Monitoring showed that a two-tick memory rule changed meaning when telemetry moved from sixty-second to thirty-second observations. Restoring the old clock repairs the release contract, but it does not answer a deeper design question: can the rule be expressed so that reasonable cadence changes do not silently retune it?
The tenth thaw is therefore narrow. It keeps arrivals, service capacity, eligibility, thresholds, fallbacks and protected outcomes fixed. It compares three memory mechanisms on common event tapes sampled at thirty, sixty and ninety seconds, then tests bursts, missing observations and clock jitter.
The Gold Standard is not a rule that produces identical decisions under every sample schedule. It is a rule whose intended elapsed-time evidence, service consequences and limits remain explainable when observation opportunities change.
A practical orientation
Use this guide when monitoring has identified sampling cadence as a real policy input. The worked replay follows 8,100 eligible decisions across nine service days and three observation schedules.
Choose a reading route
Bound the tenth thaw
Chapters 1–3 freeze the queue contract and define cadence invariance precisely.
Build faithful replays
Chapters 4–6 align event time, observations and memory-state measures.
Work and transfer the comparison
Chapters 7–9 compare contenders and change burst, gap and jitter contexts.
Write the tenth re-freeze
Chapters 10–11 define the release gate, limits and evidence package.
Continue the verified route
Chapter 12 answers practical questions and links the verified sequence.
Chapter contents
- 1. Freeze the queueing contract before changing memory
- 2. Define cadence invariance as a tolerance, not literal identity
- 3. Specify contenders with different memory semantics
- 4. Construct one continuous event tape and several observation views
- 5. Measure effective memory rather than rule labels
- 6. Pre-declare replay validity and decision rules
- 7. Worked replay: elapsed time transfers better than tick count
- 8. Transfer check: sharpen bursts and move the threshold scale
- 9. Stress missing observations, jitter and restart state
- 10. Write the tenth re-freeze as a versioned evidence contract
- 11. Limits, misconceptions and practical evidence checks
- 12. FAQs and the next reading route
1. Freeze the queueing contract before changing memory
The ninth re-freeze remains the reference: a 120-slot cap, fixed cohort definitions, the same service calendar, deterministic tie-breaks, protected override and fallback path. The tenth thaw may change only the way evidence accumulates before a threshold transfer.
Every contender receives the same decision-time classifier values and valid-observation markers. It cannot infer missing values from later outcomes, increase capacity or use a reconciled label unavailable at the decision.
Write the purpose before the mechanism: resist brief threshold crossings, respond to sustained pressure within a declared time and avoid repeated borrower-displacement reversals.
- Reference capacity and service frozen.
- One permitted change: evidence memory.
- Decision-time fields only.
- Protected outcomes and rollback retained.
2. Define cadence invariance as a tolerance, not literal identity
Two observation schedules rarely create identical paths because one may see a brief excursion that another misses. The useful invariant is therefore outcome stability under a declared family of schedules, not equality of every tick.
Pre-declare three tolerances: median transfer time changes by no more than fifteen seconds, qualifying reversals by no more than one per 100 eligible decisions, and no cohort fallback changes by more than 0.5 percentage points relative to the sixty-second replay.
A contender fails when cadence alone crosses those limits while the underlying continuous event path and service process remain fixed.
| Invariant family | Tolerance | Why it matters |
|---|---|---|
| Transfer timing | ±15 seconds median | Preserve response speed |
| Reversals | ±1 per 100 | Preserve stability |
| Cohort fallback | ±0.5 points | Protect displacement |
| Deadline success | Non-inferior | Protect service |
3. Specify contenders with different memory semantics
T60 requires the classifier to remain beyond the boundary for sixty elapsed seconds and at least two valid observations. A90 integrates positive threshold distance over a ninety-second window and transfers when the signed area exceeds a declared amount. S3 requires three confirmations but weights intervals so widely separated observations do not masquerade as continuous evidence.
The existing two-tick rule is the diagnostic reference because its meaning changes directly with cadence. An oracle built from the continuous simulated path estimates which crossings truly persist; it cannot be deployed.
Each contender states gap handling, restart state, clock source and override behaviour. Combining mechanisms without attribution would make a successful replay difficult to explain.
| Rule | Memory object | Primary risk |
|---|---|---|
| Two-tick reference | Observation count | Cadence retunes time |
| T60 | Elapsed duration plus two observations | Sparse sampling delays response |
| A90 | Threshold-distance integral | Sensitive to interpolation |
| S3 | Interval-weighted confirmations | More state complexity |
| Oracle | Continuous future-labelled path | Not deployable |
4. Construct one continuous event tape and several observation views
The base tape contains arrivals, service completions, capacity changes and a time-stamped classifier path at its highest trustworthy resolution. Observation schedules are views of that path, not different demand histories.
Generate thirty-, sixty- and ninety-second views using declared alignment and jitter. Preserve real gaps rather than filling them. Recompute policy-generated transfers independently for every contender and view.
The sixty-second reference replay must reproduce the released implementation before comparisons are interpreted. If it does not, repair the tape or state model first.
- One exogenous event tape.
- Declared observation alignment.
- Real gaps preserved.
- Policy state recomputed per run.
- Released replay reproduced exactly.
5. Measure effective memory rather than rule labels
For every transfer, record elapsed time from first qualifying evidence, number of valid observations, integrated threshold distance, reset count and whether an override fired. This reveals what the policy actually remembered.
Join those state measures to borrower completion, displaced-case delay, fallback and protected deadlines. A rule can appear cadence-stable in transfer count while changing which cohort bears the delay.
Report the full distribution and cohort slices. A median alone can hide a long tail created by ninety-second observations.
| State measure | Interpretation | Outcome join |
|---|---|---|
| Elapsed confirmation | Duration remembered | Transfer delay |
| Valid observations | Evidence opportunities | Gap exposure |
| Threshold area | Pressure strength | Useful borrowing |
| Reset count | Continuity breaks | Fallback and reversals |
6. Pre-declare replay validity and decision rules
A valid replay preserves event order, uses one monotonic clock, applies the released tie-break and produces no impossible negative intervals. Counterfactual service assumptions are shared across contenders and tested with common random numbers where simulation is necessary.
A contender must pass service non-inferiority, every cadence envelope, the missing-observation scenario and an independent implementation review. No weighted score can trade a protected failure for fewer reversals.
When a cadence view cannot observe enough evidence to decide, report indeterminate policy state or delayed action according to the contract. Do not interpolate merely to improve the result.
| Gate | Pass condition | Failure meaning |
|---|---|---|
| Reference reproduction | Exact state path | Replay invalid |
| Service | Protected outcomes non-inferior | Contender unsafe |
| Cadence envelope | All schedules within limits | Memory not invariant |
| Implementation review | Independent match | Specification ambiguous |
7. Worked replay: elapsed time transfers better than tick count
Across 8,100 eligible decisions, the two-tick reference confirms in 31, 61 and 91 median seconds under thirty-, sixty- and ninety-second views. Reversals range from 12.0 to 7.1 per 100, and cohort D fallback spans 5.0% to 3.7%. It fails every cadence envelope.
T60 confirms in 62, 63 and 94 median seconds. The ninety-second view is too sparse to meet the timing tolerance, although reversals and fallback remain stable. A90 confirms in 68, 71 and 77 seconds, with reversals between 7.6 and 8.2 and cohort fallback within 0.3 points. S3 is stable but delays protected transfers during sparse observation.
A90 is the provisional winner. Its result depends on the declared threshold-path interpolation, so it must pass changed-context checks before re-freeze.
| Rule | Median time across cadences | Reversal range | Decision |
|---|---|---|---|
| Two tick | 31–91 s | 7.1–12.0 | Fails |
| T60 | 62–94 s | 7.4–8.0 | Fails timing at 90 s |
| A90 | 68–77 s | 7.6–8.2 | Provisional |
| S3 | 79–132 s | 6.9–7.5 | Fails protected timing |
8. Transfer check: sharpen bursts and move the threshold scale
Compress arrivals into eighteen-minute bursts while preserving total volume and service. A90 accumulates evidence faster during genuine pressure, but the protected override rate remains below the three-percent review limit.
Rescale the classifier by a documented calibration change while preserving its decision meaning. The area threshold must transform with the scale. A hard-coded numeric area that is not versioned fails even if the queue outcomes look acceptable.
Move more cases close to the boundary. A90 becomes more sensitive to small interpolation choices, so the uncertainty band and alternative interpolation replay are reported rather than hidden.
| Changed context | Expected property | Risk exposed |
|---|---|---|
| Sharper bursts | Timely response to sustained pressure | Over-fast accumulation |
| Classifier rescale | Equivalent physical meaning | Unversioned numeric threshold |
| Boundary-heavy mix | Stable uncertainty envelope | Interpolation dependence |
9. Stress missing observations, jitter and restart state
Remove one observation inside an otherwise persistent crossing. A90 integrates only over supported intervals and does not bridge a gap longer than its declared maximum. T60 pauses or resets according to the same continuity rule.
Add random timing jitter around each nominal schedule. The replay uses actual timestamps, so elapsed-time and area rules change smoothly rather than at arbitrary tick counts.
Restart the allocator with and without a durable state snapshot. The snapshot path must match uninterrupted execution; the cold path enters a safe declared state. No contender reconstructs memory from future values.
- Maximum bridgeable gap declared.
- Actual timestamps used.
- Jitter distribution recorded.
- Durable restart equivalence tested.
- Cold-start state conservative and visible.
10. Write the tenth re-freeze as a versioned evidence contract
The re-freeze names A90’s window, threshold-distance units, interpolation rule, gap limit, clock source, reset state, protected override and supported cadence family. The two-tick ninth rule remains the rollback reference.
Release requires exact replay reproduction, cadence and transfer gates, shadow execution on live timestamps and independent calculation of case-level outcomes. Stop if the unsupported-gap share, override dependence or cohort displacement exceeds its limit.
The contract also states what is not promised: cadence outside thirty to ninety seconds, a redefined classifier or a different capacity regime requires new evidence.
- Versioned units and interpolation.
- Supported cadence range.
- Rollback implementation.
- Shadow and stop gates.
- Out-of-scope conditions explicit.
11. Limits, misconceptions and practical evidence checks
Cadence invariance does not make sampling irrelevant. A ninety-second observer can miss a forty-second event entirely. An integrated rule relies on the trustworthy resolution and interpolation of the base path; it is not direct access to continuous truth.
The replay cannot recover unlogged triage, abandonment or service changes. A stable policy can still produce different outcomes when demand changes. Statistical process monitoring identifies changes for investigation; it does not supply the mechanism by itself.
Preserve the base tape, observation-view generator, hashes, state-machine code, units, seeds, scenario manifest, case-level outcomes, exclusions and signed decision. Another reviewer should reproduce every table.
- No claim of literal path identity.
- Sampling blind spots acknowledged.
- Interpolation sensitivity reported.
- Unlogged work and abandonment stated.
- Independent rerun completed.
12. FAQs and the next reading route
A cadence-invariant design keeps the intended evidence object visible. It makes elapsed duration, accumulated strength and missing observations explicit instead of letting the tick counter define policy by accident.
Read the Queueing Theory root, the ninth re-freeze monitoring guide and this tenth-thaw comparison in order. The sequence moves from detecting memory drift to comparing bounded repairs.
The Gold Standard parent and How X Works hub connect this to a wider learning method: preserve the question, change one representation of evidence and test whether the conclusion transfers.
Why is A90 not automatically the permanent winner?
Its result depends on the supported cadence range, classifier scale and interpolation assumptions. New conditions require new transfer evidence.
Why require two observations in T60?
Elapsed time alone could be satisfied across an unobserved gap. The observation requirement protects evidence continuity.
Can a faster sensor always improve the queue?
No. More observations can make a count-based rule react faster than intended and increase reversals or displacement.
Continue through the verified ecosystem
Sources and scope note
- NIST/SEMATECH e-Handbook — Process control techniques — used for the distinction between establishing a reference process and monitoring later observations against it.
- NIST/SEMATECH e-Handbook — EWMA control charts — used for explicit memory weighting, representative historical data and gradual-drift monitoring.
All queue events, policies and outcomes are original teaching examples. The sources support monitoring concepts; they do not validate the example’s thresholds, service assumptions or release choice.