Wiki:Packs/Workforce Decisioning

From WFM Labs
Pack
ID CP-WFM-028
Name Workforce Decisioning
Slug workforce-decisioning
Type Pack
Domain WFM
Status Active
Version 1.0
Contents 1 instruction block + 4 reference files + briefing prompt
Builds on Workforce Decisioning
Decision Primitives for Workforce Operations
From Signal to Decision: The OVIX Case
Supply-Side Routing
Human Gates and Number Grades
Variance Routing by Horizon
Supply Elasticity Lever Inventory
Source Wiki:Packs/Workforce Decisioning

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 supports a workforce leader, a real-time operations lead or an analytics owner who wants to make the decisions between their measurements and their actions visible, typed and logged, and then to automate the cheap, reversible classes on evidence. It does four jobs: it writes an operation's decisions as primitives (proposition, choice, score) with the state each needs; it designs the three tiers (signal, decision, action) and the proposal row that passes between them; it routes a gap to a source of supply through eligibility, a fanned-out choice, pooling arithmetic and a gate; and it sets up the decision log and the calibration review that release a class of action or keep it proposed.

The problem it addresses

Between every number a workforce function produces and every action it takes there is a decision, and in most operations it is made in someone's head while reading a dashboard, or by a threshold written years ago. The decision has no type, no stated confidence and no record. Automation is then argued about in the abstract, because nothing has been measured that would say which decisions are reliable enough to release.

The pack treats the decision as a designed component. A decision is a typed question (is this true; which of these; where on this scale) answered on versioned, graded state, with a confidence that is measured against outcomes on the operation's own data. Code does the arithmetic. A rule, a classifier or a typed decision model does the judgement. A person signs at the gate until the class has a published catch rate, after which the class is released. Supply-side routing is the first application: which lever, pool or partner absorbs a gap, at which horizon.

What this pack does not restate. The grades and gates are at Human Gates and Number Grades; the horizon assignment is at Variance Routing by Horizon; the levers are at Supply Elasticity Lever Inventory; the index is at Operational Volatility Index (OVIX). The reference files summarise what the method needs from each and point back for the rest.

When to use it

  • Writing down, for the first time, the decisions a real-time or planning team actually makes, as typed questions with the state each needs
  • Designing or auditing a decision layer between a signal (an index, a variance decomposition, a capability record) and an action platform
  • Deciding which classes of action can be automated, on what evidence, and at what confidence threshold
  • Replacing a lever checklist with a routing decision: eligible options, a fanned-out choice, pooling arithmetic, a gate
  • Setting up a decision log and a quarterly calibration review

Not needed for an operation whose only real-time action is a threshold alert and a phone call, and which has no intention of logging decisions; nor for strategic decisions (operating model, partner selection, annual capacity), which are not bounded questions.

How to deploy

  1. Create a project, named for the operation rather than the method
  2. Copy Block 1 into the project's custom instructions
  3. Save Blocks 2–5 under the filenames given in their headings and upload as project knowledge, together with the operation's lever inventory or intraday playbook, its real-time alert rules, a sample of recent real-time decisions and dispositions if any exist, and the capability or skills record format
  4. Start a conversation with the briefing prompt in Block 6

Block 1 — Project instructions

This project designs and audits a workforce decision layer: the typed,
logged decisions that sit between an operation's measurements (forecasts,
indices, variance decompositions, capability records) and its actions
(schedule moves, routing changes, overtime offers, partner calls). Every
decision is one of three primitives: a proposition answered with a
probability, a choice answered with one option from a stated list plus a
probability for each, or a score answered with a position on a written
rubric. Code calculates; a rule, a classifier or a typed decision model
judges; a person signs at a gate until a class of action has a published
catch rate.

ROUTING
| Task | Open |
|---|---|
| Write a decision as a primitive; define its state, confidence, rubric | decision-primitives.md |
| Design the three tiers, the proposal row, the band-to-gate ladder | signal-to-decision.md |
| Route a gap to a source of supply; eligibility, blockers, fan-out, Erlang check | supply-side-routing.md |
| Set thresholds, release a class, design the log and calibration review | gates-log-and-calibration.md |

DISCIPLINES
- Never ask a model to do arithmetic, counting or date comparison; compute it and supply it as state.
- Every question is written positively and singly; no negations, no two-step inferences.
- Every choice list carries an explicit "none" or "other" option.
- Every score has a rubric with written criteria per level; no rubric, no score.
- Confidence is how concentrated an answer is; it is not the probability the answer is right. Gate thresholds apply to the winning answer's locally calibrated probability; confidence is reported beside it.
- A decision inherits the weakest grade of its inputs; grade and confidence are reported separately.
- No class of action is released on a maturity level or on vendor benchmarks; only on the operation's own calibration and catch rate.
- Eligibility (capability records, blockers, working-time rules, frozen zones) is computed before any choice; the model chooses only among eligible options.
- Pooling effects on both the lending and the borrowing pool are computed per proposal before it reaches the gate.
- Figures that are judgement rather than measurement are tagged [estimated]. Simulation or illustrative numbers are never presented as predictions.
- Vendor-neutral: describe the model class, never recommend a product; any product figure is labelled vendor-reported.

OUTPUT
Decision specifications as tables: question, primitive, state fields with
grade, answerer, class, gate, threshold. Proposal rows in the row format.
Routing decisions showing the eligible set, the distribution, the computed
effect on both pools, the gate outcome and what was logged. Always state
what is computed, what is judged and who signs. Refuse to produce a
confidence threshold for a class whose calibration has not been measured;
offer the measurement plan instead.

Source: Wiki:Packs/Workforce Decisioning (CP-WFM-028) v1.0

Block 2 — decision-primitives.md

# Decision primitives

Convention: figures tagged [estimated] are judgement, not measurement.
Derived from "Decision Primitives for Workforce Operations", wiki.wfmlabs.org.

## The three shapes

| Primitive | Question | Output | Confidence |
|---|---|---|---|
| Proposition | Is this true of the state? | p in [0,1] | abs(2p - 1) |
| Choice | Which one of these options? | winner + probability per option | (p_max - 1/K) / (1 - 1/K) for K options, in this series' convention |
| Score | Where on this ordered rubric? | probability per level + probability-weighted mean of level indices | from how far apart the plausible levels are |

A score is a position on a rubric, not a measurement. Keep the per-level
distribution; two different distributions can share a mean.

## Writing a question

- One statement, positive form. "The event reaches queue Q's customers
  within the pre-positioning window", not "it is not the case that...".
- Supply only the state the question needs. Extra state degrades accuracy;
  narrative text in state can steer the answer.
- Choice lists: finite, under ~20 options in practice, always with "none"
  or "other". Test under option permutation before production.
- Score rubrics: 2 to 10 levels, each with written criteria. Review rubrics
  on the same cadence as calibration.

## Confidence vs calibration

- Confidence: a statistic of one answer's distribution. Says how
  concentrated the answer is. Does not say how likely it is to be right.
- Calibration: a property of a population of answers. Among cases given
  probability near 0.8, about 80% should turn out true.
- Instrument: the Brier score, decomposed (Murphy 1973) as
  Brier = reliability - resolution + uncertainty: reliability (stated vs
  observed; lower better), resolution (how much predictions separate
  cases; higher better), uncertainty (outcome variance, fixed by the base
  rate). Picture: the reliability diagram.
- Use strictly proper scoring rules (Brier, log score) to compare
  answerers; accuracy at a fixed threshold is proper but not strictly, so
  it cannot tell a calibrated answerer from an overconfident one.
- Thresholds are set on the winning answer's calibrated probability, not
  on confidence.
- Calibration is local. Vendor benchmarks do not transfer. Measure on the
  operation's own outcomes per question class before any threshold means
  anything.

## Three answerers

| Answerer | Reads narrative state | Needs labelled data | Explanation | Use for |
|---|---|---|---|---|
| Rule | no | no | the rule itself | thresholds, eligibility, legal limits, bands |
| Fitted classifier + calibration layer (incl. fine-tuned text classifiers) | via features, or directly if fine-tuned on text | yes, per question | features and weights | breach prediction, absence patterns |
| Typed decision model | yes | no per-question training; local calibration still required | none; audit through the log | reading signals, matching events to queues, choosing among levers with narrative context |

A mature tier runs all three. Log which answerer answered each question;
calibration is measured per answerer and class.

## Mapping table (starter)

| Decision | Primitive | State needed | Acts on it |
|---|---|---|---|
| Interval will breach | Proposition | forecast, staffing, queue, AHT trend | real-time analyst; released break moves |
| Signal affects queue | Proposition | signal text, queue geography/product, sites | forecasting intake |
| Lever to absorb a gap | Choice | gap by cause + persistence, eligible levers, blockers | analyst at action gate |
| Loop that owns a variance component | Choice | sign history, window, pipeline | short-term reforecast cycle |
| Severity of exogenous event | Score (index bands) | event, proximity, time-to-impact, weights | ROC by band |
| Source grade of intake signal | Score | sender, artifact, named queues, size | planner gate |
| Reversibility of an action | Score 0-2 | action class, notice, contract | gate design |
| Capability record is stale | Proposition | evidence count, window, decay | placement eligibility |

## Composition rules

1. Code calculates; the model judges.
2. Fan out independent questions on one state version; filter in code.
3. Score dimensions separately; weight in code.
4. Keep irrelevant state out.
5. Gate on calibrated probability and consequence together.
6. Log question, state version, answerer version, answer, outcome.

## Grade vs confidence

A decision is a computed object: it inherits the weakest grade [M][C][E][A]
of its inputs. A choice on an [A] gap is an [A] choice at any confidence.
Report both; neither raises the other.

## Limits

Typed models read literally, cannot plan, return no reason, and are
unreliable at arithmetic. Their accuracy, latency and calibration claims
were vendor-reported at the time of writing (late 2026).

Block 3 — signal-to-decision.md

# Signal to decision: the three tiers

Convention: figures tagged [estimated] are judgement, not measurement.
Derived from "From Signal to Decision: The OVIX Case", wiki.wfmlabs.org.

## Tiers and their guarantees

| Tier | Publishes | Must be | Fails by |
|---|---|---|---|
| Signal | typed state, every number graded, versioned, scoped (per region / pool) | ignorant of queues, levers and gates | feed loss, staleness |
| Decision | proposal rows: question, answer + distribution, confidence, state versions, weakest grade, reason code, class, gate, expiry | idempotent, no side effects, reproducible from logged inputs | miscalibration; answering a question it was not asked |
| Action | executed change + outcome written back to the row | executes only from a signed row (person or released class) | acting without authority |

Keep the tiers separate so that a failure can be attributed to one.

## Posture is state

The right action on the same signal differs by the operation's current
posture (occupancy, service, staffing vs requirement, movable breaks and
offline work in the next hours). The decision tier reads two versioned
feeds: the signal from outside, the posture from the real-time feed.
Their refresh rates must be comparable.

## Proposal row format

| Field | Example |
|---|---|
| Question, primitive | "First lever for queue Q", Choice |
| Answer + distribution | resequence breaks 0.88, defer offline 0.07, confirm overflow 0.04, none 0.01 |
| Probability, confidence | 0.88 calibrated on the class's dispositions; confidence 0.86 |
| State versions | signal 06:20:14, posture 06:19:45 |
| Weakest input grade | [E] time-to-impact |
| Reason code | ELEVATED_PREPOSITION (controlled list, never prose) |
| Class, gate, status | break resequencing, action gate, released at 0.80 on calibrated probability |
| Expiry | 07:20; void after, must be re-asked |
| Signature, outcome | filled by the action tier |

Reason codes, not sentences: the typed model returns no reasoning, and a
sentence generated separately would explain an answer it did not produce.

## Worked fan-out on an Elevated index reading (5.4 on 0-10; illustrative)

| Question | Primitive | Answered by | Answer | Conf |
|---|---|---|---|---|
| Event reaches queue Q customers in window | Proposition | typed model | 0.91 | 0.82 |
| Event reaches queue Q supply (sites, home workers) | Proposition | typed model | 0.23 | 0.54 |
| Which repertoire configuration fits (6 + none) | Choice | rule if indexed, else model | C3 0.66 | 0.60 |
| First lever (band lever set + none) | Choice | model or classifier | resequence breaks 0.88 | 0.86 |
| Reversibility of that lever (0-2) | Score | rule | 1.9 | 0.93 |

Code then: hours moved, working-time check, gate lookup. Break
resequencing released at 0.80; winning probability 0.88 -> executes,
logged. Configuration choice at 0.66 < class threshold 0.75 -> proposed to
the analyst with top two options.

## Band to gate ladder (index 0-10)

| Band | Range | Typical action | Reversibility | Usual status |
|---|---|---|---|---|
| Nominal | 0-2 | none, logged | - | no proposal |
| Watch | 3-4 | confirm signal; ask reach propositions; brief next shift | free | released (questions and a notification) |
| Elevated | 5-6 | resequence breaks, defer offline, confirm overflow | free / low | eligible for release for break and offline moves once catch rate published; overflow proposed |
| High | 7-8 | activate reserve or overflow, voluntary overtime, relax non-critical occupancy constraints, notify site leadership | at cost | proposed; analyst inside guardrails, manager outside |
| Severe | 9-10 | declare incident, continuity | irreversible | person only |

## Versus threshold alerting

Alert fires on a number; decision answers a question with state. False
positives handled per proposal by calibrated probability, not by moving
the band (signal-detection trade: fewer false alarms, more missed windows). Attention spent on proposals ordered by consequence.
Every prompt leaves a scorable row.

## Measures

Acceptance rate by class; catch rate by class (wrong proposals caught
before execution; the release figure); disposition precision by band;
calibration by question class; lead time realised (signal version -> executed action,
versus signal -> first endogenous alert).

## Where the model fits

Only on reading tasks a rule cannot enumerate: does a narrative event reach
this queue or site; which configuration fits when the archive is not
cleanly indexed. Bands, arithmetic and reversibility stay rules or code.

Block 4 — supply-side-routing.md

# Supply-side routing

Convention: figures tagged [estimated] are judgement, not measurement.
Derived from "Supply-Side Routing", wiki.wfmlabs.org. Lever codes follow the
Supply Elasticity Lever Inventory (A hours per head, B who can do what,
C where hours come from, D pay, E demand shaping).

## Definition

Which source of capacity absorbs a gap between required and available
hours: lever, pool, partner, shift or worker, at which horizon. Contact
routing decides where a contact goes; supply-side routing decides where
the capacity comes from. Same capability record, different clocks.

## Sequence

1. Horizon: persistence routing assigns the gap component to a loop
   (interval / short-term / capacity / structural). Rule.
2. Eligibility: compute the eligible option set. Rule and records.
3. Fan-out: three primitives on one state version. Judgement.
4. Effect: hours moved x rate; Erlang on both pools. Code.
5. Gate: class, released or proposed, threshold. Table.

## Option set by clock

| Clock | Levers | Lead time | Reversibility | Cost |
|---|---|---|---|---|
| Minutes | B5 intraday reallocation; break/offline moves; routing thresholds | now | free | lender's occupancy and service; interrupted offline work |
| Hours | A4 overtime and A3 VTO offered on the day (fast form of levers the inventory lists at days' lead); D1 opt-in surge pay; B2 back office as shock absorber | same day | low until accepted | premium hours; catch-up |
| Days | C3 bench/burst hours; A7 shift trades; C6 follow-the-sun; E1 shift simple contacts to automation; E2 deferrable-work calendar | 1-3 days by notice | moderate; partner hours billed once called | bench rate; cross-region cost |
| Weeks | C1/C2 contract bands; B1 cross-trained group resize; A5 variable-hour contracts; A6 contingent and returner pools; class starts | weeks to a pipeline | low once committed | notice; training hours; hiring |

A gap routed to the interval loop takes only minute/hour levers. A gap
routed to the capacity loop may use them until the pipeline delivers but
not beyond; carrying a persistent gap on overtime for a quarter is a
capacity decision not taken.

## Eligibility

- May vs should: a capability record says what a worker may do (product,
  channel, language, platform, client, vetting; evidence grade; published
  level or INSUFFICIENT DATA). Standing says what they should do. Lending
  moves should -> may. Stale or insufficient record = not eligible,
  whatever the roster tag. Lending on a stale record = staffed-but-short.
- Blockers between two pools: HARD (language, dedication, location):
  route does not exist. SOFT (trainable): eligible at a price in training
  hours and a lead time, not today. PRICED (vetting, contract clause):
  eligible at a stated price and notice.
- Rules: working-time limits, rest, frozen zone of the published week,
  contractual notice, overtime ceiling. Remove options in code.

The model chooses only among eligible options. If it is ever offered an
ineligible one, the option list is wrong.

## The fan-out

| Step | Primitive | Content |
|---|---|---|
| Gap persists past window | Proposition | whether to prepare an hour-clock route in parallel |
| Which eligible route | Choice (eligible set, any composites the gate design allows, + none) | the judgement; keep the full distribution |
| Lender quality risk | Score 0-3 | will lending push the lender out of band |

Then code: expected cost per gap-hour closed; occupancy and service after
the move on BOTH pools by Erlang. Pooling is not monotone: lending out of
a pool near its limit can cost the lender more service than the borrower
gains; pooling queues with markedly different handle times can increase
waiting (Tekin, Hopp and Van Oyen 2009).

## What stays deterministic

Horizon assignment; cost ordering within a clock (the tier's choice is
checked against it; a dearer preference is shown at the gate); eligibility;
frozen zones and change-cost schedule; service-level differentiation
between pools (a stated policy, implementable by queue-ratio rules).

## Worked example (illustrative)

Tuesday 11:00, pool P, afternoon ~200 required hours, gap 6% = 12 hours,
cause level, first day of a level excursion this week (persistence counts
in weekly cycles) -> interval/short-term clocks.

| Route | Eligible | Price / lead |
|---|---|---|
| B5 reallocate 4 from pool R | 2 yes; 2 SOFT (refresher) | free / 4 training h each, not today |
| A4 12 overtime hours | yes | premium, same day |
| B2 defer back-office batch, lend its 3 agents certified on P | yes for the 3 certified | catch-up tomorrow |
| C3 8 partner bench hours | no today (48 h notice) | - |
| None, absorb via occupancy | yes | P occupancy ~93% for 2 h; 2 intervals at risk |

Fan-out: persists 0.41; route B5 0.58, B2 0.27, A4 0.12, none 0.03
(uncalibrated until the class has disposition history); lender risk 1.2.
Code: 2 lent for 3 h closes 6 of 12 gap-hours; R stays in band; P at-risk
intervals 2 -> 1. By cost per gap-hour: B5 and B2 free, A4 dearest; no
contradiction. Class intraday reallocation released at 0.75 on calibrated
probability; 0.58 < 0.75 -> proposal to analyst with B5 and B2. Gate
design allows composed routes: analyst signs B5 (2 agents) + B2
(residual); refreshers queued as a priced unblock. Disposition: closed by
13:30; row enters calibration set.

## Measures

Share of gap-hours by route and clock, weekly. Routing regret (route taken
vs cheapest eligible route that would have closed the gap, judged at
disposition; report as a distribution). Override rate by class (high =
wrong option list or wrong state). Staffed-but-short incidents. Calibration
of the route choice by class.

## Coupling to contact routing

Publish supply-side moves to the routing engine as state changes. A
contact-routing policy that protects one pool at another's expense is a
standing supply-side decision; put it in the eligible set as a constraint
with an owner.

Block 5 — gates-log-and-calibration.md

# Gates, the decision log and the calibration review

Convention: figures tagged [estimated] are judgement, not measurement.
Condensed from Human Gates and Number Grades and the Workforce Decisioning
series, wiki.wfmlabs.org.

## Release rule

A class of action is released when a catch rate (share of wrong proposals
the gate caught before execution) has been published for it, not at a
maturity level and not on a vendor benchmark. Parasuraman, Sheridan and
Wickens (2000) choose the automation level separately per stage, with
decision selection and action automated least where outcomes are costly
and reliability unproven. Release is per class, reviewed on a cadence,
and revocable.

## Act / propose / log ladder

| Condition | Outcome |
|---|---|
| Class released AND calibrated probability of the answer >= class threshold AND no rule violated | act; log; outcome written back |
| Class released AND probability < threshold | propose to the gate's signer with the distribution and computed effect |
| Class not released | propose, whatever the confidence |
| Question not applicable (upstream proposition below 0.5) | log only |
| State version expired | void; re-ask |

Thresholds are set per class from the class's reliability curve on local
outcomes, by reversibility: free actions lower, costly ones higher,
irreversible ones person-only. Refuse to set a threshold for a class
without a curve; propose the measurement plan instead.

## Number grades on inputs

[M] measured from a versioned ledger, definition cited. [C] computed by a
stated formula; inherits weakest input. [E] estimated: a range and the
assumption. [A] asserted: one source, no artifact; never load-bearing
alone. A proposal row carries the weakest grade of its inputs. A gate
signer sees grade and confidence as separate fields.

## Decision log schema (minimum)

| Field | Notes |
|---|---|
| row id, timestamp | |
| question id, primitive, version of the question text | questions change; keep versions |
| answerer | rule / classifier id / model id and version |
| state versions | every feed used |
| weakest input grade | |
| answer, full distribution, calibrated probability, confidence | |
| class, gate, status at the time | released or proposed, threshold |
| signer | person or "released:<class>" |
| action taken | may differ from the proposal |
| outcome at disposition | what happened; who judged; when |
| disposition verdict | correct / incorrect / moot |

## Calibration review (quarterly, or when a class accumulates ~100 rows) [estimated]

1. Per question class and answerer: reliability diagram and Brier score
   with Murphy decomposition. Compare answerers with proper scoring rules.
2. Per action class: acceptance rate, override rate and what was signed
   instead, disposition precision, catch rate.
3. Decide per class: release, keep proposed, revoke, change threshold,
   rewrite the question, change the option list, rewrite the rubric.
4. Retire questions whose answers no longer change decisions; add
   questions where the override log shows a judgement nothing captures.
5. Record the decisions of the review in the same log format.

## Failure modes to check

- Threshold applied to an uncalibrated answerer, or set on confidence
  rather than on calibrated probability.
- Model given arithmetic, counting or dates.
- Choice list without "none"; results not tested under permutation.
- Narrative feed supplied whole as state (steerable, degrades accuracy).
- Posture feed lagging the signal feed.
- Gate that approves without producing anything (decays into assent);
  require an amended figure, a reason code, or a decision per row.
- Rows without expiry acted on late.
- Class released on vendor figures or on a maturity level.

Block 6 — Briefing prompt for the first conversation

Read the project knowledge. Then, from the material I have uploaded about
this operation (lever inventory or intraday playbook, alert rules, any
logged real-time decisions, the capability record format), do the
following and stop after each numbered step for my confirmation:

1. List the decisions the real-time and planning teams actually make,
   as a table: decision, primitive, state needed (with grade), current
   answerer (person / rule), who acts. Mark any that are not bounded
   questions and say why.
2. For the two highest-frequency decisions, write the question text,
   the option list or rubric, and the state fields, following the
   composition rules. Flag anything the model would be asked to compute.
3. Propose the three tiers for one signal in this operation, with the
   proposal row format filled for one realistic case tagged [estimated].
4. Route one recent gap: eligible set with blockers, the fan-out, the
   computed effect on both pools (state the Erlang assumptions), the
   gate outcome, and the log row.
5. Draft the decision log schema and the first calibration review plan,
   including which classes cannot have a threshold yet and what evidence
   would release them.

Keep every figure that is judgement tagged [estimated]. Never recommend a
product; describe the model class. Say what is computed, what is judged,
and who signs, in every step.

Usage notes

  • Sizing. Blocks 2–5 total about 2,600 words of project knowledge; Block 1 is about 450 words and loads with every message.
  • Drift. The reference blocks are derived from the four articles named in the infobox. When an article changes, regenerate the block from it rather than editing the block in place, and bump the version.
  • Scope. The pack designs and audits a decision layer; it does not build one. Platform integration, the routing engine's API and the data pipeline for the signal tier are outside it.
  • Model class. The pack describes typed decision models as a class. Any product's latency, cost and calibration figures are vendor-reported at the time of writing and must be labelled as such in outputs.

Change history

Version Date Change
1.0 2026-11-10 Initial release: instruction block, four reference files, briefing prompt

See also