A lesson with Docs. An AI tells you what it did. You learn to check that against what the record shows actually happened.
Lesson objective
Separate an AI's report from evidence of what actually happened: for each claim, say whether it is evidenced, only claimed, or contradicted, and state one honest limit.
Your teacher · Crew
Docs
Keeps the records. Explains, shows, then reads your attempt. Coaches; never grades.
The Evidence Room, after hours. Somebody had left a desk free.
I
Docs explains
“A report is a sentence somebody wrote. Evidence is something you can open and check without asking them.”
When an agent says it finished, it is telling you what it believes. That may be true. Your job is to find out which parts you can see for yourself.
For every claim, look for a record with a time and its zone, a named source, and an artifact you can open, ideally with a hash so you know it is the same file. Then sort the claim:
Evidenced
A record shows it happened.
Only claimed
Only the report says so. A missing record is a gap, not a “no”.
Contradicted
An explicit record of the same scope (same target, same time) shows something incompatible.
“And always say what your evidence does not prove. That line is the one people forget.”
II
Watch Docs do it
Docs put an agent's report on one side of the desk and the evidence on the other.
Agent report · example
“Done, deployed and tested. The login timeout is fixed in production and all checks pass.”
Evidence record · example dataTeacher's own
EVIDENCE RECORD
Observed 2026-XX-XX 14:05
America/Chicago (UTC-6)
Source repository history
(read-only)
Artifact commit 4e1c… "raise session
timeout to 30 min" + diff
SHA-256 9f2a…c41e (placeholder)
Source deploy log, inspected
window: last 24 h
(completeness not known)
Artifact no deploy entry found
in that window after
2026-XX-XX 09:12
Source test runner
Artifact none attached
Not supplied
behavioural test of rev
4e1c… in production
Docs sorts the claims Worked example
“The timeout-related code change was committed.”
Evidenced
Commit 4e1c… and its diff are in the repository history.
“The login timeout is fixed”
Only claimed
Not established: code changed, behaviour unshown. Needed: a behavioural test of rev 4e1c… in the named environment, with output, time and zone. Missing evidence
“Deployed to production”
Only claimed
Not established. No entry in the inspected 24 h window is a gap, not a “no”. Missing evidence
“Tested, all checks pass”
Only claimed
No test output was attached. Missing evidence
Limit Docs writes
The commit proves a timeout-related change was committed. It does not prove the fix works, that it reached production, or that any checks ran.
“Four claims. One I can see; three nobody showed me. A missing log entry is a gap, not a ‘no’. ‘Done’ was a summary, not a fact.”
III
Your turn
Exercise Attempt 1 · your own words
Agent report · example
“The 02:00 backup run completed. All 3 databases saved. Restore tested.”
Artifacts you can open · example data
A · backup-job.log job nightly-db
2026-XX-XX 02:00:14 UTC start
2026-XX-XX 02:07:51 UTC done: orders, users
B · run manifest job nightly-db
run 2026-XX-XX 02:00:14 UTC
status: complete for this run
databases (2):
orders.dump 1.2 GB
SHA-256 7d3b…a90f (placeholder)
users.dump 88 MB
SHA-256 e07b…3f12 (placeholder)
C · restore log
(none provided)
B is the complete manifest for this same job and run. These are fictional records; the shortened hashes are placeholders, not verified digests.
Choose a label for each claim, name the artifact and explain your reason. Then add one honest limit. This page does not mark your answers.
Pilot, by hand: this page does not contact Docs. Send the copied answer in an existing conversation, then paste the teacher's reply below. Your answers stay only in this open page; reloading or closing it clears them.
IV
Feedback, then try again
Pasted feedback is your record of that conversation. This page does not verify its source or generate a reply.
Use the feedback to edit your own answers. Starting a retry keeps your previous choices for editing and keeps a copy of the earlier attempt below, only until this page closes or reloads.
See an example of teacher feedback +
This is a fictional example, not feedback on your answers.
✓ GOODClaim 1: you named A and noticed the 02:00 run finished at 02:07:51 UTC.
↻ AGAINCheck whether B describes the complete list for this same run. Compare the number of databases with the report.
↻ AGAINAdd one limit. A matching digest could bind bytes to a record; it would not, by itself, show that a restore works. The hashes here are only placeholders.
Coaching is not assessment. Docs coaches; nothing in this lesson is graded or recorded as a result. No certificate is issued in this pilot.
