Evidence independence as a fourth probe dimension: fixture and argument - #107
Merged
Merged
Conversation
NIST's agentic-research-evaluation-probes scores every citation with three LM-judge probes -- faithfulness, completeness, sufficiency -- and its docs/probes.md states the three are "mutually exclusive, collectively exhaustive" for how a citation is used. All three are per-citation: each judge is shown one source passage and one citing sentence. A judge that sees one citation at a time cannot observe a relation between citations, so it cannot observe that three cited documents descend from one observation. Each citation is faithful, complete and sufficient on its own, and the report still presents one observation as three. The corpus is three documents: one measurement and two secondary reports that each state in their own text that they made no measurement. Deliberately honest -- nothing misquotes, drops caveats or overreaches -- because the three existing probes are supposed to score them well. independence_demo.py runs the provenance-aware aggregator over the three citations and prints 3 citations against 1 root. It is deterministic and needs no model: the independence side requires no judge. What is NOT here, and the README says so plainly: the three existing probes have not been run on this corpus. They need an OpenAI-compatible endpoint and no run has been performed. The discriminating experiment is written down with both outcomes named -- all three scoring well supports a fourth dimension; any of them flagging the derivatives refutes it and should be recorded here rather than discarded. Until then the argument is structural, not empirical, and nothing should be proposed to NIST on it. Their architecture already admits the fourth probe: ProbeFunc receives the section's full list[Finding] and SectionVerdict exists for per-section probes. Read from the NIST repository on 18 Aug 2026. Nothing sent anywhere. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixture and argument. Nothing sent to NIST, nothing asks NIST to adopt Minority Prophet.
The gap
usnistgov/agentic-research-evaluation-probesscores each citation with three LM-judge probes — faithfulness, completeness, sufficiency — and itsdocs/probes.mdcalls them "mutually exclusive, collectively exhaustive" for how a citation is used.All three are per-citation: each judge sees one source passage and one citing sentence. A judge shown one citation at a time cannot observe a relation between citations, so it cannot see that three cited documents descend from one observation. Each is faithful, complete and sufficient on its own — and the report still presents one observation as three.
Not a defect in the probes. A dimension outside the unit they evaluate. Their architecture already admits the fix:
ProbeFuncreceives the section's fulllist[Finding], andSectionVerdictexists for per-section probes.What is demonstrated
corpus/holds one measurement plus two secondary reports, each stating in its own text that it made no measurement. The documents are deliberately honest — nothing misquotes, drops caveats, or overreaches — because the existing probes are supposed to score them well.Deterministic, no model, no network. The independence side needs no judge.
What is NOT demonstrated
The three existing probes have not been run on this corpus. They need an OpenAI-compatible endpoint; no run has been performed. This PR does not claim they score well here — only that they are structurally unable to see the derivation.
The discriminating experiment is written down with both outcomes named:
Until that run happens the argument is structural, not empirical. Proposing this to NIST beforehand would be asserting a measurement nobody made.
Verified vs not
Verified at source (18 Aug 2026): the repo exists, is public, is not archived, and its probe docs say what is quoted. Not verified: anything about NIST's intentions or receptiveness. Last push was 15 Apr 2026, so "ongoing" should not be assumed.
🤖 Generated with Claude Code