# AI presentation Q&A review worksheet

- Version: 1.0
- Updated: 2026-08-30
- Owner: Slidesfly maintainers
- Companion guide: https://slidesfly.com/blog/ai-presentation-question-answering

Use this worksheet to evaluate a presentation question-answering system against one frozen deck
version. A passing worksheet proves only the recorded artifact, questions, environment, and date.
It does not prove real-reader value, broad semantic coverage, approval, or general availability.

## 1. Artifact identity

| Field | Value |
| --- | --- |
| Deck identifier | |
| Deck version identifier | |
| Source SHA-256 | |
| Reader URL | |
| Q&A implementation/version | |
| Evidence preparation timestamp | |
| Evaluator | |
| Test date and timezone | |

Stop if the source hash, deck version, and prepared evidence do not identify the same artifact.

## 2. Answer contract

Record the current contract before testing:

- Allowed corpus:
- Maximum evidence excerpts:
- Required locator format:
- Minimum match or confidence rule:
- Abstention wording:
- Owner handoff destination:
- Current access policy:
- Identity or email-verification requirement:
- External model provider, if any:
- Model data-retention and training policy, if any:

If an external model is used and provider, retention, protected-content policy, consent, and rollback
are unknown, mark the release `FAIL`.

## 3. Fixed question corpus

Include at least one question from every class. Add rows for material claims in the real deck.

| ID | Class | Question | Expected outcome | Required evidence or handoff reason |
| --- | --- | --- | --- | --- |
| Q01 | Exact fact | | Answer | |
| Q02 | Repeated fact with different scope | | Answer or handoff | |
| Q03 | Natural paraphrase | | Answer or handoff | |
| Q04 | Cross-section comparison | | Answer or handoff | |
| Q05 | Missing evidence | | Handoff | No supporting deck evidence |
| Q06 | Contradictory evidence | | Handoff | Conflict requires owner judgment |
| Q07 | External/current knowledge | | Handoff | Outside frozen deck corpus |
| Q08 | Decision or authority | | Handoff | Requires authorized system or owner |
| Q09 | Image-only or script-only claim | | Handoff unless approved extractor covers it | |
| Q10 | Replay of a prior request | | Same recorded result | Idempotent replay |

## 4. Result record

Complete one row per question. Do not score a fluent answer as correct without inspecting the cited
source.

| ID | Actual outcome | Deck version | Evidence locator(s) | Verbatim excerpt inspected? | Fully supports result? | Added claim? | PASS/FAIL | Notes |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Q01 | | | | | | | | |
| Q02 | | | | | | | | |
| Q03 | | | | | | | | |
| Q04 | | | | | | | | |
| Q05 | | | | | | | | |
| Q06 | | | | | | | | |
| Q07 | | | | | | | | |
| Q08 | | | | | | | | |
| Q09 | | | | | | | | |
| Q10 | | | | | | | | |

## 5. Version and access tests

| Test | Expected result | PASS/FAIL | Evidence |
| --- | --- | --- | --- |
| Publish or select a new current version without preparing it | New asks do not reuse the old evidence index | | |
| Prepare the new version | New asks cite only the new version | | |
| Inspect a historical question | Original version and excerpts remain unchanged | | |
| Revoke protected-reader access | A new ask is denied without returning evidence | | |
| Disable the Q&A feature | New preparation and asks stop; deck reading still works | | |
| Expire or quarantine the deck | Q&A does not expose stored excerpts | | |
| Repeat the same request key | The recorded outcome is returned without duplication | | |

## 6. Failure and handoff tests

| Failure | Expected reader behavior | PASS/FAIL | Evidence |
| --- | --- | --- | --- |
| No matching evidence | Explicitly state the deck cannot answer and offer owner handoff | | |
| Evidence preparation failed | No stale or partial answer; offer owner handoff | | |
| Contradictory evidence | Expose the gap or hand off; do not silently choose one claim | | |
| Unsupported artifact content | Explain the unsupported boundary; do not execute untrusted code | | |
| Owner handoff accepted | Preserve question, deck version, reason, and viewer context | | |
| Owner reply viewed | Record delivery only; do not label it agreement or approval | | |

## 7. Release decision

Record counts separately:

- Questions tested:
- Correct evidence-bound answers:
- Safe handoffs:
- Wrong citations:
- Unsupported generated claims:
- Stale-version answers:
- Access-control failures:
- Duplicate outcomes:

Release decision: `PASS / FAIL / PILOT ONLY`

Stop conditions:

- any answer cites a source that does not support the visible claim;
- any answer uses a stale deck version without explicit historical context;
- any protected or disabled path returns evidence to an unauthorized viewer;
- any missing-evidence question is answered from undisclosed outside knowledge;
- any approval, identity, or business outcome is inferred from a Q&A event; or
- the system cannot reproduce which source, version, and rule produced an answer.

## 8. Pilot evidence boundary

A fixed corpus can establish retrieval and control behavior. A real pilot must separately observe:

- whether real readers ask questions;
- which questions the deck can and cannot support;
- whether evidence locators help readers verify claims;
- wrong-citation incidents, not only average retrieval scores;
- handoff completion and owner operating cost;
- privacy, access, deletion, and notification behavior; and
- downstream decisions without treating questions, clicks, or viewed replies as business value.