic_label	OpenAlex topic
n_works_in_topic	Attested works in it, ≥ 3
subfield, field, domain	The OpenAlex hierarchy
polarity	Always ATTESTED, see the next section

What it measured

Exa /answer, n=1,000 across four strata of 250, negatives verified false against OpenAlex /works before the provider was asked:

Claim	Correct	False affirmation
Attested	0.9600 [0.9279, 0.9781]	0 [0, 0.0151]
False, adjacent subfield (NEAR)	0.8320	0.1680 [0.1268, 0.2193]
False, adjacent field (FAR)	0.8680	0.1320 [0.0956, 0.1796]
False, different domain (CONTROL)	0.9160	0.0840 [0.0556, 0.1250]

It confirms real expertise 96% of the time and affirms topics the person has never published in one time in six. NEAR is 2.00× the control with non-overlapping intervals, so topic distance moves the rate; the control still does not reach zero, which puts a floor of at least 8.4 points under the 16.8 that no reading of the prompt explains. Stated confidence is 0.987 against 0.889 accuracy, ECE 0.1153.

One provider. Nothing here supports a claim about the category.

Are the false affirmations grounded in anything?

The strongest objection is that the provider may read "expertise in T" loosely as "works near T" - the wrong question rather than a wrong answer. Every answer comes back with citations, so for all 399 affirmations those citations were resolved against OpenAlex and asked a mechanical question: is any of them a work this author actually wrote? No model reads the evidence text.

Claim affirmed	n	cites a real work by that author
Attested	150	0.5733 [0.4933, 0.6497]
False, adjacent field	102	0.2843 [0.2058, 0.3784]
False, adjacent subfield	126	0.2302 [0.1653, 0.3110]
False, different domain	21	0.0000 [0.0000, 0.1546]

The control arm cites nothing verifiable at all, zero of twenty-one, and it is already the arm no generous reading of the prompt explains. But roughly one in four adjacent false affirmations is grounded in the person's real work, so the objection survives for near misses and collapses only at the extreme. Both halves are reported.

The negatives are a rule, not a file

The benchmark is a balanced discrimination between an attested claim and a constructed-false one: a topic in which the person has provably never published. Those negatives are deliberately not published. Distributing a list of real, named researchers paired with things they cannot do would publish an artefact of a test as though it were a finding about a person. The rule is published instead, and it is fully deterministic:

1. Take a topic from the same field, different subfield (NEAR), the same domain, different field (FAR), or a different domain (CONTROL).
2. Exclude catch-all labels. OpenAlex carries unfalsifiable buckets such as "Diverse Scientific Research Studies", and a negative nobody can be wrong about measures nothing.
3. Verify against /works that the author has zero works in it. This step is not optional: an author's topics field is a top-N summary (median 5 topics for people with ~120 works), not an exhaustive record. Skipping it made 13.3% of an earlier corpus's "false" claims true, the same magnitude as the effect being measured. That run was retracted.

No model participates at any point. Adjacency is a set operation over the OpenAlex hierarchy.

Known limits

• Publication record is not expertise. An engineer, a practising clinician or a lawyer who never publishes is invisible here, and that is most of who an expert-sourcing platform actually hires. This is the largest gap.
• 20,000 of ~3.66M eligible ORCID authors (0.55%). Name-collision degree measured inside the corpus is a sampling artefact and under-counts reality.
• works_count banded 20–500, excluding both merged author records (one real row carries 2.4M works across 42 affiliations) and early-career researchers.
• Domain skew: Physical Sciences dominates; Social Sciences is thin.
• OpenAlex's own author-disambiguation error rate is the contamination bound.

Intended use

Evaluating expertise verification and identity resolution. Not for locating or contacting individuals, and not as a judgement about any person's competence.

Rectification

Removal on request within 7 days, no justification required, recorded as a count and never as an identity.

Code, pre-registration and results: https://github.com/caiotheodoro/titer