Grant Evaluation Release-Gate Toolkit
Collection
Fictional grant-evaluation gates. I work on DDScore at Playful Pixels Oy. Funding decisions stay human-owned; evaluator context is linked below. • 4 items • Updated
field stringlengths 9 33 | type stringclasses 4
values | description stringlengths 36 63 |
|---|---|---|
fixture_id | string | Stable fictional fixture identifier. |
change_type | string | Human-readable evaluation-fingerprint change class. |
changed_field | string | Exact fingerprint field changed by the candidate. |
baseline_fingerprint_sha256 | string | Synthetic complete baseline fingerprint hash. |
candidate_fingerprint_sha256 | string | Synthetic complete candidate fingerprint hash. |
calibration_case_basis | string | Confirms the calibration cases are fictional. |
calibration_case_count | integer | Number of fictional cases in the bounded calibration set. |
reviewer_count | integer | Number of blind independent reviewers. |
raw_agreement | number | Recorded raw agreement; not a correctness measure. |
unweighted_cohens_kappa | number | Recorded unweighted kappa; no universal threshold is imposed. |
material_disagreements | integer | Material reviewer differences found. |
unresolved_material_disagreements | integer | Material differences still unresolved at the fixture decision. |
historical_outcome_use | string | Bounded retrospective use; never automatic ground truth. |
expected_comparability | string | Fixture-local expected comparison state. |
expected_release_status | string | Fixture-local expected release state. |
human_release_owner | string | Named fictional accountable release role. |
final_funding_decision_owner | string | Named fictional accountable funding-decision role. |
fixture_local_policy_rule | string | Explicit fictional rule that explains the expected state. |
synthetic | boolean | Always true; no real application, company or person is present. |
Grant-Scoring Release Gate Regression Fixtures
Test a grant-scoring configuration-change release gate without real applications. This CC BY 4.0 package contains 27 fully fictional regression fixtures: three local-policy scenarios for each of nine evaluation-fingerprint fields.
The changed fields are rubric version, policy context, resolved configuration, prompt bundle, tool schema, few-shot set, model identity, evidence cutoff and pipeline commit. Every row keeps the baseline and candidate fingerprint, blind two-reviewer calibration record, agreement statistics, unresolved material disagreements, historical-outcome use, expected comparability, expected release status and named human owners together.
Files
release-gate-regression-fixtures.csv: 27×19 machine-readable fixtures.release-gate-regression-fixtures.jsonl: the same fixtures as JSON Lines.data-dictionary.csv: definitions for all 19 fields.expected-summary.json: counts and coverage invariants.validate-fixtures.cjs: zero-dependency positive and negative contract checks.LICENSE.md: CC BY 4.0 licence notice.
Use
- Run
node validate-fixtures.cjs. - Import either fixture format into the release-gate implementation.
- Confirm that the implementation preserves the named expected state and human owners.
- Replace the fictional policy only after the real programme has documented its own criteria, review roles and change-control rules.
The expected outcomes are examples under each row's explicitly fictional local policy. They are not universal thresholds or recommendations. Agreement is not correctness. Historical awards are not automatic ground truth.
Boundaries
This package is not the DDScore output schema, an application quality score, a grant rubric or an accuracy/fairness benchmark. It does not define programme criteria, validate a rubric, verify application evidence, certify model accuracy or fairness, determine eligibility, rank applicants, award funding, calculate DDScore or replace programme policy, procurement, legal review, a panel or full due diligence. Programme fit and the final funding decision remain with named people.
DDScore context
DDScore is separate analytical tooling. It turns submitted private-company materials into a structured 0–100 analysis across 12 dimensions, checks relevant claims against current public sources and keeps confidence and named gaps visible. It supports the evaluator; it does not replace the evaluator or the programme's policy decision.
Disclosure: I work on DDScore at Playful Pixels Oy.
- Downloads last month
- 50