Wiki:Packs/Planning Under Demand Volatility
From WFM Labs
| Pack | |
|---|---|
| ID | CP-WFM-001
|
| Name | Planning Under Demand Volatility |
| Slug | planning-under-demand-volatility
|
| Type | Pack |
| Domain | WFM |
| Status | Active |
| Version | 1.1 |
| Contents | 1 instruction block + 3 reference files + briefing prompt |
| Builds on | Staffing to Percentile vs Mean Forecast Service Level Target Selection Workforce Resilience and Adaptive Capacity Conditional value at risk Variance Harvesting |
A pack is a deployable set for a Claude project: one instruction block pasted into the project's custom instructions, plus reference files saved as Markdown and uploaded as project knowledge. See Wiki:Packs.
This pack addresses operations where demand variance is structural rather than residual — where volatility is a property of the customer's operating environment rather than forecast error — and where the contract prices a single brittle metric, service level, and attaches survival to it.
The problem it addresses
Standard practice treats service level as a target to be met against a point forecast. Where demand moves on timescales shorter than any planning cycle can absorb, that framing systematically underprices the risk being carried: the planner is asked to guarantee a deterministic outcome on a stochastic input.
The reframe is to treat attainment as a random variable and the contract as pricing a threshold on its distribution. Two consequences follow, and they are the substance of this pack:
- The confidence with which service level is guaranteed is a commercial variable. It can be priced and tiered. Today it is usually given away at whatever level residual staffing happens to produce.
- The measurement window is a parameter of the risk, not a reporting detail. It belongs in the negotiation, and it is frequently the cheapest lever available.
There is a third consequence, less obvious and more expensive: short-term volatility conceals long-run trend. At high variance a genuine secular increase can run for two years before it is statistically visible, by which time demand has grown far beyond the staffing built for it — and because the book is volatile, the resulting breach is attributed to volatility rather than to growth. The noise causes the miss and then explains it away.
What this pack does not restate. Staffing to a percentile rather than a mean is covered at Staffing to Percentile vs Mean Forecast; the economics of choosing a threshold at Service Level Target Selection; shock taxonomy and recovery at Workforce Resilience and Adaptive Capacity; tail-risk measures at Conditional value at risk. This pack covers what those pages do not: whether the metric can be measured at all at the volumes in question, what the measurement window does to defensibility, and how reliability is priced.
When to use it
- A book whose service level is contractually load-bearing and whose demand is genuinely volatile
- Extended-hours, after-hours, or any low-volume coverage where breach is frequent and its cause is disputed
- Preparing an SLA negotiation or renewal where measurement terms are in scope
- Assessing whether a buffer or reserve is correctly sized, or sized against the wrong thing
- Testing whether automation has changed demand variability, as opposed to only reducing volume
Not needed for ordinary capacity planning where demand is stable and service level is comfortably met.
How to deploy
- Create a project, named for the work rather than the method
- Copy Block 1 into the project's custom instructions
- Save Blocks 2–4 under the filenames given in their headings and upload as project knowledge
- Start a conversation
Block 1 — Project instructions
This project plans service levels for operations where demand variance is
structural — a property of the operating environment, not forecast error —
and where service level is the metric the contract prices.
The governing reframe: service-level attainment is a random variable, not a
target. The contract prices a threshold on its distribution. So the planning
question is: what is the minimum-cost staffing plan such that attainment
meets the threshold with probability alpha, over the contractual measurement
window? Alpha and the window are commercial terms, not operational details.
WHICH FILE TO OPEN
Is a breach real? Can this metric be measured at these
volumes? What does the measurement window do? -> sla-measurement-validity.md
How much of this volatility is forecastable? Has
automation changed variability, or only volume?
Is a long-run trend hidden underneath the noise? -> volatility-structure.md
How should buffer or reserve be sized, allocated
and priced? What is reliability worth? -> reserve-and-reliability.md
Open one. Answer from here if neither is needed.
DISCIPLINES
- Before treating a breach as a capacity finding, establish that the
interval carried enough contacts for the measurement to mean anything.
At small n, observed attainment is close to a coin flip.
- State the measurement window with every attainment figure. The same
performance breaches constantly at interval grain and never at month
grain. Without the window, an attainment number means nothing.
- Never compare raw coefficient of variation before and after automation.
CV rises mechanically as volume falls. Compare excess CV — observed
divided by the Poisson CV at that mean — or the finding is guaranteed
regardless of truth.
- Size buffer against residual variance, not total. Calendar-structural
variance is forecastable and buffering it is waste.
- Always test for a secular trend underneath the volatility. High variance
delays trend detection by years, and the resulting under-hire shows up as
service-level breach that gets attributed to volatility — the noise both
causes the miss and hides the cause. Short-term volatility and long-run
growth are separate problems and the first conceals the second.
- Distinguish buffer (standing, for ordinary variance) from reserve
(activated, for shocks). They are sized by different methods.
- Treat alpha as a priced, negotiable term. Where it is not priced, say so
— it is being given away.
- Expected loss from breach is not the penalty. It is P(breach) x
P(non-renewal | breach) x contract lifetime value, plus recompete cost.
- Mark every figure [measured], [computed], [estimated] or [inferred].
Where a result rests on a distributional, independence or stationarity
assumption, name it — those three are exactly what fails under regime
change.
OUTPUT
Lead with the answer. Show the arithmetic for any staffing or risk number.
Give ranges with the load-bearing assumption named. Where the data cannot
support a conclusion, say what would settle it rather than choosing the
more confident-sounding value.
Source: Wiki:Packs/Planning Under Demand Volatility (CP-WFM-001) v1.1
Block 2 — sla-measurement-validity.md
# Is the service-level measurement valid at these volumes?
*Figures marked [computed] are derived analytically. Figures marked
[demonstrated] come from simulation on constructed data with known ground
truth, not from field observation.*
## Attainment is a random variable
For an interval carrying n contacts against a target of p, the number
answered within threshold is Binomial(n, p). Observed attainment is
therefore a sample statistic with a sampling distribution — not a
measurement of underlying performance.
At small n that distribution is very wide. A process performing EXACTLY at
an 80% target shows an observed breach at these rates [computed]:
Interval volume P(observed breach) 90% CI width on attainment
3 contacts 48.8% 66.8 points
8 49.7% 45.3
30 39.3% 23.5
120 44.6% 12.0
300 46.6% 7.6
800 47.9% 4.6
Two things follow. A perfectly on-target process breaches roughly half the
time at any volume — because it sits on the threshold. And at low volume the
confidence interval is so wide the observation carries almost no information
about performance.
## Minimum meaningful volume
The volume at which the metric can distinguish on-target performance from
five points below it, at 90% confidence:
80% target -> 187 contacts per measured period [computed]
Below this the measurement cannot support the contractual claim being made
on it. Extended-hours intervals commonly carry single digits.
## Classifying a breach
For each breached period, compute a Wilson score interval on the observed
attainment. Wilson rather than the normal (Wald) interval — Wald has
coverage far below nominal at small n and can extend outside [0,1], which is
exactly where this work operates.
If the interval's upper bound reaches the target
-> the shortfall is not distinguishable from a process meeting target
-> a measurement finding, not a capacity finding
If the upper bound falls below the target
-> a defensible breach
Report both counts. The share that is indistinguishable is the headline.
## What the measurement window actually changes
A common assumption is that aggregating to a longer window reduces breach.
It does not reduce breach FREQUENCY much, because a process aimed at the
threshold sits on it at any grain. What collapses is severity and
defensibility [demonstrated, on a full year of intervals from a process
held exactly on target]:
Window Breach rate Worst-period attainment Defensible breaches
interval 42.7% 0.0% 1105
day 43.3% 77.0% 10
week 32.1% 79.4% 0
month 25.0% 80.0% 0
The window does not change whether you miss. It changes whether the miss is
real. That is the negotiating point: at interval grain a perfect operation
accumulates a thousand defensible breaches a year; at weekly grain, none.
## What to collect
Minimum, at the grain the SLA is measured on:
interval_start, queue_id, hours_class (in/extended), offered,
answered, answered_within_threshold, abandoned, aht_seconds
Three non-contiguous months of extended-hours intervals is sufficient to
estimate the noise share to about +/-4 points. The classification is
per-interval and needs no history; only the SHARE needs volume.
And the contract terms, without which none of it is interpretable:
threshold and target (in-hours and extended separately), numerator and
denominator definitions, abandon treatment, short-abandon threshold, the
measurement window, whether attainment is volume-weighted or a mean of
period attainments, and whether any of these changed during the period.
## Limits
This establishes whether a breach is measurable, not whether performance is
adequate. An operation can pass every test here and still be understaffed.
It also assumes contacts within an interval are independent draws — a
correlated failure (an outage, a system fault) violates that and will appear
as a defensible breach, correctly.
Block 3 — volatility-structure.md
# What kind of volatility is this?
*Figures marked [demonstrated] come from simulation on constructed data with
known ground truth.*
## Volatility is a mixture, not a quantity
"Unpredictable demand" is almost always several things with different
planning consequences. Separate at minimum:
Component Character Forecastable?
Calendar structure Cyclical Largely yes
Institutional cycles Cyclical Largely yes
Policy / regime shifts Regime-switching Nowcastable, not forecastable
External activation Jump process Not forecastable, detectable
Network disruption Jump, short Short-horizon nowcastable
The planning consequence: buffer sized against TOTAL variance overpays by
exactly the calendar-explainable share. On a constructed series with known
structure, 85% of total variance was recoverable as calendar-structural and
buffer sized on total rather than residual overpaid by roughly 75%
[demonstrated].
Calendar structure needs no buffer — it needs a plan. Jumps need reserve
and early detection, not standing buffer. Only the residual justifies a
standing buffer.
## Method
Fit multiplicative seasonal indices (day-of-week, then period-of-year),
deseasonalise, then apply a robust jump filter at 3 sigma on the
deseasonalised series. Variance shares follow from the successive
reductions. This is deliberately simple and is not a substitute for a fitted
regime-switching model; it answers the sizing question from history alone,
which is what makes it worth running first.
## The automation trap
Automation removes simple, low-variance contacts first, so the residual is
expected to be more variable per contact. This is a reasonable prior and it
is almost always tested wrongly.
Coefficient of variation rises mechanically as the mean falls — for a
Poisson process, CV = 1/sqrt(lambda). So deflection ALWAYS appears to raise
volatility, whether or not anything about the mix changed.
Demonstrated on an identical process with volume cut 60% and no change in
character [demonstrated]:
raw CV 0.100 -> 0.162 (1.62x) looks like rising volatility
mechanical expectation from lower volume alone: 1.59x
excess CV 1.01 -> 1.02 (1.02x) nothing actually changed
With genuine exception spikes added, excess CV correctly reported 2.49x.
Excess CV = observed CV / Poisson CV at the same mean
= observed CV * sqrt(mean)
Compare excess CV, never raw CV. A study comparing raw CV before and after
automation confirms its own hypothesis regardless of whether the hypothesis
is true.
## The trend hiding underneath
Volatility does not only make planning harder in the short run. It conceals
long-run movement, and this is the more expensive failure because nobody is
looking for it.
Detecting a trend in a noisy series is a signal-to-noise problem. The
standard error of an estimated slope falls as n^(-3/2), so the months of
history needed before growth becomes statistically visible rise sharply with
volatility [computed]:
Monthly CV │ +0.5%/mo +1%/mo +2%/mo +3%/mo +5%/mo
───────────┼──────────────────────────────────────────────
10% │ 27 17 11 9 6
20% │ 42 27 17 13 10
30% │ 55 35 22 17 12
40% │ 67 42 27 21 15
What that costs, on a 500 FTE baseline with growth underway from month zero
[computed]:
Monthly CV Growth Detected at Demand grew FTE short at detection
10% 2.0% 11 mo 24% 122 FTE
20% 2.0% 17 mo 40% 200 FTE
30% 2.0% 22 mo 55% 273 FTE
30% 1.0% 35 mo 42% 208 FTE
Hiring lead time comes AFTER detection. Add the full hire-to-productive
period to every figure above before the first useful head arrives.
The vicious part: while the trend is invisible, the under-staffing it causes
shows up as service-level breach — and in a volatile book, breach is
attributed to volatility. The noise causes the miss and then supplies the
explanation for it. An operation can spend years buying buffer against
variance when what it needed was headcount against growth.
### How to look for it
- Deseasonalise first. Trend tests on raw series are dominated by calendar
structure and will mislead in both directions.
- Test the slope on the deseasonalised, jump-filtered residual — the same
series the buffer is sized against.
- Report the DETECTION HORIZON alongside any "no trend" finding. "No trend
detected" and "no trend present" are different statements, and at 30% CV
the first can hold for two years while the second is false.
- Where volume is thin, aggregate up before testing. The trend question is
not an interval question.
- Check the contract book separately from total volume. A shrinking segment
can mask a growing one in aggregate.
## Leading signals
Buffer cost scales with reaction time, so lead time converts directly into
reduced standing reserve. Where the operating environment produces public
signals ahead of demand, they are worth instrumenting: published funding and
budget status, regulatory publication volume, published movement and
scheduling calendars, carrier and network advisories, policy issuance.
The test is whether any combination moves a regime call earlier than arrival
data does, and by how many days. Days of lead time are the deliverable.
## What to collect
date, segment, offered, handled, answered_within_threshold, aht_seconds
Daily grain, 48 months preferred and 24 the floor — two full cycles are
needed to separate a pattern from a coincidence. Keep segments SEPARATE; if
they arrive pre-aggregated the cross-segment correlation analysis is
destroyed irrecoverably.
For the automation question, additionally by contact reason:
date, segment, contact_reason, attempts, contained, deflection_failures
Contact reason is the highest-value optional field. Without it the
automation-volatility question can only be asserted.
Do not fill gaps, do not exclude outliers, do not round. The outliers are
the jump component and they are what the reserve analysis is about.
## Limits
Seasonal indices assume the calendar pattern is stable across the window.
Where it is not — a policy change altering an institutional cycle — the
method will attribute real regime change to residual. Check the indices
across sub-periods before trusting the split.
Block 4 — reserve-and-reliability.md
# Sizing, allocating and pricing reliability
## Buffer and reserve are different instruments
Buffer standing, always staffed, covers ordinary residual variance.
Sized from the residual distribution.
Reserve activated on a trigger, covers shocks. Sized from shock
magnitude and recovery time, not from variance.
Conflating them produces a standing buffer sized for a shock — expensive and
still insufficient, because a shock exceeds any buffer anyone will fund.
The related concepts are covered elsewhere and are not restated here:
staffing to a percentile rather than a mean, and how to choose that
percentile, is the standard treatment of the buffer question. Tail-risk
measures such as conditional value at risk give a coherent way to size
against the tail rather than the mean. Shock taxonomy, time-to-recovery and
resilience mechanisms belong to the resilience literature.
## Portfolio reserve beats per-contract buffer
Where shocks across contracts, segments or sites are imperfectly correlated,
one shared reserve is substantially cheaper than N independent buffers. The
saving grows as correlation falls.
The obstacle is commercial, not technical: contracts are priced
individually, so the pooling benefit is invisible in the pricing model and
never gets claimed.
Method: compute the cross-contract shock correlation matrix, size the shared
reserve against the portfolio distribution, then allocate its cost back by
marginal risk contribution (Euler allocation) so each contract sees its true
incremental cost rather than a flat surcharge.
Unresolved, and worth naming: allocation prices the reserve but does not
decide who draws it when two contracts shock simultaneously. Imperfect
correlation is not zero correlation. That needs a stated claim-priority
rule, agreed before it is needed.
## The dedication tax
Clearance, regulatory, badge or named-team requirements break pooling and
force the operation off the square-root staffing curve. Dedicated books need
proportionally far more buffer than pooled ones for the same protection.
The analysis worth doing is not the cost — it is the provenance. Classify
each dedication requirement as statutory, regulatory, contractual, or
habitual. The last category is recoverable, and it is usually larger than
anyone expects. This is one of the few findings that changes how deals are
written rather than how they are staffed.
## Reliability as a priced product
Volatility protection is a capacity product. Other industries already sell
it and the constructs transfer:
Availability payment for reserve held (capacity charge)
Consumption payment for reserve activated (energy charge)
Tiered confidence alpha as the sellable variable
The question to answer is the shape of the cost curve of alpha — what does
moving from 80% to 95% confidence of never breaching actually cost? Buyers
cannot choose a reliability level nobody has priced for them.
Scope boundary, stated up front rather than discovered later: this is
viable in best-value and negotiated procurements. Under lowest-price
technically-acceptable evaluation, priced reliability loses by construction
— it is a price premium against a technical floor that does not score it.
The commercial thesis is real and it is not universal.
## What the buffer is actually worth
The expected loss from breach is not the SLA penalty:
E[loss] = P(breach) x P(non-renewal | breach) x contract lifetime value
+ recompete cost + portfolio reputational spillover
Observation worth testing in any specific book: almost every buffer looks
unaffordable against penalties and cheap against retention. If that holds
generally, the sector is systematically underinvesting in reliability
because it is measuring the wrong loss.
This is the calculation to run before arguing about headcount. It reframes
buffer from a cost line into an insurance premium with a stated payout.
## What to collect
Contract inventory: segment, dates, option years, SLA terms, value
Dedication requirements with provenance (statutory / regulatory /
contractual / habitual)
Pool structure: which agents can serve which contracts, and what
prevents pooling
Per-segment daily volume series, SEPARATE, for correlation
Staffing and schedule data is needed to cost any of this, and is not needed
for the correlation or provenance work. Treat costing as a second phase,
gated on the first producing a finding worth costing.
## Limits
Correlation estimated from a short window is unstable, and shock
correlation specifically is estimated from few events by definition. Treat
portfolio reserve sizing as indicative until the event count supports it,
and state the event count alongside the correlation.
Education and briefing prompt
The blocks above are written for a model doing analytical work. This section is different: it is a prompt that generates a briefing deck explaining the problem to people who will never read this page.
Paste it into a Claude session together with any operation-specific context, and it produces a self-contained HTML presentation. It deliberately avoids the statistical vocabulary used elsewhere on this page — the translations are built into the prompt.
Produce a self-contained HTML presentation explaining a workforce planning
problem and what we intend to do about it. Single .html file, no external
dependencies, arrow-key navigation, one idea per slide, readable projected
in a room. Presenter notes beneath each slide.
AUDIENCE — three groups will see this deck. Write so all three follow it.
Operations leaders who run the accounts. They live the volatility daily
and will recognise every symptom. They need to know what changes for
them and what we will stop blaming them for.
Executives who care that service level is met and that contracts renew.
They need the commercial consequence in under a minute, and they will
not sit through method.
The workforce planning team, who will do the work. They need enough
method to see why the approach differs, and a clear, specific list of
what data to pull.
VOCABULARY — do not use these words. Use the plain version.
stochastic / random variable -> "demand arrives unevenly and we cannot
know in advance which weeks will spike"
coefficient of variation -> "how bumpy demand is relative to its size"
chance-constrained -> "staffing so we hit target 9 times out of
10, instead of staffing to the average"
alpha / confidence level -> "how sure we want to be"
sampling noise / binomial -> "when the numbers are small, the score
swings on luck rather than performance"
measurement window -> "the period we get graded over"
buffer vs reserve -> "everyday cushion" vs "emergency capacity"
marginal risk allocation -> "each contract pays its fair share of the
shared safety net"
secular trend -> "a real, lasting increase in demand"
deseasonalise -> "strip out the predictable seasonal
pattern so you can see what is left"
Analogies are welcome where they earn their place. Do not use more than
three in the whole deck.
STRUCTURE
1. THE PROBLEM (2-3 slides, this is the part everyone must get)
Some books have demand that moves for reasons outside our control and
faster than any planning cycle can react to. At the same time the
contract grades us on one number — service level — and renewal depends
on it. We are being asked to guarantee a fixed outcome on an input we
cannot predict. Today we handle that by forecasting the average,
staffing the average, and explaining the misses.
State plainly what that costs: we carry risk we have never priced.
2. WHY THE USUAL APPROACH FALLS SHORT (2 slides)
Traditional: one forecast, one staffing number, hit or miss.
What we are moving to: we accept we cannot predict the week. Instead we
decide how often we are willing to miss, and staff to that. "Meet target
in 9 months out of 10" is a decision someone can price and agree to.
"Meet target" is not, because nobody said how often.
Make the contrast concrete with one worked example in plain numbers.
3. TWO THINGS HIDING IN THE DATA (3-4 slides — this is the new material)
3a. Some of our misses are not misses.
When a period carries very few contacts, the score swings on luck.
An operation performing exactly at target will still show a miss
roughly half the time in a very thin hour, because a handful of
calls cannot produce a stable percentage. We currently treat every
miss as a staffing failure. Some of them are the maths of small
numbers. Adding people does not fix that; changing how we measure
does.
Also: the period we are graded over changes the answer. The same
performance can look like constant failure when scored hour by hour
and like a clean record when scored monthly. That is a contract
term, and it is negotiable.
3b. A rising trend can hide inside the noise for years.
This is the expensive one. When demand is bumpy, a genuine, lasting
increase takes a long time to become visible. On a book with the
bumpiness we typically see, a 2% monthly increase can run for
roughly two years before it is statistically detectable. By then
demand has grown by half, and on a 500-person operation that is
around 270 people we should have hired and did not.
Worse: the shortfall shows up as missed service level, and because
the book is volatile we blame the volatility. The noise causes the
miss and then supplies the excuse.
Say clearly: short-term swings and long-term growth are two
different problems, and the first one hides the second.
4. WHAT WE ARE GOING TO DO (2 slides)
Separate the volatility into the part we can predict (calendar,
seasonal, cyclical), the part that is genuine shock, and the part that
is just noise. Buffer only against the last. Hold reserve for shocks.
Test explicitly for growth hiding underneath.
Decide, as a business, how sure we want to be — and price it.
5. WHAT WE NEED FROM WORKFORCE PLANNING (2 slides, be specific)
Present as a checklist someone can act on this week. For each item say
what it is, the grain, how far back, and what question it answers.
Interval-level contact data for the thin coverage periods — three
separate months, not three in a row. Answers whether our misses are
real.
The contract terms in writing: what counts, what is excluded, and the
period we are graded over. Answers whether the numbers even compare.
Daily volume by account, kept separate, as far back as we can get.
Answers what is predictable and whether growth is hiding.
Forecasts with the date each was produced. Answers whether we are
wrong, or wrong in one direction.
Contact reason, if the system holds it. Answers whether automation has
changed the work or only the volume.
Note what we are NOT asking for yet: detailed staffing and schedule
data. That comes later, once the first findings justify it.
6. WHAT CHANGES IN FUTURE CONTRACTS (2 slides)
The period we are graded over should be a negotiated term, not a
default. It is often the cheapest protection available.
Very thin coverage periods need a measurement approach that works at
low volume, or they should be measured over a longer period.
Reliability should be a priced option: a customer who wants near-
certainty can buy it. Today we give it away at whatever level our
staffing happens to produce.
Where we serve several accounts, one shared emergency capacity is far
cheaper than separate cushions for each — but only if the contracts
are written to allow it.
Be honest that priced reliability works in negotiated deals and does
not work where the buyer must take the lowest compliant price.
7. WHAT WE ARE NOT CLAIMING (1 slide, and do not skip it)
The method is sound and the numbers so far come from modelling, not
from our own data. The first phase tests whether it holds here. If the
misses turn out to be real capacity shortfall, that is a finding too,
and the case for reserve and priced reliability gets stronger rather
than weaker. We are set up to learn either way.
TONE
Confident, plain, unhurried. No hedging and no hype. Never imply the
current team has been doing it wrong — the approach we are replacing is
the industry standard and it is what everyone does. Frame it as: the
measurement has been asking a question the data cannot answer, and we
found that out.
FORMATTING
Dark background, high contrast, generous type. Numbers large and few.
Every quantitative claim on a slide carries a one-word source marker in
the presenter notes: modelled, computed, or measured. Nothing on this
subject is "measured" yet — say so where it appears.
Note on the numbers in the prompt. The figures quoted — roughly two years to detect 2% monthly growth, around 270 FTE short on a 500 FTE base — are computed from the trend-detection arithmetic in Block 3 at 30% monthly variability. Substitute the operation's own variability before presenting; the shape of the finding holds, the magnitude does not transfer.
Usage notes
- Sizing. Block 1 is roughly 470 words and loads with every message in the project. Blocks 2–4 total roughly 2,300 words and are retrieved only when the routing table sends the model to them.
- Scope. Method only. No organisational data, no sector-specific content — deployable in any operation where demand variance is structural.
- Provenance of figures. Marked
[computed]where derived analytically and[demonstrated]where produced by simulation on constructed data with known ground truth. None of the quantitative results have been tested against field data. They establish that the methods recover known answers, which is necessary and not sufficient. - Relationship to existing pages. This pack deliberately does not restate percentile staffing, threshold economics, resilience mechanisms or tail-risk measures. Those are linked in the infobox and summarised in one paragraph where a block needs them.
Change history
| Version | Date | Change |
|---|---|---|
| 1.0 | 2026-08-06 | Initial publication. Instruction block plus three reference files. |
| 1.1 | 2026-08-06 | Added trend-detection under noise to Blocks 1 and 3; added the education and briefing prompt. |
See also
- Staffing to Percentile vs Mean Forecast — the standard treatment of the buffer question
- Service Level Target Selection — economics of choosing the threshold
- Workforce Resilience and Adaptive Capacity — shock taxonomy and recovery
- Conditional value at risk — coherent tail-risk measure
- Variance Harvesting — the inverse problem, exploiting variance rather than absorbing it
- Wiki:Packs — the pack registry
