From Signal to Decision: The OVIX Case
From signal to decision describes the path by which a reading of the Operational Volatility Index becomes an action on the operation, and uses that path to show the three-tier pattern of Workforce Decisioning end to end. The index page explains how outside events are scored and banded. This page starts where that one stops: a number has been published, and the question is what has to happen, and in which component, for the number to move a break, a routing threshold or an overtime offer while the operation is still meeting service. The answer is a signal tier that publishes typed, graded state; a decision tier that answers bounded questions about that state with a stated confidence and no side effects; and an action tier that executes only from a signed proposal and writes the outcome back.
The pattern is drawn from the index's own design, from the situation-awareness literature's separation of perceiving, comprehending and projecting, and from the human-factors finding that the level of automation should be chosen separately for each stage of a task, with decision selection and action automated least where outcomes are costly and reliability unproven.[1][2] It is worked through with the index because the index is the clearest case of a signal that arrives before the operation's own metrics move, which is exactly when a decision has to be made on something other than a queue. In Endsley's terms the signal tier is perception, the decision tier is comprehension, and the time-to-impact and expiry fields are projection.
Why a score is not yet a decision
An index reading of 5.4 in the Elevated band tells a Resource Optimization Center that something outside the operation is likely to raise demand soon. It does not say which queues, which sites, how many hours, which of the prepared responses fits, or whether the right move is to act now or to watch for an hour. The index page's own framing, "sensing plus repertoire", names the gap: detection without a prepared and selected response converts a warning into a shorter deadline.[3]
Between the reading and the response there are, in practice, four judgements and two calculations:
- Does this event reach the customers of queue Q inside the pre-positioning window? (judgement)
- Does it reach the sites or home-working population that supplies queue Q? (judgement)
- Which configuration in the repertoire fits the affected queues and sites? (judgement)
- Which lever deploys first, and how reversible is it? (judgement)
- How many hours does the chosen configuration move, and does it stay inside working-time rules? (calculation)
- Which gate does the resulting proposal fall under, and is the class released? (calculation, a table lookup)
In most operations all six are performed in a person's head in the minutes after the alert. The three-tier pattern makes each one a component with an interface.
Tier 1: the signal
The signal tier publishes state. For the index, the state is more than the number.
| Field | Content | Grade |
|---|---|---|
| Reading | The composite, 0–10, and its band | [C], inheriting the weakest component or event grade |
| Components | Severity, proximity to customers, proximity to supply, recency and time-to-impact, each on its own scale | Severity [M] or [E] by feed; proximity and recency [C] from the event's location and timing |
| Contributing events | The source items, their category, their geo-anchor, their source grade | [A] or better per item |
| Affected scope | Sites and customer concentrations inside the geo-anchor radius | [C], inheriting the event's grade |
| Time to impact | The expected onset window for demand | [E] |
| Version | Timestamp and feed versions | — |
Three properties are required of the tier. Scoping is argued on the index page; versioning and grading follow from its requirement to log every reading for calibration. The state is versioned, so that a decision can be reproduced from what was known at the time. Every number is graded in the four grades of Human Gates and Number Grades, because a proposal built on an asserted contributing event is an asserted proposal, and the proposal row's "weakest input" field can read [E] only when every contributing event is graded above [A]. And the state is scoped: a global reading for a multi-region operation averages away the locality that makes it actionable, so the tier publishes per region, per affected scope.[3]
A signal tier does not decide. It does not know about queues, levers or gates. Keeping it ignorant of them is what allows the same signal to serve the forecasting loop, which uses it as an external regressor, and the ROC, which uses it as a trigger.
Tier 2: the decision
The decision tier answers typed questions about the published state and about the operation's current posture, and returns proposals. It has no side effects: it does not move a break, send an offer or change a threshold. Its outputs are rows, each carrying the question, the answer, the confidence, the state version used and a reason code.
The questions
Written as the primitives of Decision Primitives for Workforce Operations, the morning's questions on an Elevated reading are:
| Question | Primitive | State supplied | Answered by |
|---|---|---|---|
| The event reaches the customers of queue Q inside its pre-positioning window | Proposition | The contributing events, queue Q's geography and product, time to impact | Typed model (a model constrained to return a probability, a choice or a score with no free text; see Decision Primitives for Workforce Operations), because the events are narrative |
| The event reaches the supply of queue Q (sites, home workers) | Proposition | The contributing events, queue Q's site list and home-working share | Typed model |
| Which repertoire configuration fits | Choice over the precomputed archive plus "none" | The affected scope, the current posture, the configurations' preconditions | Rule where the archive is indexed by scope; typed model where the match is narrative |
| Which lever deploys first | Choice over the band's lever set plus "none" | Posture: occupancy, service, staffing against requirement, break and offline schedule for the next three hours | Typed model or classifier |
| Reversibility of the chosen lever | Score, 0–2 | The lever's class and notice terms | Rule; reversibility is a property of the class, not of the day |
The questions are asked together, on one state version, and code discards the ones that do not apply: if the first proposition comes back below 0.5 the lever questions are moot for queue Q, and their answers are logged but not acted on. This fan-out is cheap when the answerer is low-latency, and it avoids the sequential dependence of asking one question at a time; the exemplar's third-party reference describes it as the intended pattern.[4]
Posture is state too
The most common error in building the tier is to supply the signal and not the posture. The right action on a 5.4 reading at 06:20 with the queue clear and breaks movable is different from the right action on the same reading at 11:40 with occupancy at ninety-four percent and the next break window gone (illustrative). The decision tier therefore reads two state sources: the signal, from outside, and the posture, from the operation's own real-time feed, each versioned. Real Time Threshold Alerts and Escalation Protocols describes the posture feed; the point here is that it is an input to the decision rather than a competing alarm.
What a proposal row carries
| Field | Example |
|---|---|
| Question and primitive | "First lever for queue Q" · Choice |
| Answer and distribution | Resequence breaks 0.88 · defer offline 0.07 · confirm overflow 0.04 · none 0.01 |
| Probability and confidence | 0.88 for the winning option, calibrated on the class's disposition history; confidence 0.86 by the series' convention for a choice of six |
| State versions | Signal 06:20:14 · posture 06:19:45 |
| Grade of weakest input | [E], the time-to-impact estimate |
| Reason code | ELEVATED_PREPOSITION, a code from a controlled list, not prose
|
| Class and gate | Break resequencing · action gate · released, threshold 0.80 on the calibrated probability |
| Expiry | 07:20, after which the proposal is void and must be re-asked |
A reason code rather than a sentence is deliberate. The typed models return no reasoning, and a sentence generated separately would be an explanation of an answer the generator did not produce. The code is what the decision tier can honestly state; the person at the gate can ask for the distribution and the inputs, which are the actual grounds.
Tier 3: the action
The action tier executes a change on a platform from a signed proposal row and writes the outcome back to the same row. The signature is either a person's, at the action gate, or the automatic signature of a released class. The rule for release is the one the agentic protocol states: a class is released when a catch rate, the share of wrong proposals the gate caught before execution, has been published for it, not at a maturity level; and following Parasuraman, Sheridan and Wickens, action and decision selection are the stages automated last where errors are costly.[5][2]
For the index's bands this produces a natural ladder.
| Band | Range | Typical action | Reversibility | Usual gate status |
|---|---|---|---|---|
| Nominal | 0–2 | None; logged only | — | No proposal |
| Watch | 3–4 | Confirm the signal; ask the reach propositions for the affected queues and sites; brief the incoming shift | Free | Released: the actions are questions and a notification |
| Elevated | 5–6 | Resequence breaks away from projected onset; defer discretionary offline activity; confirm overflow availability | Free to low cost | Eligible for release for break and offline moves once their catch rate is published; overflow confirmation proposed to the analyst |
| High | 7–8 | Activate reserve or overflow, voluntary overtime, relax non-critical occupancy constraints, notify site leadership | Reversible at cost | Proposed; the analyst signs inside guardrails, a manager outside them |
| Severe | 9–10 | Declare under Event Management; continuity plans | Irreversible in practice | Person only |
The ladder makes explicit what the index page argues: the Elevated band is where the index earns its cost, because its actions are cheap, reversible and taken while service is still nominal.[3] Those are precisely the actions whose class can be released first.
What changes relative to threshold alerting
A threshold alert and a decision tier both produce a prompt to act. They differ in four ways that matter to the people operating them.
- An alert fires on a number; a decision answers a question with state. The alert cannot say which queue, which lever or how sure. The decision row says all three.
- False positives are handled per proposal, not by threshold tuning. Signal-detection theory makes raising a band threshold a trade of missed events for fewer false alarms, and on an exogenous index a missed event is a lost pre-positioning window.[6] A decision tier keeps the band where it is and lets the calibrated probability of each proposal decide what is executed, what is proposed and what is merely logged.
- The attention budget is spent on proposals, not readings. An analyst reviews rows that carry an action and a confidence, in the order of consequence, rather than a scrolling list of readings.
- Every prompt leaves a record that can be scored. Disposition review, which the index page already requires for tuning the bands, becomes the calibration set for the decision tier.
Measurement
Four measures tell an operation whether the pattern is working, and all four come from the proposal rows.
- Acceptance rate by class: the share of proposals the gate signed as proposed. Variance Harvesting names the same quantity its automation acceptance rate.
- Catch rate by class: the share of wrong proposals the gate caught before execution, the published figure on which release depends.
- Disposition precision by band: the share of Elevated-and-above proposals a post-hoc review judged correct, the measure the index page specifies for tuning.
- Calibration by question: the reliability curve of stated probability against observed outcome, by class, reviewed on a fixed cadence.
- Lead time realised: the interval between the signal version and the executed action, compared with the interval between the signal and the first endogenous alert. The difference is the window the pattern bought.
Where the typed model fits and where it does not
The bands are rules and stay rules. The arithmetic of hours moved stays in code. The reversibility of a lever is a property of its class and is a lookup. The typed model earns its place on two questions only: whether a narrative event reaches a particular queue or site, which is a reading task no rule can enumerate, and, where the repertoire is not cleanly indexed, which configuration fits. Those two are also the questions on which the model's known weaknesses bite. News text can steer the answer, so the state supplied is the contributing events and the queue's geography, not the whole feed; and the model reads literally, so the propositions are written positively and singly.[4]
The exemplar of the class, TypeSafe AI's Jev, reports latency of roughly 70 to 500 milliseconds and vendor-measured calibration; independent measurements did not exist at the time of writing, which is one more reason the calibration curve is computed locally before any class is released.[7]
Limitations
- Feed reliability binds harder here than on the index page. A decision tier that fans out questions on a stale signal version produces confident proposals about a world that has moved. The expiry field and the state version are the defences, and both must be enforced in the row format, not in policy.
- Rare events give few calibration points. The Severe band may produce a handful of proposals a year. Calibration there is unknowable for a long time, which is why the band stays person-only.
- Posture latency. If the posture feed lags the signal by minutes, the decision is made on a posture that no longer holds. The two feeds need comparable refresh rates.
- The pattern adds operational surface. Three services, a row format and a gate table are more to run than one alerting rule. For an operation at Level 2 the index page's advice applies: start with one category and one rule.
Maturity Model Position
- Level 3 — Progressive: signals are monitored on the daily ROC checklist; the six judgements and calculations are written down as questions the analyst answers; proposals and dispositions are logged. No tier is automated.
- Level 4 — Advanced (The Ecosystem Emerges): the index is scored and banded; the decision tier runs; propositions about reach are answered by a model and calibrated locally; Elevated-band break and offline classes become eligible for release, and release itself follows each class's published catch rate, not the level; posture and signal are versioned feeds.
- Level 5 — Pioneering: the repertoire is regenerated from dispositions; the calibration review retunes thresholds; the same decision tier serves the ROC, the forecasting loop's intake step and the placement engine.
See Also
- Workforce Decisioning
- Decision Primitives for Workforce Operations
- Supply-Side Routing
- Operational Volatility Index (OVIX)
- Resource Optimization Center (ROC)
- Daily ROC Routine
- Real Time Threshold Alerts and Escalation Protocols
- Intraday Management
- MAP-Elites
- Variance Harvesting
- Human Gates and Number Grades
- Event Management
- External Regressors in WFM Forecasting
References
- ↑ Endsley, M. R. (1995). "Toward a theory of situation awareness in dynamic systems". Human Factors 37 (1), 32–64. doi:10.1518/001872095779049543.
- ↑ 2.0 2.1 Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). "A model for types and levels of human interaction with automation". IEEE Transactions on Systems, Man, and Cybernetics – Part A 30 (3), 286–297. doi:10.1109/3468.844354.
- ↑ 3.0 3.1 3.2 Operational Volatility Index (OVIX), this wiki: design, action bands, and the "sensing plus repertoire" argument.
- ↑ 4.0 4.1 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. Third-party reference, not vendor documentation. https://gist.github.com/pjburnhill/adf8d28efcad9df037bfdece178ef965
- ↑ Human Gates and Number Grades, this wiki: the action gate, who signs, and the rule that a gate is released by published evidence.
- ↑ Swets, J. A. (1988). "Measuring the accuracy of diagnostic systems". Science 240 (4857), 1285–1293. doi:10.1126/science.3287615.
- ↑ TrueFoundry (2026). "TypeSafe AI's Jev: What 'System One Models' Actually Are". Third-party analysis; notes that benchmark, latency and calibration figures are vendor-reported. https://www.truefoundry.com/blog/typesafe-ai-jev
