ggenie-docs

AI Quality Scorecard — build 2026081700

Generated 2026-08-17 from our offline judge, which inspects real research output produced by this build.

Every category answers the same seven questions

An earlier draft of this card graded each category on a different metric. That was wrong, and we would rather say so than quietly change it: if one surface is scored on task success and another on groundedness, the two letters are not comparable, and whoever chooses the metric can choose a flattering one. We now follow the design of Stanford’s HELM — apply one shared rubric to every category and report a cell wherever it applies. 30 of 42 cells are graded on this build; the rest read NA with the reason stated. NA is a published result here, not a gap we hide.

Grades and certainty are separate

We report the observed result as a letter, and how much weight it can bear as a separate certainty rating, following GRADE — the evidence-grading standard used by Cochrane, the WHO and NICE. Certainty starts at High and is downgraded for:

A cell can therefore read A / Very Low: an excellent result you should not yet lean on. That is more useful, and more honest, than folding the uncertainty into the letter and publishing a quietly deflated grade with no explanation of what deflated it.

The rubric

# Dimension Question asked of every category Modelled on
1 Productive work Did it actually produce substantive output, rather than terminating empty? Modelled on agentic task-success rate (tau-bench, Yao et al. 2024; WebArena, Zhou et al. 2023) — the unit is work that achieved its goal, not work that merely finished. We compute a success rate over our own runs; we do not run those benchmarks.
2 Substantiation Can every claim be traced to evidence the system actually holds? Modelled on AIS (Rashkin et al., TACL 2023) and RAGAS faithfulness (Es et al., 2023). NOTE: AIS is a human-evaluation protocol and RAGAS scores claim-level entailment; we do neither. We check structural traceability — whether a claim has evidence attached at all.
3 Reference integrity Do the references resolve, and point at the right record? Modelled on ALCE citation precision (Gao et al., EMNLP 2023). NOTE: ALCE judges whether a citation SUPPORTS its statement, via entailment. We do not do that. We check that a reference is well-formed and specific — a weaker property, and we would rather name the gap than imply the stronger one.
4 Specificity Is the output actionable and particular, rather than boilerplate? Modelled on RAGAS answer relevance (Es et al., 2023). We use pattern-based detectors for boilerplate rather than RAGAS’s model-scored relevance.
5 Instruction adherence Does it honour the machine-checkable output contract it was given? IFEval-style strict accuracy (Zhou et al., 2023) — verifiable instructions scored pass/fail with no partial credit. This is the closest genuine match in the table: our contracts are machine-checkable, so the method really applies.
6 Calibration Does stated confidence match the evidence actually behind the claim? Modelled on the QUESTION HELM’s calibration metric asks — does stated confidence match reality. We do NOT compute expected calibration error: our outputs are discrete states and grades, not probabilities, so there is no calibration curve to score.
7 Non-redundancy Is the same thing emitted more than once? Modelled on the MQM error typology (themqm.org), which scores redundancy by user impact. We apply the idea, not MQM’s weighted scoring formula.

What we borrow, and what we do not implement

The rubric is modelled on published evaluation work; it does not reimplement it. That distinction matters in a document arguing for honest measurement, so here it is explicitly.

Borrowed as structure, used faithfully: HELM’s one-suite-across-every-scenario design; GRADE’s separation of estimate from certainty (three of its five downgrade domains; the sample-size thresholds are ours); IFEval’s strict pass/fail scoring of machine-checkable instructions, which is the closest genuine match here.

Computed exactly: the Wilson score interval behind every certainty rating.

Named but deliberately NOT implemented — each row states the weaker thing we actually do:

If a number here matters to you, the definition behind it is the one in this document — not the one in the cited paper.

Results

Category Productive work Substantiation Reference integrity Specificity Instruction adherence Calibration Non-redundancy Overall
Research runs A A NA NA A NA NA A
Research work log A A A A B A A B
Sources & citations A A A A B NA A B
Person identification A A A A A B A B
Insights A A A A A A A A
Historical places NA NA NA NA NA NA NA NA

Letters are the observed rate: A ≥ 95% · B ≥ 85% · C ≥ 70% · D ≥ 50% · F < 50%. Overall is the worst cell in the row, not an average — a system that substantiates nothing is not redeemed by tidy formatting. Per-cell certainty follows.

Research runs

Records from this build: 5 (capture holds 49).

Dimension Result Grade Certainty Why not higher
Productive work 3/3 = 100.0% A Low serious imprecision: only 3 records in scope
Substantiation 3/3 = 100.0% A Low serious imprecision: only 3 records in scope
Reference integrity NA A run is not itself an authored claim; its references are graded on the sources and insights it wrote.
Specificity NA Not applicable to a run record.
Instruction adherence 3/3 = 100.0% A Low serious imprecision: only 3 records in scope
Calibration NA No records in scope for this dimension on this build.
Non-redundancy NA Run ids are deterministic and rows are reused by design, so a duplicate run cannot occur.

Research work log

Records from this build: 57 (capture holds 270).

Dimension Result Grade Certainty Why not higher
Productive work 19/19 = 100.0% A Moderate imprecision: 19 records in scope, short of the 30 a tight interval needs
Substantiation 19/19 = 100.0% A Moderate imprecision: 19 records in scope, short of the 30 a tight interval needs
Reference integrity 19/19 = 100.0% A Moderate imprecision: 19 records in scope, short of the 30 a tight interval needs
Specificity 19/19 = 100.0% A Moderate imprecision: 19 records in scope, short of the 30 a tight interval needs
Instruction adherence 18/19 = 94.7% B Moderate imprecision: 19 records in scope, short of the 30 a tight interval needs
Calibration 19/19 = 100.0% A Moderate imprecision: 19 records in scope, short of the 30 a tight interval needs
Non-redundancy 19/19 = 100.0% A Moderate imprecision: 19 records in scope, short of the 30 a tight interval needs

Sources & citations

Records from this build: 20 (capture holds 78).

Dimension Result Grade Certainty Why not higher
Productive work 20/20 = 100.0% A Moderate imprecision: 20 records in scope, short of the 30 a tight interval needs
Substantiation 20/20 = 100.0% A Low imprecision: 20 records in scope, short of the 30 a tight interval needs; indirectness: the presence of a source URL is checked, not whether the record at it supports the claim
Reference integrity 20/20 = 100.0% A Low imprecision: 20 records in scope, short of the 30 a tight interval needs; indirectness: link syntax and specificity are checked, but the URLs are never fetched, so reachability is unverified
Specificity 20/20 = 100.0% A Moderate imprecision: 20 records in scope, short of the 30 a tight interval needs
Instruction adherence 18/20 = 90.0% B Moderate imprecision: 20 records in scope, short of the 30 a tight interval needs
Calibration NA A source’s only confidence signal is provisional, which asserts that its URL was found in the research corpus. The capture cannot see corpus membership, so the claim cannot be checked against evidence.
Non-redundancy 20/20 = 100.0% A Moderate imprecision: 20 records in scope, short of the 30 a tight interval needs

Person identification

Records from this build: 19 (capture holds 68).

Dimension Result Grade Certainty Why not higher
Productive work 19/19 = 100.0% A Moderate imprecision: 19 records in scope, short of the 30 a tight interval needs
Substantiation 19/19 = 100.0% A Moderate imprecision: 19 records in scope, short of the 30 a tight interval needs
Reference integrity 19/19 = 100.0% A Moderate imprecision: 19 records in scope, short of the 30 a tight interval needs
Specificity 19/19 = 100.0% A Moderate imprecision: 19 records in scope, short of the 30 a tight interval needs
Instruction adherence 19/19 = 100.0% A Moderate imprecision: 19 records in scope, short of the 30 a tight interval needs
Calibration 17/19 = 89.5% B Moderate imprecision: 19 records in scope, short of the 30 a tight interval needs
Non-redundancy 19/19 = 100.0% A Moderate imprecision: 19 records in scope, short of the 30 a tight interval needs

Insights

Records from this build: 7 (capture holds 25).

Dimension Result Grade Certainty Why not higher
Productive work 7/7 = 100.0% A Low serious imprecision: only 7 records in scope
Substantiation 7/7 = 100.0% A Low serious imprecision: only 7 records in scope
Reference integrity 7/7 = 100.0% A Low serious imprecision: only 7 records in scope
Specificity 7/7 = 100.0% A Low serious imprecision: only 7 records in scope
Instruction adherence 7/7 = 100.0% A Low serious imprecision: only 7 records in scope
Calibration 7/7 = 100.0% A Low serious imprecision: only 7 records in scope
Non-redundancy 7/7 = 100.0% A Low serious imprecision: only 7 records in scope

Historical places

All dimensions NA. Our own measurement is defective, and we would rather say so than publish grades from it. The capture reads 93 records; the database held 20,355 for this account at capture time, the newest written minutes earlier. It joins each place to its originating event and silently drops those whose event our test reset had already deleted, so it has re-reported the same frozen hundred records from February across nineteen builds. Tracked as GGENIE-1136.

Records from this build: 93 (capture holds 93).

Dimension Result Grade Certainty Why not higher
Productive work 93/93 = 100.0% NA Low risk of bias: a known defect in our own capture
Substantiation 93/93 = 100.0% NA Low risk of bias: a known defect in our own capture
Reference integrity 93/93 = 100.0% NA Very Low risk of bias: a known defect in our own capture; indirectness: URL presence is checked; the URLs are never fetched
Specificity 93/93 = 100.0% NA Low risk of bias: a known defect in our own capture
Instruction adherence 93/93 = 100.0% NA Low risk of bias: a known defect in our own capture
Calibration 65/65 = 100.0% NA Low risk of bias: a known defect in our own capture
Non-redundancy 89/93 = 95.7% NA Low risk of bias: a known defect in our own capture