Demonstration
Five records, open to be argued with.
A company, a person, a clinical trial, a counterparty and a facility. These entities are fictional and the evidence behind them is written for this page. What is real is the shape: every value carries the sentence it was read from, the source, the locator and the date; every value that could not be established says so in words; and the whole record can be replayed at an earlier date.
Records differ in kind and in how much is known about them. What does not differ is what each value has to carry before it is shown to anybody.
Halden Therapeutics AG
Lead program
HLD-204
“Following the January reprioritisation, HLD-204 is our lead candidate and our principal development priority.”
- Source
- Corporate presentation, February 2026
- Read at
- slide 4
- Observed
- 11 March 2026
chosen over
- HLD-118observed 11 March 2026
haldentx.com/pipeline, page not revised since August 2025
A dated presentation outranks an unrevised web page.
Things worth trying
Four ways to test it in thirty seconds
Switch the date control on Halden Therapeutics. Four of six values differ between the two dates, and the rows that changed are marked. Neither state was overwritten to produce the other.
Open the lead program. The value that lost is still on the record, with its own source and date and a single line saying on what grounds it was not carried.
Open the Phase 2 readout date, or Northgate’s tier-three input origin. Neither is blank, and neither says no. Both list what was searched and when.
Open Meridian’s issuer rating. A rating exists; this reader is not cleared to see it; the cell says so and implies nothing about its level or direction.
Open Dr Marianne Ek’s identity. Three spellings resolve to one person and both disambiguation datasets disagree. The record carries the names, dated and attributed; none of them is the identity.
On trial NCT06712233, read the estimated primary completion and the readout date together. One is cited and has moved by four months; the other is declared not found on both dates. Carrying the first as the second is the defect this page exists to show.
On the same trial, replay to November 2025 and open the interim analysis. At that version the protocol specified none, so the property does not arise: not applicable, which is a different negative from not found.
The five records
Four kinds of thing, one set of obligations
A company, a person, a clinical trial and a facility do not raise the same questions. Each record here was written to carry a different kind of difficulty.
A value that changed for a reason: the lead program, the phase and the chief executive all differ between the two dates, and the value each was chosen over is still on the record.
Identity where a name answers worst. Three spellings, one person, two disambiguation datasets that split the record, and a date of birth this reader is not cleared to see.
Two quantities that are routinely confused. The estimated primary completion is cited and has moved; the readout date is declared not found, on the same record, on both dates.
A group reorganization. The parent changes, the former name is retained and still resolves, and an issuer rating exists behind a licence this reader does not hold.
Three transliterations resolving to one permit, a parent that arrives only when the accounts are filed, and a tier-three origin that stays declared not found after five searches.
The same answer, read by a program
What an answer looks like
A value never arrives on its own. It arrives with the sentence it was read from, where that sentence is, when it was read, when it entered the record, what it was chosen over, and which reader the answer was assembled for. A declared absence arrives in the same shape, and so does a restriction.
This is an answer, not a schema. The properties are the ones the four guarantees require; there is no arrangement of them that omits the evidence.
$ rc get company/haldentx.com --as-of 2026-03-14
record Halden Therapeutics AG
anchor haldentx.com
reader Research analyst, public-sources clearance
as of 14 March 2026
values 6 (4 cited, 1 not found, 1 blocked)
lead_program
value HLD-204
status cited
excerpt "Following the January reprioritisation,
HLD-204 is our lead candidate and our
principal development priority."
source Corporate presentation, February 2026
locator slide 4
observed 2026-03-11
entered 2026-03-11
chosen_over HLD-118
haldentx.com/pipeline, not revised since
August 2025, observed 2026-03-11
"A dated presentation outranks an
unrevised web page."
phase_2_readout_date
value not_found
status declared
reason No source we hold states a readout date.
The registry's estimated primary
completion is a different quantity.
searched 4 sources, last 2026-03-13
mechanism
value blocked
status declared
source J. Med. Chem. 69(4), subscription
policy licensed source, not permitted for
reader clearance public-sources
attempted 2026-03-12Downloads
Take a sample away.
The audit export of the fictional Halden Therapeutics AG record, as a reader would be sent it, and the written scope that would precede it.
Sample, fictional data
Sample written scope
The document fixed in writing before work starts, for the fictional Halden Therapeutics AG: first record, permitted sources, reader clearance, deployment boundary, model routing, retention, export. No price or duration.
Sample, fictional data
Sample audit export
What an auditor receives for the same fictional record: each value with its excerpt, source, locator, dates, what it was chosen over and the reader clearance, plus declared negatives and blocked values.
Bring your own entity.
The demonstration is fictional on purpose. The useful version of this conversation starts with one entity you already argue about internally, and the sources you already pay for.
Further reading
Articles on this page’s subject.
How to Evaluate LLM Citations at the Claim Level
Evaluating LLM citations by atomic claim, with support and completeness metrics, confidence intervals and perturbation tests.
Valid Time vs Transaction Time: Which Clock Proves What?
Valid, transaction, observation and publication time distinguished through a delayed correction and four as-of tests.
W3C PROV for LLM Context Provenance: A Testable Profile
An application profile of W3C PROV for reconstructing the evidence behind LLM context, with mandatory mappings, source-state fields and mutation tests.