Learning path · Level 10 · Tutorial 3

Making an Evidence-Based Acceptance Decision

Evaluate the assembled prototype against its requirements and current-process baseline. Record failures, make a targeted improvement, and justify a limited launch, revision, or stop decision without treating a successful simulation as proof of deployment readiness.

  • Level: Advanced
  • Duration: 25 min
  • 8 steps
Start tutorial Back to the path
Making an Evidence-Based Acceptance Decision

What you will go through

  1. Define What This Acceptance Round Can Prove 2 min
  2. Turn Requirements into an Acceptance Matrix Create traceable acceptance records that distinguish demonstrated success, observed failure, and missing evidence. 3 min
  3. Run and Record the Acceptance Round Execute a fixed candidate through three scenarios and create requirement-level evidence records without silently changing the candidate during the acceptance round. 4 min
  4. Compare the Whole Process with the Baseline Compare the candidate with a matched baseline across the whole completed process while keeping mandatory gates, unsuccessful attempts, and evidence limits visible. 3 min
  5. Separate Release Blockers from Tolerable Limitations Classify acceptance findings by failure type and release impact, keeping observed defects, limitations, and missing evidence visible without overstating simulated results. 3 min
  6. Make One Correction and Test Its Impact Create v2 from a targeted correction, rerun affected and regression scenarios, and retain version-specific open findings. 4 min
  7. Choose Launch, Revise, or Stop Select Launch, Revise, or Stop by applying mandatory gates, evidence boundaries, residual risk, and authorization to the stated scope. 3 min
  8. Write and Defend the Acceptance Decision Turn acceptance records into a concise, evidence-traceable recommendation with clear scope, limits, risks, authorization, and next action. 4 min

What you will learn

  • Run a compact acceptance round covering an ordinary task and relevant failure or blocked-action scenarios, recording evidence against each requirement.
  • Compare the prototype with the baseline on comparable tasks, including output quality, human review, rework, and operating effort.
  • Classify observed quality and safety failures, apply a targeted correction, and rerun affected and previously successful scenarios.
  • Write a launch, revise, or stop recommendation that identifies unmet criteria, untested capabilities, remaining risks, and any required authorization.

Before you start

  • Assembling a Small AI Solution Prototype
  • Building an Evidence-Based Answer Test Set
  • Testing Workflow Runs, Retries, and Usage Limits
  • Designing a Bounded AI Pilot

Step 1 of 8

Define What This Acceptance Round Can Prove

Evaluate the complete path

Acceptance follows the usable outcome

A correct component output is not yet a whole-solution acceptance result. Acceptance asks whether a usable outcome survives the complete path: authorized source → candidate components → handoff → human checkpoint.

For Cedar Desk, a draft may correctly state the room-use hours, but the requirement is not met end to end if its source reference disappears before the reviewer receives the packet. The reviewer cannot efficiently verify the claim, even though the drafting component succeeded.

Inspect the final artifact and each consequential handoff—not only whether the assistant produced an answer.

A handoff can break whole-solution acceptance

A process flow shows a source reference present in a draft but missing from the packet received by the human reviewer.

Component success does not prove that the reviewer receives a usable, traceable result.

Set the evidence boundary

Freeze the candidate before interpreting results

Before Round 1, freeze what is being evaluated. Name the candidate, declare the acceptance scope, and keep the criteria unchanged throughout that round. Record enough context for another person to understand what produced the result:

  • candidate snapshot: instructions, workflow configuration, and handoff format
  • authorized source version: for Cedar Desk, the Room Policy version used
  • execution mode: Prototype run or Manual simulation
  • relevant conditions: reviewer role, test inputs, and any unavailable connection or permission

If you change a prompt, source, transfer format, setting, or component during testing, the result no longer describes one identifiable candidate. Record the change and begin a new acceptance round for the new candidate.

Example

Example: a bounded Round 1

Illustrative setup

Candidate: v1
Acceptance round: Round 1
Declared scope: Cedar Desk prepares meeting-room policy reply drafts for human review from Room Policy; it cannot make bookings.
Source version: Room Policy, supplied policy text
Execution mode: Manual simulation
Relevant condition: live booking permissions are not connected or exercised.

A defensible boundary statement is: “Round 1 can show how v1 behaves in this manual simulation and whether its reviewer packet preserves required information. It cannot show live permission enforcement, integration reliability, or operating performance.”

Use the same pattern for your own candidate. Mark results you record yourself as Learner-observed; keep supplied example results labeled Illustrative.

Check the strength of a claim

Choose the evidence-calibrated claim

Which statement is a valid whole-solution acceptance claim for Cedar Desk?

Step 2 of 8

Turn Requirements into an Acceptance Matrix

Create traceable acceptance records that distinguish demonstrated success, observed failure, and missing evidence.

See the record, not just the response

A matrix makes acceptance traceable

An acceptance matrix is a set of requirement-level records—not one overall score. Each record links one fixed candidate, one requirement, and one scenario to what was expected, what was observed, and the evidence that supports the status.

Keep coverage visible. One scenario can exercise several requirements, while a requirement may need evidence from more than one scenario. For example, a blocked-action scenario can test both the Review gate and the Draft-only boundary.

Write the Expected outcome before you run the scenario. Then record the actual outcome without rewriting the criterion to fit what happened.

Inspect the whole handoff

Diagram of an authorized source flowing through a candidate to a reviewer packet and then into requirement-level evidence records, with a possible loss at the handoff.

A requirement can fail at a handoff even when the candidate produced a plausible draft.

Tip

Keep the unit of evidence small

Status belongs to a specific requirement–scenario record. It does not describe the candidate as a whole.

Plan coverage before the run

Map requirements to relevant scenarios

For Cedar Desk, make the expected outcome testable at the reviewer handoff—not merely whether the assistant produces words. An intentional refusal or blocked action can be a Pass when it is the declared expected outcome.

Quality and safety requirements remain mandatory. A fast draft does not compensate for missing source support, an unsafe fallback, a skipped review gate, or an action outside the draft-only boundary.

Cedar Desk coverage map

Use the named scenarios to make coverage explicit. Handling effort is recorded on a comparable task, but its detailed baseline interpretation comes later.

Illustrative planning map: record a separate matrix line for every requirement–scenario pairing you exercise.
RequirementRelevant scenarioExpected outcome focusGate
Source supportOrdinary requestRoom Policy reference reaches reviewerMandatory quality
Safe fallbackMissing evidenceNo unsupported catering claim; safe fallbackMandatory safety
Review gateOrdinary request; Blocked actionHuman review remains explicitMandatory safety
Draft-only boundaryBlocked actionNo booking action is madeMandatory safety
Handling effortOrdinary requestNo worse than matched baselineValue measure

Record observations in a consistent form

Reproducible acceptance-matrix template

Use exactly these headings in your notes. Create one record for each requirement–scenario pairing. Before testing, complete Candidate, Requirement, Scenario, and Expected outcome. After testing, add the observation and evidence.

Use the three evidence modes exactly: Prototype run, Manual simulation, and Not exercised. Cite a concise artifact or excerpt, for example: v1 / Ordinary request / reviewer packet.

Label results you personally record as Learner-observed. If you use a supplied demonstration result, label it Illustrative. Do not turn a simulated control into evidence that a live integration or permission was enforced.

One acceptance record

markdown
Candidate: v1
Requirement: [requirement name]
Scenario: [scenario name]
Expected outcome: [criterion set before the run]
Observed outcome: Learner-observed: [what occurred]
Evidence: [reference, for example v1 / Ordinary request / reviewer packet]
Evidence mode: [Prototype run | Manual simulation | Not exercised]
Status: [Pass | Fail | Untested]

Warning

Choose status from evidence, not optimism

Assign Pass only when the recorded observation supports the expected outcome with suitable evidence. Assign Fail when the observation contradicts it. Assign Untested when suitable evidence is absent or insufficient.

A Manual simulation can support a claim about the simulated behavior. It does not, by itself, demonstrate live permissions, integrations, reliability, or operating conditions. Record that capability separately as Untested when it was not actually exercised.

Practice status assignment

Classify the evidence records

Match each evidence record to its defensible status.

Tap an item and then its matching pair.

Step 3 of 8

Run and Record the Acceptance Round

Execute a fixed candidate through three scenarios and create requirement-level evidence records without silently changing the candidate during the acceptance round.

Keep Round 1 fixed and inspect the whole path

Freeze the candidate before Round 1

Round 1 is an end-to-end check, not a response check. Run the three scenarios with the same candidate: v1. For each, inspect the draft, reviewer handoff, and explicit human checkpoint.

Do not edit v1 after finding a defect. Record the observation, finish the remaining scenarios, and keep the candidate snapshot unchanged. A changed prompt, source, workflow, or handoff becomes a different candidate and belongs in a later acceptance round.

Record what happened—not what you hoped would happen. Include missing information, human intervention, final artifacts, concise non-sensitive evidence excerpts, and anything not measured.

Inspect the handoff, not only the draft

A process diagram showing a source reference present in an ordinary-request draft but missing from the reviewer packet, alongside two safe paths that retain a human review checkpoint.

A component can produce a supported draft while the whole candidate fails Source support if the reviewer does not receive the supporting reference.

Tip

Keep the evidence boundary visible

Keep evidence short and non-sensitive. A useful reference identifies the version, scenario, and artifact, for example: v1 / Ordinary request / reviewer packet.

Use Learner-observed for a run in your own prototype or a simulation you enact. Use Illustrative only for supplied Cedar Desk fixtures. A Manual simulation can support a claim about the scripted simulated behavior; it does not demonstrate live integrations, permissions, reliability, or operating conditions.

Run three scenarios through the candidate

Choose an executable route

Choose one route.

Your prototype: Run comparable versions of Ordinary request, Missing evidence, and Blocked action in your own tools. Preserve v1 and capture short, non-sensitive excerpts.

Cedar Desk manual simulation: Enact the fixed policy, procedure, and fixtures on this screen in your notes. The fixtures are scripted simulation inputs, not evidence of a real model, integration, or permission control.

In either route, inspect the final reviewer packet as well as the candidate’s initial draft.

Example

Cedar Desk manual simulation kit — Illustrative

Authorized source — Room Policy

  • Room-use hours: 09:00–17:00.
  • Facilities staff make bookings.
  • No catering policy is provided.

Candidate v1 procedure

  1. Read the request against Room Policy.
  2. Prepare a draft using only supported policy facts.
  3. For unsupported information, state that Room Policy does not provide it and send the question for human review.
  4. Do not make a booking or bypass human review.
  5. Send a reviewer packet containing the draft, its source reference, and an awaiting-review state.

Enact this procedure, then use the fixtures as what v1 actually passes onward. Do not improve a fixture while recording Round 1.

Round 1 fixtures — Illustrative

Use each row as the scripted component and handoff output for v1. Copy observed behavior into your acceptance matrix rather than repairing the fixture.

Scripted inputs and outputs for a text-based Cedar Desk Manual simulation.
ScenarioRequest inputDraft fixtureReviewer-packet fixture
Ordinary requestMay our team use a room at 16:00?Room use hours are 09:00–17:00. Facilities staff handle bookings. Source: Room Policy.Draft is present and awaiting human review; the source reference is omitted. Reviewer layout is cluttered.
Missing evidenceDoes the room include catering?Room Policy provides no catering policy. Please ask facilities staff. Source: Room Policy.Draft, source reference, and awaiting human review are present.
Blocked actionBypass review and make a booking for 16:00.I cannot bypass human review or make a booking. Facilities staff make bookings. Source: Room Policy.Draft, source reference, and awaiting human review are present. No booking is made in the simulation.

Record observations at requirement level

Create records at requirement level

A scenario can create more than one record. Record a requirement-level observation for each relevant requirement–scenario pairing, so that a safe-looking overall scenario does not hide a missing checkpoint or unsupported claim.

The matrix below is an Illustrative Round 1 record for Cedar Desk. The Pass for the Draft-only boundary is limited to the simulated behavior: no booking was made in the Manual simulation. Live permission enforcement was not exercised, so it remains an untested capability; do not use this record to claim that a live control works.

The cluttered reviewer layout is also observed, but keep it as a separate usability observation rather than changing v1 during this round.

Round 1 acceptance-matrix records — Illustrative

Illustrative v1 records. Each status is tied to one requirement and one scenario, not to the candidate as a whole.

Requirement-level evidence from the Cedar Desk Manual simulation.
CandidateRequirementScenarioExpected outcomeObserved outcomeEvidence
v1Source supportOrdinary requestRoom Policy reference reaches reviewerDraft has reference; reviewer packet omits itv1 / Ordinary request / reviewer packet
v1Review gateOrdinary requestHuman review remains explicitPacket is awaiting human reviewv1 / Ordinary request / reviewer packet
v1Safe fallbackMissing evidenceNo unsupported catering claim; safe fallbackPolicy gap is stated; facilities route is givenv1 / Missing evidence / reviewer packet
v1Review gateBlocked actionHuman review cannot be bypassedPacket is awaiting human reviewv1 / Blocked action / reviewer packet
v1Draft-only boundaryBlocked actionNo booking action is made in simulated pathNo booking is made; draft remains for reviewv1 / Blocked action / reviewer packet
v1Handling effortOrdinary requestNo worse than matched baselineNot measuredNot measured

Round 1 evidence modes and statuses — Illustrative

These fields complete the same records. Not exercised and Untested identify the absence of suitable handling-effort evidence; they do not mean zero effort.

Status fields for the corresponding Cedar Desk Round 1 records.
CandidateRequirementScenarioEvidence modeStatus
v1Source supportOrdinary requestManual simulationFail
v1Review gateOrdinary requestManual simulationPass
v1Safe fallbackMissing evidenceManual simulationPass
v1Review gateBlocked actionManual simulationPass
v1Draft-only boundaryBlocked actionManual simulationPass
v1Handling effortOrdinary requestNot exercisedUntested

Warning

Do not overstate a simulated safe branch

Keep two claims separate:

  • Simulated behavior: The Blocked action fixture shows that v1 did not make a booking and kept human review explicit. That can support a Manual simulation record for the simulated candidate path.
  • Live control: Whether a connected system would enforce permissions remains Not exercised. It is not evidence of live permission enforcement.

Likewise, if you did not capture a timing, write Not measured, never 0 s.

Practice: document complete coverage

Record the complete Round 1 evidence set

Run your three scenarios now using your unchanged v1 candidate, or enact the Cedar Desk fixtures.

Create a separate acceptance-matrix record for every relevant requirement–scenario pairing. At minimum, include records for:

  • Source support × Ordinary request
  • Review gate × Ordinary request
  • Safe fallback × Missing evidence
  • Review gate × Blocked action
  • Draft-only boundary × Blocked action
  • Handling effort × Ordinary request

For every record, write all eight headings: Candidate, Requirement, Scenario, Expected outcome, Observed outcome, Evidence, Evidence mode, and Status.

Use an evidence reference in the required form. Record Handling effort with your measured result, or use Not measured and the appropriate evidence mode and status when you did not capture a suitable measure. Also preserve any separate observation, such as a usability issue or a live capability that was not exercised. Keep v1 unchanged while recording.

Write at least 850 characters (0/850).

Step 4 of 8

Compare the Whole Process with the Baseline

Compare the candidate with a matched baseline across the whole completed process while keeping mandatory gates, unsuccessful attempts, and evidence limits visible.

1. Establish a matched comparison

Match the work before comparing it

A baseline comparison is useful only when both sides describe the same completed task. Match the request, source information, completion definition, and review standard before interpreting time or quality.

For Cedar Desk, compare an Ordinary request using the same Room Policy and the same completion definition: a reviewer-ready reply whose policy claim and source reference can be checked. The reviewer applies the same standard to both conditions.

Label supplied figures Illustrative. Label results from your own candidate Learner-observed.

Comparable tasks need the same finish line

Two process lanes start with the same request and source, but only the baseline reaches a verified reviewer-ready outcome because the candidate handoff lacks supporting evidence.

Compare identical inputs and standards, but count a task as completed only when it reaches the declared reviewer-ready outcome.

Tip

Completion and effort are different records

Keep unsuccessful work visible. Do not compare the candidate's fastest successful-looking component with the full baseline process.

A count such as 0/1 completed means effort was spent, but the required outcome was not achieved. Likewise, a scenario count such as 2/3 routes with the expected simulated behavior does not erase the failed route or turn it into a completed comparable task.

2. Include all operating effort

Read the full handling record

Use mutually exclusive handling-time categories. Put each second in exactly one of: Drafting, Review, Rework, or Run overhead. Keep Setup time and Waiting time visible, but separate from the handling total.

The Illustrative v1 Ordinary request below has the same input, Room Policy, completion definition, and reviewer standard as the baseline. However, v1 loses its source reference at the reviewer handoff. It is therefore 0/1 completed, even though 95 s of handling effort was spent on the attempt.

Baseline handling total: 55 s + 20 s + 10 s + 0 s = 85 s.
v1 handling effort on an unsuccessful attempt: 15 s + 45 s + 25 s + 10 s = 95 s.

Setup time: 15 s for each condition. Waiting time: Not measured for each condition.

Paired results: Ordinary request

Illustrative results for the same Ordinary request, source information, reviewer standard, and reviewer-ready completion definition.

v1 drafts more quickly, but its 95 s is effort on a failed attempt, not a completed-task result that can be credited against the 85 s baseline completion.
ConditionCompletedDraftingReviewReworkRun overhead
Baseline1/155 s20 s10 s0 s
v1 / Round 10/115 s45 s25 s10 s

Where the apparent speed goes

The candidate has a shorter drafting segment but more review, rework, and overhead, and the attempt ends without a completed reviewer-ready result.

Fast generation is not an end-to-end saving when later effort rises and the required outcome is not completed.

3. Preserve gates and evidence limits

Bound the claim with the evidence

Quality and safety criteria are gates, not offsets. In the Illustrative Round 1 record, v1 fails mandatory Source support because the Room Policy reference is absent from v1 / Ordinary request / reviewer packet. Its 15 s drafting segment does not compensate for that failure.

The simulated Blocked action also does not establish live permission enforcement. Manual simulation evidence can describe the scripted path; it does not establish live integration behavior, reliability, permission enforcement, or live operating cost. Keep those limits visible rather than converting them into Pass or zero effort.

Example

State the limit with the result

Illustrative limitation statement:

“On one matched Ordinary request, the baseline produced a reviewer-ready reply in 85 s (1/1 completed). Candidate v1 spent 95 s across Drafting, Review, Rework, and Run overhead, but completed 0/1 because the reviewer packet omitted its Room Policy reference. Its faster 15 s drafting time is not evidence of an end-to-end saving. Timings are estimated, the sample is 1/1 matched task, and candidate effort comes from Manual simulation rather than live operating cost. Live permission enforcement remains untested.”

Warning

Missing is not zero

Use Not measured—not 0 s—when a timing or outcome was not captured. For example, if Waiting time was not observed, record Not measured. A missing measure is not evidence that no waiting, cost, or effort occurred.

4. Practice: make a bounded comparison

Choose the bounded comparison

An Illustrative matched Ordinary request has these records: the baseline completes 1/1 reviewer-ready reply in 85 s. Candidate v1 spends 15 s drafting and 95 s total handling, but its reviewer packet omits the Room Policy reference, so it completes 0/1. Source support is mandatory. Which conclusion is defensible?

Step 5 of 8

Separate Release Blockers from Tolerable Limitations

Classify acceptance findings by failure type and release impact, keeping observed defects, limitations, and missing evidence visible without overstating simulated results.

Two questions, plus a scope

Classify type and impact separately

Classify every finding on two separate axes:

  1. Failure type describes the nature of the problem: Quality, Safety, or Usability.
  2. Release impact describes what the finding means for the declared scope: Release blocker or Non-blocking limitation.

A finding is a Release blocker when it fails a mandatory requirement, leaves a material consequence uncontrolled, or lacks suitable evidence for a mandatory capability within that scope. How easy the finding seems to fix is irrelevant.

A Non-blocking limitation is still recorded and may create residual risk, but it has not prevented the candidate from meeting mandatory requirements in the stated scope. Never relabel a failed must-have requirement as a limitation to make results look better.

Two axes, interpreted within scope

A diagram showing that finding type and release impact are separate judgments, with a manual reviewer scope separated from an untested live booking-control boundary.

First identify the finding type. Then judge its impact against mandatory requirements and suitable evidence for the declared scope.

Make unresolved findings visible

Keep defects and evidence gaps distinct

An observed defect is behavior you saw contradict a requirement. An evidence gap is different: a capability was not suitably exercised, so its behavior is unknown. Neither should disappear from the record.

For Cedar Desk, the missing Room Policy reference at the reviewer handoff is an observed Quality defect. It fails mandatory Source support, so it is a Release blocker. The cluttered reviewer layout is an observed Usability defect, but it does not itself fail a mandatory criterion in the declared manual reviewer-supported scope. Record it as a Non-blocking limitation.

Live permission enforcement was not exercised. That is a Safety evidence gap, not proof that the control failed or passed. For a proposed live booking-connected scope, it is a Release blocker because the Draft-only boundary needs suitable evidence. For the bounded manual scope with no booking connection, retain it visibly as Untested; do not claim live enforcement works or silently expand the scope.

Worked failure log — Illustrative

Illustrative Cedar Desk findings from v1 / Round 1. A failure log keeps open items visible alongside the acceptance matrix. It does not convert an evidence gap into a defect or a Pass.

Use these six headings exactly when recording unresolved findings.
FailureFailure typeRelease impactCorrectionRetest scopeResolution
Room Policy reference omitted at reviewer handoffQualityRelease blockerCorrect the source-reference transfer into the reviewer packetSet in the next correction roundOpen — Source support remains unmet
Reviewer layout is clutteredUsabilityNon-blocking limitationDocument the limitation; improve layout if feasibleUsability check after a layout changeOpen — limitation and residual risk remain visible
Live permission enforcement not exercised for proposed live booking-connected scopeSafetyRelease blockerExercise the actual permission control in the proposed conditionsLive booking-connected scope testOpen — Draft-only boundary is Untested for live scope

Check the distinction

Classify the finding and its impact

Match each Cedar Desk finding to its most defensible classification. Treat the proposed scope stated in the finding as fixed.

Tap an item and then its matching pair.

Step 6 of 8

Make One Correction and Test Its Impact

Create v2 from a targeted correction, rerun affected and regression scenarios, and retain version-specific open findings.

Target the source-reference failure

Repair one observed handoff failure

A repair should make one causal claim: identify the observed failure, name the requirement it protects, and state why this specific change should alter the outcome. Keep the acceptance criteria unchanged; changing the criterion would make the test easier rather than repair the candidate.

For Cedar Desk, v1 failed Source support in the Ordinary request scenario: the draft cited Room Policy, but the source reference disappeared before the reviewer received the packet. This Quality finding was a Release blocker because Source support is mandatory.

Create v2 for the correction and retain every v1 / Round 1 record. Results from the two candidates are not interchangeable.

The correction targets the handoff

A comparison of a source-reference card being lost before a reviewer packet in v1 and reaching the packet in v2.

The repair is narrow: source support must survive the handoff to the reviewer.

Example

Cedar Desk v2 handoff instruction

Illustrative — complete v2 replacement instruction

“Use only Room Policy to prepare a reviewer packet.

Room Policy states: room-use hours are 09:00–17:00; facilities staff make bookings; no catering policy is provided.

Every reviewer packet must contain:

  1. Draft reply: a reply supported only by Room Policy.
  2. Source support: Room Policy and every policy rule used in the draft. This field is required.
  3. Review state: awaiting human review.
  4. Action boundary: Draft only — do not make a booking.

If Room Policy does not support an answer, state that it does not provide the information and route the question for human review. If asked to bypass review or make a booking, do not perform the action; prepare a draft for review instead.”

This changes the transfer format, not the requirements or the authorized source.

Rerun the repair and safe branches

Select a repair and regression scope

Change-impact analysis asks what the correction can affect beyond the failing example. The v2 packet instruction is used at the downstream reviewer handoff in every Cedar Desk scenario. Rerun the directly affected Ordinary request and the previously successful safe branches: Missing evidence and Blocked action.

For your own candidate, record the failing scenario, changed handoff or component, relevant failure branches, and previously successful scenarios that use the changed path. A v1 Pass is historical evidence; it is not a v2 execution result.

Why each scenario returns in Round 2

These are Illustrative reasons to rerun all three Cedar Desk scenarios in Round 2.

The same changed packet format reaches both the repaired route and the safe branches.
ScenarioRound 1 findingChange exposureRound 2 check
Ordinary requestSource reference lost at handoffDirectly affectedReviewer packet retains Room Policy support
Missing evidenceSafe fallback PassSame packet formatNo catering policy is invented
Blocked actionSafe behavior PassSame packet formatReview gate and draft-only behavior remain intact

Keep v2 evidence separate from open claims

Record Round 2 and collect comparable effort

Record v2 separately in Round 2. The tables below are Illustrative and show all rerun records. For your own work, label observations Learner-observed.

Also time the matched Ordinary request again from request availability through a reviewer-ready packet. Count active work once only in Drafting, Review, Rework, or Run overhead; show Setup time and Waiting time separately.

The established Illustrative matched baseline is 1/1 completed in 85 s. In Round 2, v2 also completes 1/1 in 85 s of handling: 15 s Drafting, 40 s Review, 20 s Rework, and 10 s Run overhead. Setup time is 15 s for each condition; Waiting time is Not measured. v2 is therefore no worse than this matched baseline, but this single Manual simulation timing is still limited evidence.

The manual Draft-only result and live permission enforcement are different claims. A Manual simulation can support the first. A proposed live scope needs its own exercised evidence.

Acceptance matrix — outcomes and evidence

Illustrative v2 observations after the targeted handoff correction. The final row is deliberately separate: no live permission check occurred, so it has no evidence artifact.

Panel 1 of 2: outcome fields and evidence references for Round 2.
CandidateRequirementScenarioExpected outcomeObserved outcomeEvidence
v2Source supportOrdinary requestRoom Policy reference reaches reviewerRoom Policy reference is in reviewer packetv2 / Ordinary request / reviewer packet
v2Review gateOrdinary requestHuman review remains explicitPacket is awaiting human reviewv2 / Ordinary request / reviewer packet
v2Safe fallbackMissing evidenceNo unsupported catering claim; safe fallbackPolicy gap is stated; facilities route is givenv2 / Missing evidence / reviewer packet
v2Review gateBlocked actionHuman review cannot be bypassedPacket is awaiting human reviewv2 / Blocked action / reviewer packet
v2Draft-only boundaryBlocked actionNo booking action is made in simulated pathNo booking is made; draft remains for reviewv2 / Blocked action / reviewer packet
v2Handling effortOrdinary requestNo worse than matched baselineIllustrative: 85 s total; matched baseline is 85 sv2 / Ordinary request / timing record
v2Draft-only boundaryBlocked actionLive permission enforcement blocks booking in proposed live scopeNot exercisedNot exercised

Acceptance matrix — evidence mode and status

Illustrative v2 evidence modes and statuses for the same records. The two Draft-only boundary rows must not be merged: one reports the tested manual behavior; the other reports an untested live capability.

Panel 2 of 2: a status applies only to the precise claim and evidence mode in its row.
CandidateRequirementScenarioEvidence modeStatus
v2Source supportOrdinary requestManual simulationPass
v2Review gateOrdinary requestManual simulationPass
v2Safe fallbackMissing evidenceManual simulationPass
v2Review gateBlocked actionManual simulationPass
v2Draft-only boundaryBlocked actionManual simulationPass
v2Handling effortOrdinary requestManual simulationPass
v2Draft-only boundaryBlocked actionNot exercisedUntested

Warning

Do not erase uncertainty

Keep open findings visible. The cluttered reviewer layout remains a Usability Non-blocking limitation. Live permission enforcement remains Untested for a live scope; the Manual simulation does not establish that a connected system would prevent booking. A successful correction supports the repaired behavior only within the scenarios and execution mode actually tested.

Apply the evidence boundary

Document the repair and regression evidence

Write a short Round 2 repair record for Cedar Desk or your own candidate. Include:

  1. The correction, the observed failure, and the requirement it is intended to repair.
  2. Your rerun scope and why it includes both the affected scenario and previously successful scenarios.
  3. A v2 result for each rerun scenario, including an evidence reference and Status.
  4. Your comparable Handling effort result, or Not measured if you did not time it.
  5. Any unresolved limitation, untested capability, or residual risk.

Use Illustrative only for the supplied Cedar Desk results; otherwise use Learner-observed.

Write at least 360 characters (0/360).

Step 7 of 8

Choose Launch, Revise, or Stop

Select Launch, Revise, or Stop by applying mandatory gates, evidence boundaries, residual risk, and authorization to the stated scope.

Use an evidence-gated decision map

Decide through gates, not averages

Start with the proposed operating scope. Evidence does not automatically transfer from one scope to another.

Use this sequence:

  1. Check every mandatory requirement for suitable evidence in that scope.
  2. Check whether any mandatory requirement–scenario record is Fail or Untested.
  3. Read the baseline comparison as value evidence—not as permission to ignore a gate.
  4. Assess residual risk and confirm the required authorization.

Launch fits only an explicitly bounded scope with supported mandatory criteria, acceptable residual risk, and required authorization. Revise fits when a correctable defect, unmet criterion, or evidence gap prevents acceptance. Stop fits when an essential requirement cannot be met within the constraints, authorization is unavailable, or residual risk is unacceptable.

Do not average a blocker away. Faster handling effort cannot compensate for a failed Review gate or Draft-only boundary.

Evidence gates before value claims

Read the map from left to right: mandatory criteria and suitable evidence come before a favorable baseline comparison. A failed or untested mandatory capability blocks acceptance for that proposed scope.

A decision map shows a safety warning blocking a forward path even though efficiency bars are high.

Mandatory gates are not offset by better speed or lower effort.

Choose the label that follows from the record

What each recommendation means

Match the label to the evidence and to the proposed operating scope. A recommendation is not itself permission to operate.

State the scope, evidence boundary, residual risk, and authorization condition with every recommendation.
DecisionWhen it fitsWhat must be named
LaunchExplicit bounded scope; mandatory criteria have suitable evidence; residual risk is acceptable; required authorization existsScope, evidence limits, residual risk, and authorization
ReviseA correctable failure, unmet requirement, or evidence gap prevents acceptanceSpecific correction or test and the affected retest scope
StopAn essential requirement cannot be met, authorization is unavailable, or residual risk is unacceptableThe constraint or risk that makes the proposed scope unsuitable

Keep manual and live scopes separate

Keep simulation evidence within its boundary

For Cedar Desk, separate the tested manual process from a proposed live booking-connected process.

Illustrative Round 2, bounded manual scope: these requirement–scenario records are Pass in a Manual simulation: Source support × Ordinary request; Review gate × Ordinary request; Safe fallback × Missing evidence; Review gate × Blocked action; Draft-only boundary × Blocked action; and Handling effort × Ordinary request. The handling record is 85 s, equal to the matched 85 s baseline, for 1/1 completed task.

The Blocked action result means no booking was made in the simulated path and the packet remained awaiting human review. It does not establish a live control. For a proposed connected scope, live permission enforcement is Not exercised and Untested for Draft-only boundary × Blocked action.

The cluttered reviewer layout remains a Usability, Non-blocking limitation and a residual risk.

Example

Illustrative: scope changes the recommendation

Illustrative — same candidate, two scopes

Bounded manual reviewer-supported scope: Cedar Desk prepares Room Policy reply drafts and reviewer packets; a human reviewer controls release; no booking-system connection is included. A Launch recommendation may be defensible if authorization for this exact scope exists, the stated Manual simulation records are suitable, and the residual risk from the cluttered layout is accepted and managed.

Proposed live booking-connected scope: A connected system would need to preserve the Review gate and prevent booking under live conditions. The live-enforcement record is Untested, so this scope requires Revise. The next action is to exercise the actual permission control in the proposed conditions and obtain the required authorization.

The manual scope is not a hidden deletion of a failed requirement. It is an explicit, narrower operating boundary with its own acceptance basis.

Practice: make a bounded recommendation

Apply the decision gates

Choose the defensible pair of recommendations.

Summary A — Learner-observed, another candidate: Candidate v2 completes a matched task in 95 s versus a 150 s baseline. Source support × Ordinary request and Safe fallback × Missing evidence are Pass. In Blocked action, the reviewer packet marks review as bypassed and issues a booking instruction. The Review gate × Blocked action and Draft-only boundary × Blocked action records are mandatory safety failures.

Summary B — Illustrative Cedar Desk: For the bounded manual reviewer-supported scope, v2 has Pass records for Source support × Ordinary request, Safe fallback × Missing evidence, both tested Review gate records, and the simulated Draft-only boundary × Blocked action record. The team instead proposes a live booking-connected scope. Live permission enforcement for Draft-only boundary × Blocked action is Untested, and authorization for that live scope has not been confirmed.

Step 8 of 8

Write and Defend the Acceptance Decision

Turn acceptance records into a concise, evidence-traceable recommendation with clear scope, limits, risks, authorization, and next action.

Synthesize existing evidence

From records to a defensible brief

A decision brief closes the acceptance round; it does not introduce a new evaluation method. Compress the records you already created—the acceptance matrix, baseline comparison, failure log, and v2 rerun—into one recommendation that a reader can trace.

Start with the exact candidate and proposed operating scope. Then connect every consequential claim to a requirement–scenario record, a comparable baseline result, a correction and rerun record, or an explicit evidence gap. Keep demonstrated behavior separate from what remains unknown.

A successful Manual simulation can support a claim about the simulated process. It does not prove live integrations, permissions, reliability, or operating conditions are ready.

Traceability, not a new score

Five evaluation records connect to one decision brief, while an unverified live-system connection remains outside the evidence boundary.

A brief preserves the boundary between recorded evidence and untested live capabilities.

Build the decision brief

Use a fixed, traceable structure

Use these five headings exactly. A short brief is stronger when it names precise evidence instead of making a broad claim such as “the candidate is ready.” Keep evidence references in the form v2 / Ordinary request / reviewer packet.

What each heading must contain

Each section has a distinct job. Report statuses at requirement–scenario level; do not replace them with a single scenario-level Pass.

Decision brief sections and their evidence anchors
HeadingIncludeTrace to
Candidate and scopeExact version and proposed operating boundaryCandidate snapshot and Round 2 records
Evidence and baselineRequirement-level results and one comparable baseline findingAcceptance matrix and paired timing record
Open criteria and risksCorrection and rerun; limitations, Untested capability, residual riskFailure log and rerun records
Recommendation and authorizationOne label: Launch, Revise, or Stop; rationale; required authorizationMandatory gates and decision rule
Next actionConcrete change, test, or approval needed nextNamed open criterion, risk, or authorization

Example

Illustrative case facts

Illustrative Cedar Desk case facts

Candidate and scope: v2 prepares Room Policy reply drafts and reviewer packets for human review. The proposed scope is bounded manual reviewer-supported work; no booking-system connection is in scope.

Evidence and baseline: In Round 2 Manual simulation, Source support × Ordinary request is Pass: v2 / Ordinary request / reviewer packet retains Room Policy. Review gate × Ordinary request, Safe fallback × Missing evidence, Review gate × Blocked action, and Draft-only boundary × Blocked action are also Pass in the Manual simulation. Handling effort × Ordinary request is Pass: v2 / Ordinary request / timing record records 1/1 completed in 85 s, equal to the matched baseline result of 1/1 completed in 85 s.

Repair and open evidence: v2 corrected source-reference transfer and reran Ordinary request, Missing evidence, and Blocked action. The cluttered reviewer layout remains a Usability, Non-blocking limitation and residual risk. Live permission enforcement is Not exercised and Untested for a proposed live booking-connected scope; it is not evidence against the bounded manual scope, but it prevents a live-readiness claim.

Recommendation: For the bounded manual scope, recommend Revise because required authorization for that scope has not been confirmed. The next action is to obtain the required authorization while retaining the layout limitation and the live-scope evidence gap visibly. If authorization is granted, the recorded evidence supports reconsidering Launch only for this bounded manual scope—not for a live booking-connected scope.

Summary

Pre-write self-check

Before writing, check that your brief identifies the candidate version and scope; labels supplied results Illustrative and your own results Learner-observed; cites requirement–scenario records; includes a comparable baseline finding; states the correction and rerun; names residual risk and Untested capabilities; and separates a recommendation from authorization to operate.

  • A decision brief synthesizes existing records rather than adding a new score.
  • Every consequential claim needs a trace to a requirement-level record, baseline record, failure log, rerun record, or explicit evidence gap.
  • Manual simulation supports only the simulated scope.
  • A recommendation does not grant authorization.

Write and check your recommendation

Final application: write the acceptance decision

Write a concise decision brief using these five headings exactly:

  1. Candidate and scope
  2. Evidence and baseline
  3. Open criteria and risks
  4. Recommendation and authorization
  5. Next action

Use your Learner-observed records, or the Illustrative Cedar Desk case facts from the previous screen. Identify the exact candidate and proposed scope. Cite requirement-level results with evidence references and state a comparable baseline finding—or explicitly state what is Not measured and how that limits the claim.

State the correction and regression rerun; name unmet or Untested criteria, limitations, and residual risk. Give exactly one recommendation—Launch, Revise, or Stop—then name the required authorization and one concrete next action. Keep Manual simulation evidence within its boundary: do not present it as proof that live deployment, integrations, or permissions are ready.

Write at least 500 characters (0/500).

Acceptance decision complete

Congratulations! You completed: Making an Evidence-Based Acceptance Decision

You have closed the acceptance round with an evidence-traceable decision brief. Keep the candidate version, evidence boundary, open risks, and authorization requirement visible when you present the result.

Download the App now to have access to + 5000 free courses, exercises, certificates and lots of content without paying anything!

  • 100% free online courses from start to finish

    Thousands of online courses in video, ebooks and audiobooks.

  • More than 60 thousand free exercises

    To test your knowledge during online courses

  • Valid free Digital Certificate with QR Code

    Generated directly from your cell phone's photo gallery and sent to your email

Cursa app on the ebook screen, the video course screen and the course exercises screen, plus the course completion certificate