# TAVONEL Pilot Evidence Worksheet

- Version: 1
- Locale: en
- Status: BLANK — NOT RUN

## About this worksheet

This is a blank template for preparing a small document pilot and for recording what happens if one is run. As downloaded, it contains no results.

- It is not a benchmark, an evaluation result, a quote, a permission, a license check, or an agreement of any kind.
- Downloading it only retrieves this blank file. Neither downloading nor filling it in starts processing, enables a connection, or creates a payment. Filling it in does not by itself send what you write to TAVONEL.
- Do not write document content, customer names, credentials, API keys, email addresses, production identifiers, or file paths into it. Use logical identifiers that only your team can map back.
- It is not legal or data-protection advice. Where a section asks about rights, retention, or deletion, it records what your own reviewers decide. It does not decide it for them.

## How to fill it in

- An empty field stays `Not provided`. Do not replace it with a guess.
- A result field stays `Not run` until an attempt has actually been made and recorded.
- Never enter 0 as a placeholder. Zero is a measurement: write it only when something was measured and the answer was zero.
- A rate with no attempts behind it is `N/A — no attempts`. Do not write a percentage for it.
- A status in capitals (UNCONFIRMED, UNMEASURED, NOT RUN) changes only after the named person has done the check and recorded the date.

## 1. Task and owner

| Field | Entry |
| --- | --- |
| Task the AI should help with | Not provided |
| Who will use the answers | Not provided |
| Pilot owner (role, not name) | Not provided |
| Independent reviewer (role) | Not provided |
| Worksheet prepared on (date) | Not provided |
| Planned pilot window | Not provided |

## 2. Input rights and processing permission

Status: UNCONFIRMED

Writing an answer here is not a permission, a signature, or a verified license. Processing stays closed until an agreement is made separately.

| Field | Entry |
| --- | --- |
| Who holds the rights to the documents | Not provided |
| Permission decision (granted, refused, or pending) | Not provided |
| Processing that is allowed | Not provided |
| Processing that is not allowed | Not provided |
| Where data may be sent and processed | Not provided |
| Retention period | Not provided |
| Exit and deletion terms | Not provided |
| Who confirms this section (role) | Not provided |
| Date confirmed | Not provided |

## 3. Input manifest

List identifiers only. Do not attach or paste the documents.

| Logical document ID | Source version | Content hash (algorithm and value) | Rights reference (section 2) | Notes |
| --- | --- | --- | --- | --- |
| Not provided | Not provided | Not provided | Not provided | Not provided |

## 4. Questions, expected answers, and evidence anchors

Write each expected answer by hand before any system is run, and point to where in the documents it comes from.

| Question ID | Question | Expected answer (written by a person) | Evidence anchor (document ID and location) | Written by (role) | Independent review status |
| --- | --- | --- | --- | --- | --- |
| Not provided | Not provided | Not provided | Not provided | Not provided | Not reviewed |

## 5. Planned comparison

Status: UNCONFIRMED

| Field | Baseline | Candidate |
| --- | --- | --- |
| Configuration name | Not provided | Not provided |
| System or model | Not provided | Not provided |
| Version | Not provided | Not provided |
| Settings that affect answers | Not provided | Not provided |
| Input snapshot ID (section 3) | Not provided | Not provided |
| Where the configuration came from | Not provided | Not provided |

Comparison conditions. Tick a box only after checking it for both sides.

- [ ] Same questions and the same expected answers
- [ ] Same input snapshot and document versions
- [ ] Same scoring rule, written down before the run
- [ ] Same way of measuring time and cost
- [ ] Any remaining differences are listed below

Remaining differences: Not provided

## 6. Change scenario

Describe one planned change to the documents and what you expect to see. Record behavior that can be checked. Do not describe what a model "pays attention to" or why it answers as it does.

| Field | Entry |
| --- | --- |
| Change to be made (for example, a new version of one document) | Not provided |
| Questions expected to be affected | Not provided |
| Expected change in the answer or its evidence | Not provided |
| How the change will be checked | Not provided |
| Observed outcome | Not run |

## 7. Cost ceiling and stopping conditions

Status: UNCONFIRMED

This worksheet sets no price, credit, provider, or quota.

| Field | Entry |
| --- | --- |
| Planned cost ceiling | Not provided |
| Unit of measurement | Not provided |
| Currency | Not provided |
| Who approves the ceiling (role) | Not provided |
| Approval status | UNCONFIRMED |

Stopping conditions. The pilot stops when any one of these is reached.

- Cost ceiling reached: Not provided
- Time limit reached: Not provided
- A rights or permission question is raised: Not provided
- Other: Not provided

## 8. Results ledger

Status: NOT RUN

Count every attempt. The denominator is all attempts, including failed, refused, and unanswered ones, not only the answers that succeeded. Time and cost spent on failed or refused attempts belong in the totals too. If a configuration has no attempts, each of its rates is `N/A — no attempts`.

### 8a. Outcome counts

| Outcome | Baseline | Candidate |
| --- | --- | --- |
| Attempts (the denominator) | Not run | Not run |
| Correct and supported by the cited evidence | Not run | Not run |
| Correct but not supported by the cited evidence | Not run | Not run |
| Incorrect | Not run | Not run |
| Refused | Not run | Not run |
| Failed (error or timeout) | Not run | Not run |
| Unanswered | Not run | Not run |
| Marked uncertain by the system | Not run | Not run |
| Time per attempt (unit stated) | Not run | Not run |
| Cost (unit and currency stated) | Not run | Not run |
| Support effort (people and time) | Not run | Not run |

Each rate is an outcome divided by attempts. With no attempts: N/A — no attempts.

### 8b. Attempt log

| Attempt ID | Question ID | Configuration | Outcome | Evidence returned | Does the evidence support the answer? | Time | Cost | Notes |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Not run | Not run | Not run | Not run | Not run | Not run | Not run | Not run | Not run |

## 9. Support, missing evidence, and rejection reasons

| Question or attempt ID | What was missing or unsupported | Rejection reason (if the answer was not accepted) | Who decided (role) |
| --- | --- | --- | --- |
| Not run | Not run | Not run | Not run |

## 10. Reuse and purchase intent

Status: UNMEASURED

| Field | Entry |
| --- | --- |
| Would the team reuse these answers? | Not provided |
| Under what conditions | Not provided |
| Purchase intent stated by the decision-maker (role) | Not provided |
| Date asked | Not provided |

## 11. Exit, retention, revocation, deletion, and backups

Status: UNCONFIRMED

This checklist records what must be reviewed and who verifies it. It does not delete, revoke, or promise anything.

| Item to review | Who verifies (role) | How it is verified | Date | Status |
| --- | --- | --- | --- | --- |
| How the pilot ends | Not provided | Not provided | Not provided | UNCONFIRMED |
| Retention period for inputs and outputs | Not provided | Not provided | Not provided | UNCONFIRMED |
| How access is revoked | Not provided | Not provided | Not provided | UNCONFIRMED |
| How inputs and derived data are deleted | Not provided | Not provided | Not provided | UNCONFIRMED |
| Whether backups or other copies exist, and how they are handled | Not provided | Not provided | Not provided | UNCONFIRMED |
| Logs or records that remain after the pilot ends | Not provided | Not provided | Not provided | UNCONFIRMED |

## 12. What this worksheet does not show

- Whether document structure and evidence anchors were preserved is a separate question from whether the answers are correct and useful. Record them separately. One is not evidence for the other.
- Neither of them shows a benefit for your business. That is recorded only in section 10, and only after asking.
- Public or historical datasets, benchmark scores, research results, mirrored data, and patent data are not results for this pilot. Only attempts logged in section 8 count.
- A completed worksheet records one pilot under the conditions in sections 2, 5, and 7. It does not change any product setting, permission, or price.