Test 01 — The input test
Did the model receive the information needed to do the assigned job? Before evaluating an answer, inventory the inputs. Separate source facts from preferences, constraints, and unknowns. A model cannot be expected to preserve distinctions that were never supplied.
Pass condition: another reviewer can identify where the prompt’s factual material came from. If the input fails, do not “prompt harder.” Correct the source packet first.
For a production workflow, version the source packet, retrieval query, model release, system instructions, and sampling configuration. This provenance record makes a later failure diagnosable: the operator can distinguish a source defect from retrieval drift, instruction conflict, or model behavior rather than treating every error as generic “hallucination.”
- List the source materials
- Label each as verified, provisional, or missing
- Remove conflicting instructions
- State what the model must not infer
Test 02 — The transformation test
Is the model transforming information or quietly inventing it? Specify the permitted operation: summarize these notes, compare these options, convert this approved outline into a draft, or identify gaps without filling them.
Pass condition: every factual statement is traceable to supplied material or separately verified evidence. Style can be generated. Evidence must be located.
Evaluate entailment, not citation decoration. A source link beside a sentence is insufficient if the source does not actually support the proposition. Preserve the epistemic status of each statement: extracted fact, model inference, operator judgment, and unresolved uncertainty should not collapse into one fluent paragraph.
Test 03 — The decision-rights test
Did AI cross from assistance into authority? Mark the points at which a person must decide, especially when work involves budgets, creator selection, public claims, partnership terms, publication, or outreach.
Pass condition: each consequential decision has a named human owner and a recorded approval state. Generated and approved should never be treated as synonyms.
Risk-tier the handoff. Reversible internal drafts can tolerate broader experimentation than external messages, contractual language, access changes, or publication. The workflow should encode that difference through permissions and release gates, not rely on a reviewer remembering an unwritten rule at the final moment.
Test 04 — The readback test
Does the final artifact match what was approved? Compare the delivered version with the reviewed version. Check links, names, dates, amounts, permissions, disclosures, and destination. If work was posted or sent, verify the actual destination rather than relying on a draft or queue.
Use deterministic comparisons where possible: identifiers, totals, dates, links, permissions, and content hashes are better checked by exact readback than by another generative pass. Record the material diff between approved and delivered state so operational observability extends beyond a successful API response or queue confirmation.
Recollective promotes practical, human-reviewed AI workflow design. This is its own service description, not a promise of results. The lab conclusion is simpler: a prompt produces an output; a process establishes inputs, boundaries, ownership, and proof.