# Prompt, RAG and fine-tuning evaluation scorecard

Proposed evaluation record. No results supplied. Keep restricted examples in approved storage; reference IDs here.

Workflow/owner: [value]. Costly mistakes: [definition]. Data-use approval: [reference]. Base model/revision/hash: [value]. Runtime/hardware/quantisation/context: [values]. Training/development/final-test split rule: [value]. Locked final-test reference: [ID].

| Candidate | Prompt/tools version | Corpus/embedding version | Adapter/training-data version | Changed component |
| --- | --- | --- | --- | --- |
| Existing baseline | To record | As applicable | As applicable | Baseline |
| Improved prompt/examples/tools | To record | Same evidence where applicable | None | To record |
| RAG | To record | To record | None | Retrieval |
| Fine-tuning/adapter | To record | Same evidence where applicable | To record | Parameters |

## Agree acceptance before execution

Task correctness threshold: [value/rubric]. Citation/support rule: [value]. Refusal/clarification rule: [value]. Latency ceiling: [value]. Review-effort limit: [value]. Minimum useful improvement: [value]. Experiment budget: [value]. Permission violations block expansion.

| Task ID/family | Expected outcome and authorised evidence | Candidate/repetition | Correctness/format/support | Refusal or costly error | Latency/review time | Result reference |
| --- | --- | --- | --- | --- | --- | --- |
| To record | To record before run | Not run | Not measured | Not measured | Not measured | None |

Use development data to select configurations. Keep near-duplicates and related records within one split. Report final-test denominators and variation across repeats; do not repeatedly tune on final-test failures.

## Lifecycle comparison over the same period

Record data preparation, implementation, training/indexing, allocated compute/power, licence/provider charges, maintenance, evaluation, human review, incident and retirement work. State currency, date, tax treatment and each unknown input. Released staff time is not automatically cash saved.

Decision/assumptions: [value]. Stop/rollback owner and known-good configuration: [value]. Remaining uncertainty: [value]. Recheck data rights, network paths and evaluation-provider handling separately.