Decision Primitives for Workforce Operations

From WFM Labs


Decision primitives for workforce operations are the three shapes into which a bounded workforce decision can be cast so that software can consume the answer: a proposition, answered with a probability that it is true; a choice, answered with one option from a stated list and a probability for each; and a score, answered with a position on an ordered rubric and a probability for each level. Every answer carries a confidence, a measure of how concentrated the answer is, which is distinct from the probability that the answer is correct. The three shapes are not new; they are the output types of classification, selection and ordinal rating that statistics has used for a century. What is new is a class of software model, so far represented by one vendor's early-access product, that is reported to answer all three on unstructured state in a single pass with probabilities trained toward calibration. If independent evaluation confirms the reports, the primitives become practical to ask at the cadence of an operation rather than at the cadence of a report.

The page defines the three primitives, states what confidence and calibration mean, maps the common decisions of a workforce function onto the primitives, compares the three ways a primitive can be answered, and gives the composition rules that keep arithmetic in code and judgement in the model. It is the vocabulary page of the Workforce Decisioning series.

Why a decision needs a type

A decision that software will act on must arrive as a value whose type is known in advance. A string of text is not such a value: it must be parsed, the parse can fail, and the text carries no usable statement of how sure its author was. The distinction is sometimes drawn with Kahneman's two systems: the fast, pattern-matching System 1 and the slow, deliberate System 2.[1] A generative model that reasons through a problem and writes an explanation is doing System 2 work. A bounded judgement, is this true, which of these, how severe, is System 1 work, and a vendor has adopted the phrase "System One model" for software built to do only that.[2]

The metaphor should not be pushed far. What matters operationally is the contract: a question with a bounded answer space goes in, a value from that space comes out with a probability attached, and the caller never parses prose.

The three primitives

Proposition

A proposition is a single statement that is either true or false of the supplied state. The answer is a probability p that it is true. Values near 0 or 1 are strong predictions; values near 0.5 say the state does not settle the question.

Workforce question State supplied Answer shape
Will the 14:30 interval in pool P breach its service threshold? The interval's forecast, staffing, current queue, handle-time trend p = 0.78
Does this transport-disruption bulletin affect the customers served by queue Q? The bulletin text, the queue's geography and product p = 0.91
Is this absence part of a pattern rather than an isolated event? The worker's absence history, the team's, the calendar p = 0.34

In the convention used by the model class, the confidence of a proposition is |2p − 1|: zero at p = 0.5, one at either extreme.[3]

Choice

A choice is a question with a stated, finite list of options, of which exactly one is to be selected. The answer is the winning option, a probability for every option and a confidence derived from how far the winner leads.

Workforce question Options Answer shape
Which lever absorbs this afternoon's hours gap in pool P? Intraday reallocation · overtime · defer back-office work · bench hours · none reallocation 0.58, defer 0.27, overtime 0.11, bench 0.01, none 0.03
Which planning loop owns this week's handle-time variance? Interval · short-term · capacity · structural capacity 0.71, short-term 0.22, interval 0.05, structural 0.02
Which team should receive this intake signal? Forecasting · scheduling · real-time · placement · other forecasting 0.83, …

Two design rules follow from how the primitive behaves. First, the list should include an explicit "none" or "other" option wherever the taxonomy may be incomplete, because the model will otherwise select the least wrong option with unwarranted confidence. Second, the result can change with the order of the options, so a production question should be tested under permutation.[3]

Score

A score is a position on an ordered rubric of two to ten levels, each level defined by written criteria. The answer is a probability for each level and a single score computed as the probability-weighted mean of the level indices. The score is a position on the rubric, not a measurement of anything: two quite different distributions over levels can produce the same mean, and the per-level probabilities should be kept.[3]

Workforce question Rubric Answer shape
How severe is this exogenous event for the operation? 0 Nominal · 1 Watch · 2 Elevated · 3 High · 4 Severe, each with the criteria of the OVIX action bands 1.6, with P(Watch) = 0.50, P(Elevated) = 0.40, P(High) = 0.10
How well sourced is this intake signal? The source grades of Human Gates and Number Grades 2.1
How reversible is this proposed action? 0 Irreversible · 1 Reversible at cost · 2 Freely reversible 1.8

The rubric is the whole of a score question. Levels without written criteria produce answers that cannot be audited, and a rubric copied from another operation produces a score calibrated to that operation.

Confidence and calibration

Two words carry most of the weight in this vocabulary and they are routinely confused.

Confidence is a statistic of the answer's distribution. For a proposition it is |2p − 1|. For a choice over K options this series uses (pmax − 1/K) / (1 − 1/K), which is zero when the winner is no better than chance and one when it is certain; a vendor's implementation may differ and should be stated. For a score it reflects how far apart the plausible levels are. Confidence says how concentrated the answer is. It does not say how likely the answer is to be right.[3]

Calibration is a property of a population of answers, not of one answer. A set of predictions is calibrated when, among all the cases to which the predictor assigned a probability near 0.8, about 80 percent turn out true. Probability verification was formalised for weather forecasts by Brier in 1950; Murphy's decomposition later isolated calibration as a separate component.[4] Murphy's decomposition separates the Brier score into three parts, Brier = reliability − resolution + uncertainty: reliability (how far stated probabilities are from observed frequencies; lower is better), resolution (how much the predictions separate the cases; higher is better) and uncertainty (the variance of the outcome, ō(1 − ō), fixed by the base rate and beyond any predictor's control).[5] A reliability diagram, stated probability on one axis and observed frequency on the other, is the standard picture; a calibrated predictor lies on the diagonal.[6]

Three consequences follow for a workforce function.

  • Calibration is measured on the operation's own outcomes. A model that is calibrated on a vendor's benchmark set can be badly miscalibrated on an operation's queues, and modern classifiers are known to be systematically overconfident unless corrected.[7] The decision log, the subject of a later page in the series, exists so that this measurement can be made.
  • A threshold is meaningless until calibration is known. "Act automatically above 0.8" is a policy about the calibrated probability of the chosen answer. Applied to an uncalibrated source it is a policy about nothing. For this reason the gate thresholds in this series are set on the winning answer's probability after local calibration, with confidence reported beside it rather than used as the gating statistic.
  • Strictly proper scoring rules should be used to compare answerers. A rule is proper when reporting the true belief is among the predictor's best strategies, and strictly proper when it is the unique best. The Brier and logarithmic scores are strictly proper. Accuracy at a fixed threshold is proper but not strictly: it rewards every report on the correct side of the threshold equally, so it cannot tell a calibrated answerer from an overconfident one.[8]

Mapping workforce decisions onto the primitives

Decision Primitive State that must be supplied Who acts on the answer
Interval will breach Proposition Interval forecast, staffing, queue, handle-time trend Real-time analyst; released automation for break moves
Signal affects queue Proposition Signal text, queue geography and product, site list The intake step of the forecasting loop
Lever to absorb a gap Choice Gap by cause and persistence, eligible levers with cost and lead time, blockers Real-time analyst or planner at the action gate; see Supply-Side Routing
Loop that owns a variance component Choice Component sign history, window length, pipeline length The short-term reforecast cycle; see Variance Routing by Horizon
Severity of an exogenous event Score Event, proximity to customers and sites, time to impact, vertical weights The ROC, by action band; see From Signal to Decision: The OVIX Case
Source grade of an intake signal Score Sender, artifact presence, named queues, direction and size The planner gate
Reversibility of a proposed action Score The action class, notice given, contractual terms The gate design, which sets the confidence threshold by reversibility
Capability record is stale Proposition Evidence count, window, decay, recent outcomes The placement engine's eligibility step

Decisions that do not appear in the table are usually not primitives. "What should the capacity strategy for next year be" is a planning problem with many coupled choices; "why did service fail on Tuesday" is a causal question. Both can be served by many primitive questions but neither is one.

Three ways to answer a primitive

A primitive is a contract, and three quite different components can honour it.

Answerer Reads unstructured state? Needs training data? Auditability Typical latency Typical use
Hand-written rule No No Complete; the rule is the explanation Microseconds Thresholds, eligibility, legal limits, the action bands of an index
Fitted classifier (logistic regression or gradient boosting on engineered features; a fine-tuned text classifier on raw text; either with a calibration layer such as Platt scaling, isotonic regression or temperature scaling) Through engineered features, or directly if fine-tuned on text Yes, labelled per question The features and weights are inspectable; calibration can be measured and corrected[9] Milliseconds Breach prediction, absence patterns, any question with a labelled history
Typed decision model Yes, text and structured records No per-question training; calibration must still be measured locally Returns no reasoning; auditable only through the log of inputs, answers and outcomes Tens to hundreds of milliseconds, as reported by the vendor Reading signals, matching events to queues, choosing among levers where the state includes narrative

The third row describes the class whose first instance, TypeSafe AI's Jev, entered early access in September 2026. The model takes state as text or structured data plus named questions, answers the three primitives above, samples all answers in one parallel pass rather than token by token, and was trained with an objective the vendor calls reinforcement learning for calibrated decisions.[10] Its reported latency of roughly 70 to 500 milliseconds and its reported parity with frontier generative models on decision-shaped tasks come from the vendor's own evaluation suite; as of November 2026 no independent benchmarks or calibration measurements existed, and the product was closed and waitlisted.[2] The page treats the class, not the product, as the subject.

The three answerers are not rivals. A mature decision tier runs all three: rules for everything a rule can express, classifiers for questions with a labelled history, and a typed model for questions whose state is narrative. The choice of answerer for a question is itself logged, because it changes how calibration is measured.

Composition rules

Six rules govern how primitives combine into a decision. Each answers a failure mode reported in third-party usage notes for the one product so far available, which is a single source; the rules are nonetheless conservative enough to apply to any answerer.[3]

  1. Code calculates; the model judges. Arithmetic, counting, date and interval comparison and Erlang staffing are done in code and supplied as state. The typed models are unreliable at all of them.
  2. Fan out, then filter. Independent questions on the same state are asked together, in parallel, and code discards the answers that turn out not to apply. Asking one question, reading the answer and then asking the next wastes the latency advantage and introduces ordering effects.
  3. Score dimensions separately; weight in code. A composite such as the severity of an event is built from separate scores for severity, proximity and recency, combined with explicit weights the operation owns, not from one question asking for the composite.
  4. Keep irrelevant state out. Extra state degrades accuracy, and text in the state can steer the answer. A question about queue Q is given queue Q's state, not the day's whole feed.
  5. Gate on probability and consequence together. A near-certain answer about an irreversible action is still a proposal. The threshold for automatic action is set per class of action by its reversibility and on the calibrated probability of the chosen answer for that class.
  6. Log the question, the state version, the answerer's version, the answer and the outcome. Without all five the decision cannot be calibrated, audited or released.

Relationship to number grades

The Agentic WFM Protocol grades every number as measured, computed, estimated or asserted, and a computed figure inherits the grade of its weakest input.[11] A decision is a computed object and the same rule applies: a choice made on an asserted gap is an asserted choice, whatever its confidence. Confidence and grade answer different questions. Grade says how the inputs are known; confidence says how concentrated the answer is given those inputs. A decision record carries both, and neither raises the other.

A worked example

At 06:20 the exogenous index for a region reads 5.4 on its 0–10 scale, in the Elevated band (the figures in this example are illustrative). The real-time analyst's morning routine, written as primitives and fanned out on the same state:

Question Primitive Answer Confidence
The flagged transport disruption affects the customers served by queue Q within its pre-positioning window Proposition 0.91 0.82
Which prepared configuration from the repertoire fits the affected sites and queues? Choice (six options plus none) Configuration C3, 0.66 0.60
Which lever is first to deploy? Choice (the Elevated band's lever set plus none) Resequence breaks away from projected onset, 0.88 0.86
Reversibility of that first lever Score (0–2) 1.9 0.93

Code then does what code should: it computes the hours the configuration would move, checks the break resequencing against working-time rules, and reads the gate table. Break resequencing is a released class with a published catch rate and a threshold of 0.80 on the winning option's locally calibrated probability; at 0.88 the proposal executes and is logged. The repertoire choice, at 0.66 against a class threshold of 0.75, goes to the analyst as a proposal with the two leading options and their probabilities. The analyst's decision is logged against the proposal; at the end of the day the disposition review records whether the disruption did reach queue Q and whether the configuration fitted, and the record enters the calibration set for the reach proposition and the repertoire choice.

Limitations

  • A primitive is only as good as its question. A proposition that embeds a negation or a two-step inference ("is it not the case that this signal fails to name a queue") is read literally and answered badly by the model class.[3] Questions are written positively and singly.
  • Rubrics drift. A score's levels are written once and the operation changes around them. A rubric review belongs in the same cadence as the calibration review.
  • No explanation is returned. A typed model gives an answer and a distribution, never a reason. Where a reason must accompany a decision at a gate, the reason is produced by a separate generative step or by the person signing, and is labelled as such.
  • The model class is two months old. As of November 2026, every performance claim on this page about typed decision models is vendor-reported and should be re-read when independent evaluations exist.

Maturity Model Position

  • Level 3 — Progressive: decisions are written as propositions, choices and scores and answered by rules or by people; answers and outcomes are logged. The primitives are a documentation discipline before they are a technology.
  • Level 4 — Advanced (The Ecosystem Emerges): fitted classifiers and typed models answer the questions whose state is narrative or high-volume; calibration is measured per class on the operation's outcomes; thresholds are set from the reliability curve.
  • Level 5 — Pioneering: the question set itself is maintained from the log, with questions retired when their answers stop changing decisions and added when a gate's override log shows a judgement no question captures.

See Also

References

  1. ↑ Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux. ISBN 978-0-374-27563-1.
  2. ↑ 2.0 2.1 TrueFoundry (2026). "TypeSafe AI's Jev: What 'System One Models' Actually Are". Third-party analysis of the vendor's launch; notes that benchmark, latency and calibration figures are vendor-reported. https://www.truefoundry.com/blog/typesafe-ai-jev
  3. ↑ 3.0 3.1 3.2 3.3 3.4 3.5 Burnhill, P. (2026). "Comprehensive project reference for TypeSafe Jev: concepts, architecture, primitives, strengths, limitations, use cases, patterns, and practical guidance". GitHub Gist, reviewed 9 October 2026 against jev-1.13.0. Third-party reference, not vendor documentation. https://gist.github.com/pjburnhill/adf8d28efcad9df037bfdece178ef965
  4. ↑ Brier, G. W. (1950). "Verification of forecasts expressed in terms of probability". Monthly Weather Review 78 (1), 1–3. doi:10.1175/1520-0493(1950)078<0001:VOFEIT>2.0.CO;2.
  5. ↑ Murphy, A. H. (1973). "A new vector partition of the probability score". Journal of Applied Meteorology 12 (4), 595–600. doi:10.1175/1520-0450(1973)012<0595:ANVPOT>2.0.CO;2.
  6. ↑ DeGroot, M. H., & Fienberg, S. E. (1983). "The comparison and evaluation of forecasters". Journal of the Royal Statistical Society: Series D (The Statistician) 32 (1–2), 12–22. doi:10.2307/2987588.
  7. ↑ Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. Q. (2017). "On calibration of modern neural networks". Proceedings of the 34th International Conference on Machine Learning, PMLR 70, 1321–1330. https://proceedings.mlr.press/v70/guo17a.html
  8. ↑ Gneiting, T., & Raftery, A. E. (2007). "Strictly proper scoring rules, prediction, and estimation". Journal of the American Statistical Association 102 (477), 359–378. doi:10.1198/016214506000001437.
  9. ↑ Niculescu-Mizil, A., & Caruana, R. (2005). "Predicting good probabilities with supervised learning". Proceedings of the 22nd International Conference on Machine Learning, 625–632. doi:10.1145/1102351.1102430.
  10. ↑ MarkTechPost (19 September 2026). "TypeSafe AI Releases Jev: A System One Model That Returns Typed, Calibrated Decisions Instead of Text". Press coverage of the vendor's launch announcement. https://www.marktechpost.com/2026/09/19/typesafe-ai-releases-jev/
  11. ↑ Human Gates and Number Grades, this wiki: the four number grades [M], [C], [E], [A] and the rule that a computed figure inherits its weakest input's grade.