How the review is produced
GrantWise is a pre-submission quality-assurance tool. It produces a simulated assessment aligned with published evaluation criteria. It is not an official evaluation and does not predict funding.
1. Review contract
The selected programme loads a version-controlled review profile, evaluator method and drafting method. Uploaded official call material raises the grounding level from programme-level to call-material.
2. Deterministic preflight
Before AI assessment, GrantWise checks the document for expected structural and evidence signals. A missing search signal is never treated as proof of failure; it only directs the reviewer to verify the source.
3. Evidence ledger
Each weakness must include a proposal location, supporting evidence, applicable requirement or call basis, severity, correction and re-review acceptance condition. Findings without proposal support are prohibited.
4. Scoring and challenge
Horizon scores use the most conservative supported result from the model judgement, graded findings and atomic scorecard. An unusually favourable result triggers a separate challenge pass. A finding count is not a target: a strong proposal may legitimately have a short ledger.
5. Confidence and limitations
Reports show high, moderate or limited confidence based on scorecard traceability and call grounding. Eligibility, scientific validity, legal claims and financial assumptions still require qualified human verification.
6. Revision comparison
A later version can be matched using non-reversible fingerprints and similarity statistics. The report separates resolved, persisting, worsened and new issues without retaining proposal text.
Benchmark plan
Status: validation in progress. GrantWise does not yet claim a statistically sufficient accuracy result.
- Collect 50–100 proposals with official ESRs under explicit permission.
- Blind GrantWise to the official result and run the applicable call profile.
- Measure criterion-score error, threshold classification, unsupported-finding rate and run-to-run variance.
- Ask two experienced reviewers to blind-rate usefulness and factual grounding.
- Publish aggregate results, sample size, programme mix and known failure modes.