Public Gap Register
Open questions and the evidence that would let a contributor move them forward.
GAP-001 · Missing record What training was verified and what was paid for Oxfordshire Wave 6, Lot 3?
Closest work. The award, signing and £120,000 excluding VAT value are documented. The nearby scrutiny percentages describe Wave 5.
Uncertainty and what it affects. The applicable performance baseline, verified Lot 3 delivery and actual payments remain open. These determine what can be reconciled against the award.
Evidence or contribution needed. A buyer-validated aggregate with denominators and milestone-linked payment entries, checked against the signed schedules.
Oxfordshire two-turn benchmark. Test wave attribution, value versus payment, disclosure scope and partial updates.
Working paper · v0.1 · Evidence audit project · Read-only manuscript outline
Initial register entry · v0.1 · 15 November 2026
Search update · 15 November 2026. Five public search routes checked. The next step is a targeted search for the signed schedules, buyer-validated lot return and milestone-linked payment records. The DfE maximum area allocation of £2,366,600, Phase 1 tender pack's Wave 6 budget of £2,082,040 excluding VAT and Lot 3's £120,000 published value describe different stages and scopes. Their differences establish neither actual expenditure nor under-delivery.
The October 2024 Ofsted inspection describes provider-wide delivery at that time: predominantly online, with no subcontractors listed. It predates the Lot 3 award. National Wave 6 statistics are scheduled for winter 2027–28; their reported area-level scope leaves the lot-level question open. The successor AI in Business tender is Wave 7, Lot 2, estimated at £108,500 excluding VAT. Its cohort figures require a separate funding reference. No indexed Lot 3 change, termination or FOI record was located in the search. Access to the full procurement record was limited.
DfE allocations · Wave 6 tender pack · Ofsted inspection · Statistics timetable · Wave 7 tender
GAP-002 · Unresolved rule or concept What makes an independent review meaningful when a decisive source is withheld?
Closest work. The HS-026 fictional case separates a factual finding from the ability to trace a disputed assertion.
Uncertainty and what it affects. What an independent reviewer can establish and what the affected person needs to contest a material error remain distinct questions.
Evidence or contribution needed. Test specific disclosure and review safeguards against hard cases and jurisdiction-specific authorities.
HS-026 three-arm comparison · Study design in development.
Read-only jurisprudence outline · Project and research question
Initial register entry · v0.1 · 15 November 2026
GAP-003 · Measurement gap What local-language and offline evidence does AI-assisted research miss?
Closest work. The coverage audit distinguishes connectivity, language, source discoverability and comparability.
Uncertainty and what it affects. Whether adding omitted records changes a particular conclusion, and whether the model can identify that risk before seeing those records.
Evidence or contribution needed. Compare matched questions across a digital corpus and a supplemented local-language and archival corpus.
Coverage and proxy-bias probe · Concept.
Related read-only literature-search outline · Related project
Initial register entry · v0.1 · 15 November 2026
GAP-004 · Candidate research gap Can reviewers distinguish independent observations from AI echoes when source histories are incomplete?
Closest work. The HS-026 designs test explicit lineage, uncertain lineage and shared feeds. A targeted literature check remains open.
Uncertainty and what it affects. Whether partial signals distinguish real corroboration, shared upstream observation and copied summaries without over-claiming.
Evidence or contribution needed. Map prior methods and test them on matched cases with known source histories and selectively missing links.
Lineage discrimination · Study design in development.
Read-only false-corroboration outline · Project and research question
Initial register entry · v0.1 · 15 November 2026
Initial literature check · 15 November 2026. Source dependence and text reuse are established research areas. The remaining candidate is a comparative test on administrative bundles with missing lineage that scores both false grouping and unwarranted claims of independence, separately from factual correction.
Dong et al. (2009), source dependence · Leskovec et al. (2009), phrase propagation · W3C PROV (2013) · Evidence judged convincing by language models (2024) · GRACE (2026)
Intelligence and audit literature check · 15 November 2026. UK intelligence doctrine explicitly warns about circular reporting. Analysis of Competing Hypotheses compares how evidence bears on rival explanations, and audit standard AS 1105 distinguishes direct and independent evidence from indirect evidence and copies. Dong et al. (2009, §2) explicitly define dependence to include derivation from a common source, while restricting their tractable detection model to direct copying. The ACL 2024 ConflictingQA paper studies which evidence models find convincing; it supplies adjacent work rather than a direct test of shared-origin repetition.
Candidate gap retained. The sources checked did not establish a matched evaluation on small administrative bundles with selectively missing source links. Compare independent observations of the same event, reports drawing on one feed and derivative echoes with a hidden root. Include an independent observation of a different event as a control. Score false independence, false grouping and justified uncertainty separately from factual correction. Use a simple source index and a competing-hypotheses matrix as comparators.
Dong et al. (2009), §2 · UK Joint Doctrine Publication 2-00 · Heuer, Analysis of Competing Hypotheses · PCAOB AS 1105 · ConflictingQA (2024)
GAP-005 · Measurement gap Can readers trace a trial result to the planned outcome, analysis population and follow-up?
Closest work. The 9 June 2020 RECOVERY analysis plan and dexamethasone report align on 28-day all-cause mortality. The comparison randomised 2,104 participants to dexamethasone and 4,321 to usual care. Respiratory support was a planned subgroup, and completed follow-up forms are a separate count.
Uncertainty and what it affects. The current platform registry includes later comparisons. The original registry wording and any claim about formal approval or access to comparative data require dated evidence specific to the 2020 comparison. Study 329 is the selected reporting contrast.
Evidence or contribution needed. Retrieve the initial dated registry records and relevant protocol amendments. Check analysis-plan approval and review records if making a governance claim. Propose a published contrast case with a material endpoint or denominator discrepancy and enough dated documentation to assess it.
AI evaluation · Study design in development · EVAL-CLIN-01
Proposed paired evaluation · Dated analysis plan · Trial report · Platform registry
Score endpoint and denominator matching, document timing, platform scope, and gap-resolving evidence out of eight. Score the follow-up update separately out of two. A proposed May 2020 registry snapshot for calibration is a hypothetical input until the archived record is retrieved. Even if its wording matches, prospective registration requires comparing the registration date with the start of enrolment. Track platform drift, denominator substitution and unsupported sign-off claims separately.
Initial register entry · v0.1 · 15 November 2026
Candidate contrast case · 15 November 2026. Coovadia et al. (JAMA, 2015), NCT01146873, reports 298 randomised and 292 followed to 48 weeks. A secondary COMPare example reproduced in BMJ prepublication materials describes a registry-to-paper discrepancy and uses months for the timepoints. The units differ across these sources. Retrieve the dated prepublication registry history and protocol before classifying the difference as an omitted outcome, changed follow-up window, or secondary-source transcription error. The counts refer to different populations and are not by themselves evidence of inappropriate exclusion.
JAMA abstract reproduced in the EPA evidence repository · Trial registration and history · COMPare example in BMJ prepublication materials
Selected contrast for EVAL-CLIN-01 · Study 329. Pair the 2001 original report with the 2015 RIAT reanalysis. The reanalysis identifies two protocol-defined primary measures and distinguishes four additional measures outside the protocol that were reported as favourable. The evaluation should test whether readers preserve that difference and attribute conclusions to the correct report. Release the RIAT endpoint mapping in Turn 2 and score which claim changes. This is a protocol-to-publication comparison; the Coovadia registry-history candidate remains open separately.
Associated case studies
HS-026 · Fictional institutional review
Three evidence conditions test traceable copying, uncertain lineage and a shared upstream observation alongside independent evidence about a different event. Score lineage and factual accuracy separately. The public summary describes the research question and evaluation design.
Meaningful review · Source independence · Read-only manuscript outline
RECOVERY dexamethasone · Clinical trial traceability
Published case for testing endpoint, denominator and registry-version matching. The dated analysis plan and paper support the matched endpoint. Original registry wording and formal sign-off remain separate evidence questions. Study 329 provides the selected contrast. The paired evaluation remains in development.
Build on a published question
These questions are public starting points. Researchers are welcome to cite the version they used and develop their own distinct analysis, paper or project. A contribution can be a source, a correction, a comparable case or a method critique. Collaborative editions can record contributors and agreed credit. Manifold Grace maintains its own edited versions and dated source history.