A proposed third queue state earns release only when it identifies a coherent mechanism, changes a useful action and leaves uncertain cases a safe route. The dangerous shortcut is to publish P3 because the old two-child split has an awkward remainder.
This guide develops a prospective admission trial for P3, measures whether the parent fallback actually protects cases the new rule cannot classify, and freezes the boundary before an untouched release check. It treats coverage as a safety property, not a target to maximise by forcing labels.
Examples are hypothetical teaching models. A real service needs its own system boundary, protected outcomes, capacity limits, retention rules and decision authority. A statistical signal opens investigation; it does not authorise a new live route by itself.
eduKateSG · The Gold Standard · Queueing Theory
Find your next learning step
Start with positive evidence for P3, test its proposed action in shadow mode, prove that unresolved cases reach the parent safely, then freeze and challenge the whole control on untouched cases.
- ROUTE 1 · Define admission — turn a remainder into a testable mechanism claim
- ROUTE 2 · Run the trial — compare parent, two-child and three-state controls
- ROUTE 3 · Measure fallback — check coverage, delay and protected outcomes
- ROUTE 4 · Freeze and release — lock the boundary before holdout review
- ROUTE 5 · Transfer carefully — limits, questions and evidence routes
Full chapter index · Queueing Theory root guide · Learning that transfers into life · How X Works hub
Chapter contents
Chapters 1–3
Chapters 4–5
Chapters 6–8
Chapters 9–10
CHAPTER 1 OF 12
1. P3 must be more than the unresolved bin
The preceding guide on re-split shadow validation and third-state escalation identified cases that were unstable when forced into P1 or P2. Admission now asks a stricter question: do some of those cases share evidence available at decision time, a plausible mechanism and an action that differs from the parent?
Write positive inclusion criteria for P3. Also write exclusions from P1 and P2, the evidence that is mandatory, and the cases that must remain unknown. “Anything else” is not a mechanism; it is an invitation for the class to absorb missing data, reviewer uncertainty and future changes.
CHAPTER 2 OF 12
2. State the decision before counting cases
Name the route, priority, capacity reservation or diagnostic action P3 would change. Then name the outcome and harm that will judge that action. If the proposed label never changes a defensible decision, it may still be useful analytically, but it has not earned live operational control.
Record the live parent action as the comparator. The relevant question is not whether P3 cases look unusual after the fact. It is whether information available at arrival supports a reproducible assignment whose action improves a declared result without breaching a protected outcome.
CHAPTER 3 OF 12
3. Design a prospective shadow trial
Freeze a candidate rule and run it beside the live control. For every arrival, retain the fields visible at that moment, the shadow state, proposed action, parent fallback decision, later outcome and reviewer version. Do not relabel the main trial after seeing outcomes.
Use a development window to clarify the rule, then an untouched confirmation window. Cover cycles and channels that matter to the decision. Sample near the boundary deliberately, because high overall agreement can hide instability exactly where P3 competes with P1 or P2.
CHAPTER 4 OF 12
4. Worked example: capacity makes coverage matter
A fictional confirmation window contains 1,200 parent-family arrivals. The frozen rule assigns 360 to P1, 390 to P2, 250 to candidate P3 and 200 to parent fallback. P3 has 92% repeat assignment and its proposed specialist route reduces its 90th-percentile delay from 76 to 44 minutes.
The specialist can accept only 220 cases per window. Thirty low-confidence P3 assignments therefore return to the parent, making effective fallback 230 cases. If those cases wait for specialist rejection before fallback, their protection may be nominal rather than timely. Admission must test both correct destination and transfer delay.
| State | Initial cases | Live treatment | Evidence question |
|---|---|---|---|
| P1 | 360 | Existing child route | Boundary remains stable |
| P2 | 390 | Existing child route | Boundary remains stable |
| Candidate P3 | 250 | 220 specialist; 30 fallback | Action value and capacity |
| Unknown/fallback | 200 | Parent route | Coverage and delay |
CHAPTER 5 OF 12
5. Compare three controls on the same arrivals
Evaluate the parent-only control, the P1/P2 control and the P1/P2/P3 control on the same prospective cases. Compare service tails, reassignment, capacity use, recurrence and protected outcomes. Keep a continue-shadowing result available when evidence is promising but incomplete.
A weighted score can summarise ordinary trade-offs, but it should not cancel a hard safety or access condition. Report counts and uncertainty beside averages. Rare high-consequence cases can disappear inside a favourable mean.
CHAPTER 6 OF 12
6. Define parent-fallback coverage precisely
Fallback coverage is the proportion of cases that fail or abstain from the new classifier and still receive a valid parent decision within the allowed time. The denominator should include missing required fields, low confidence, rule errors, capacity rejection and explicit unknowns—not merely cases voluntarily sent to the parent.
Track at least destination coverage, transfer delay, duplicate handling, lost cases and the protected outcome after fallback. One hundred per cent routing coverage is not enough if handoff delay creates the harm the parent route was supposed to prevent.
CHAPTER 7 OF 12
7. Challenge the fallback rather than admiring it
Run failure injections appropriate to the service: remove one evidence field, exhaust P3 capacity, make the rule service unavailable and present a boundary case with conflicting signals. Confirm that each case reaches the parent once, carries its context and can be recovered without manual guesswork.
Also test recovery. When the specialist route returns, queued cases should not be silently reclassified under a newer rule. The event record must show which version made each decision and whether the parent action was final or provisional.
- Mandatory evidence missing.
- Candidate P3 confidence below the approved limit.
- Specialist capacity exhausted.
- Assignment service unavailable.
- Duplicate or late event arrives after fallback.
CHAPTER 8 OF 12
8. Use sequential monitoring as an investigation trigger
NIST describes CUSUM charts as useful for detecting smaller persistent shifts in a process mean. For an admission trial, a chart might monitor the P3 share, fallback delay or boundary disagreement, but the baseline, reference value and decision interval must be declared before a persuasive sequence appears.
A signal is not proof that P3 failed. Investigate channel mix, missing fields, reviewer learning, policy changes and capacity. Preserve signal history when the rule changes; start a new control version instead of resetting the evidence into apparent stability.
CHAPTER 9 OF 12
9. Freeze the boundary before the release check
Freeze P1, P2, P3 and fallback definitions, required fields, tie-breakers, thresholds, capacity rules and action mappings. Record a version, time and owner. If a material rule changes, close the holdout and begin another confirmation window.
The release set should be untouched by boundary tuning. Review performance by state, channel and distance from the boundary. A model that succeeds only after repeated edits on the same cases has demonstrated fit to those cases, not transfer.
CHAPTER 10 OF 12
10. Release gates and rollback belong together
A release package should pass assignment agreement, action value, parent-fallback coverage, capacity, boundary stability and protected-outcome gates. Set the effective time and a new monitoring baseline while retaining the shadow history.
Define whether a failure rolls back only P3 or the whole split. Drill the rollback to the parent and confirm that historical P3 cases remain queryable. Reversibility means restoring a safe control without erasing why the attempted improvement was made.
| Gate | Pass evidence | Failure response |
|---|---|---|
| Assignment | Reproducible P3 and explicit unknowns | Revise rule; restart holdout |
| Action | Material improvement without protected harm | Keep P3 analytical only |
| Fallback | Complete, timely parent route under challenge | Repair handoff |
| Boundary | Stable on untouched cases and channels | Continue shadowing |
| Rollback | Parent restoration drill passes | Do not release |
CHAPTER 11 OF 12
11. Limits, misconceptions and changed-context transfer
High P3 coverage is not automatically desirable. It can be produced by a boundary so broad that P3 swallows uncertainty. Low fallback use is not proof of safety if failure conditions were absent from the trial. Admission supports only the studied boundary, action and operating range.
Transfer check: a new intake channel lacks the field that separates P3 from P2. Do not impute it merely to preserve coverage. Validate another observable bridge, route those cases to the parent, or create a channel-specific rule with its own evidence and release gates.
Should every unresolved case be sent to P3?
No. P3 requires positive, reproducible evidence and a useful action. Cases with missing or conflicting evidence belong in an explicit unknown state with a safe parent fallback.
What makes parent fallback genuinely protective?
The case reaches a valid parent decision once, within the allowed time, with enough context to act—and challenge tests show that missing data, capacity loss and service failure do not strand it.
CHAPTER 12 OF 12
12. Evidence route and connected reading
NIST SP 800-137 frames continuous monitoring as ongoing visibility into control effectiveness and risk response. The NIST CUSUM guide explains sensitivity to smaller persistent shifts, while the NIST remedial-actions route separates a signal from investigation and action.
Continue locally through post-reactivation monitoring and re-split triggers, then the immediate shadow-validation guide. The Queueing Theory root connects the wider subject.
Contents · Previous chapter · Return to the Queueing Theory root