Baseline
Ordinary source-packet research without an added evidence-control scaffold.
Active empirical study
Do structured workflow scaffolds reduce unsupported claims?
The study tests whether procedural controls for provenance, uncertainty, disconfirmation, and final claim auditing improve source support in AI-assisted research without winning by making answers empty, vague, or excessively cautious.
Research question
Is unsupported output partly a workflow reliability problem?
The proposal treats unsupported claims as errors that can survive several stages: source review, synthesis, inference, drafting, and editing. The intervention is procedural rather than rhetorical.
The study does not assume that extra structure works. It tests whether the controls improve actual claim support or merely produce more convincing process artifacts.
Controls under test
- Source-to-claim binding and provenance
- Explicit uncertainty and inference labels
- Counterevidence or disconfirmation checks
- Final claim audit with removal or downgrade of weak claims
Experimental design
Four conditions separate formatting from evidence discipline
The official measurement unit is the substantive claim extracted from the visible final answer. Model-generated scaffold tables are treated as model self-report, not as the authoritative claim registry.
Format only
Visible claim structure without provenance, audit, or disconfirmation discipline.
Provenance scaffold
Claims identify source, inference, uncertainty, or lack of support.
Full scaffold
Provenance, disconfirmation, uncertainty labels, final audit, and removal or downgrade of weak claims.
Measures
The study must not win by saying less
Support measures are paired with usefulness, coverage, false-caution, inspectability, and cost measures.
Primary measures
- Unsupported-claim rate
- Source-attribution accuracy
- Overconfidence
- Counterevidence recall
- Claims corrected or removed during audit
Guardrail measures
- Usefulness and coverage
- False caution
- Inspectability
- Reviewer time-to-problem
- Tokens, model calls, and runtime
Claim verdicts
- Supported
- Partially supported
- Reasonable inference
- Unsupported
- Contradicted or not checkable
Measurement architecture
Generation, candidate evidence, and support judgment remain separate.
The Research Scaffold Harness runs experimental conditions and preserves traceable run artifacts. Evidence Bundler nominates candidate passages and seals evidence bundles. Claim Audit Lab applies controlled support verdicts for later human calibration.
The harness does not verify claims. Candidate passage coverage is not a support verdict. CAL is a measurement channel, not automatic ground truth.
Integrity requirements before a result
- The official answer surface is defined
- Zero-answer and malformed outputs are dispositioned
- Claim extraction is calibrated against human coding
- Support verdicts are human-calibrated
- Coverage and false-caution tradeoffs are measured
Current interpretation
No valid scaffold-effect conclusion is claimed yet.
The central causal question remains open. The study will not report a confirmatory difference until the answer-surface protocol, claim extraction, candidate evidence bundles, support judgments, and human calibration are sufficiently settled.
A positive, negative, mixed, or false-caution result would all be informative. The objective is to identify whether the bundled workflow creates real reliability gains and, if so, which components are necessary.
Publication restraint: completing the apparatus or producing a matrix of outputs does not establish that the intervention works. Method progress and causal evidence are different claims.
Need a research question treated with the same controls?
The Evidence Intelligence Research Sprint applies a bounded protocol, source register, competing interpretations, explicit uncertainty, and a decision-focused briefing to a client question.