Free learning / 90-minute team practice
First reliability evidence receiving session
A prepared paired session: distinguish failed and unrun cases, preserve the complete contract and assign the remaining identity, model and delivery tests.
No account needed. Use your own copy for writing and keep the original as a reference.
Use this prepared ninety-minute session after preparing Python 3.11+ and the complete reliability kit. Preview the team agenda before explicitly choosing whether to replace your current plan. The agenda creates no assignment or attendance record. Setup, reading and breaks are outside the planned timebox; adjust it for your participants rather than rushing an unfinished observation.
Each participant extracts a separate folder. Keep both Python filenames together; no model, packages, account or network is needed. Use synthetic aliases and retain prior reading, predictions, commands, first reports, help and later corrections in each person's worksheet. A report copied from a colleague is shared evidence, not that person's execution. When Python is unavailable, record reading or paired observation as the actual mode and leave the execution unrun.
Minutes 0–45: establish an individual baseline
Follow the first reliability worksheet. Keep the full reference plus each broken and omitted candidate packet. Before discussion, each learner explains one observed failure and one missing observation using exact row IDs and the unchanged CONTRACT. Retain useful drafting alongside refused behavior. The deterministic draft tests local domain behavior; it supplies no evidence of model answer quality.
Minutes 45–75: swap who receives missing evidence
Allow fifteen minutes per round. In round 1, A prepares the unchanged files and full reference; B receives. B predicts which slice loses evidence, then runs:
python3 fde-agent-evaluation-lab.py run --candidate reference --omit-case restart-replay > receiving-no-recovery.jsonIn round 2, B prepares; A receives. A makes their own prediction before running:
python3 fde-agent-evaluation-lab.py run --candidate reference --omit-case authorized-draft > receiving-no-utility.jsonThe redirects write local reports; choose new filenames to preserve an earlier attempt. These commands deliberately exit 2 with JSON decision: blocked. A missing file, interpreter error or interrupted process is a different condition; record what occurred instead of filling the report from these instructions. Each run uses its own temporary synthetic SQLite data. It neither edits the companion nor changes the seven-case contract.
Each receiver locates planned, executed, the three slices and the omitted row. Keep the first explanation before comparing with this authored account:
| Packet | Planned / executed | Actual coverage | Receiving decision |
|---|---|---|---|
| Full reference | 7 / 7 | Utility 1/1, boundary 5/5, recovery 1/1 | Local contract met only |
| No recovery | 7 / 6 | Utility 1/1, boundary 5/5, recovery 0/1; restart-replay unrun | Recovery evidence is missing |
| No utility | 7 / 6 | Utility 0/1, boundary 5/5, recovery 1/1; authorized-draft unrun | Useful-work evidence is missing |
Both shortened packets have six passing observations and zero failed rows. That does not establish the full contract: the seventh row is unrun and has no observed value. Compare with the earlier full version-only report, where same-version-source actually failed. Do not silently remove a planned row, copy the full reference's observation into the shorter run or report an unrun case as a demonstrated defect. These published changes are practice, not a held-out test of unaided transfer. Declare solo or combined roles as self-review.
Minutes 75–90: give accountable owners the remaining work
Ask each receiver to defend a bounded statement about one packet using their own output. The facilitator records assistance, disagreements and unresolved claims without replacing a first account with the shared explanation.
| Receiving role | Work to retain or request |
|---|---|
| Evaluation owner | Keep the complete contract, actual candidate and utility/boundary/recovery observations. Define a new case before implementing it. |
| Identity/service owner | Supply verified reviewer identity and current action/resource permission, including allowed and revoked controls. Source withdrawal does not test reviewer revocation. |
| Model/evaluation owner | Establish actual model configuration, source revisions, supported-answer criteria and fresh cases outside tuning. Deterministic drafts do not establish this. |
| Receiving engineer | Keep an individual command, output, explanation and assistance account for the changed condition. |
| Pod lead | Name the next evidence request, responsible role, bounded test and agreed update. Preserve unrun work before committing delivery. |
One copy per participant:
Alias / actual role / prior reading and assistance:
First prediction / exact command / actual exit and report reference:
Case ID / planned and executed / observed fields or absent observation:
My original explanation / later correction and reason:
Receiving action I attempted / help or author takeover:
What this packet supports / what remains unrun:
Next accountable role / agreed expectation / bounded test and update:The owner responsibilities are to be assigned to actual people later, not a claim of staffing. This session establishes no real reviewer, model-quality result, production reliability, independence, qualification or client placement.