Net Conduct Score
Net Conduct Score (NCS) is the human-experience input to a Conduct Quotient. One eligible worker who directly participated in a verified interaction answers one bounded question about how the customer treated people.
PLAIN-LANGUAGE SUMMARY
The Net Conduct Score in plain language
NCS is a single-item measure. That is a deliberate trade and it carries known costs: reliability cannot be estimated from one item the way it can from a scale, and collapsing responses discards information. The question wording, the response weights, and the decision to use one item are all research candidates rather than settled choices.
This page carries the normative requirements. The explanatory NCS page covers the same ground for a wider audience. Where the two differ, this page governs, and the working specification governs over both.
7.1
The question and the response set
One question, four scored responses, and one unscored abstention. The wording is normative and must be identical everywhere it appears.
CANDIDATE
The lead prompt is: “Based only on how this customer interacted with you and others, would you want to serve them again?”
The four scored choices are Definitely yes, Probably yes, Probably not, and Definitely not. A separate Not enough interaction to say action records an unscored abstention and MUST remain distinguishable from a nonresponse.
CANDIDATE
The lead graduated model assigns +10, +3, −3, and −10 response points respectively. The 10:3 ratio is a hypothesis about conviction, not a measured quantity.
7.2
Who may answer, and about what
A response counts only when the rater was present and the interaction was real.
REQUIRED FOR COHERENCE
Every response MUST be linked to one verified interaction, and the rater MUST have participated directly with sufficient contact to answer.
CONSTITUTIONAL
A rater MUST consider only conduct they directly experienced or observed during that interaction. Reports, summaries, internal notes, and other ratings are ineligible.
CONSTITUTIONAL
The worker MUST NOT see the customer’s existing NCS, VCI, CQ, tier, or benefit eligibility before responding.
A response MUST reflect willingness to serve the person again based only on how they treated people. Spending, status, tipping, profitability, service complexity, and complaint activity are all ineligible as inputs to the judgment.
CONSTITUTIONAL
Every response MUST be a deliberate act by a person who participated in the interaction. A response MUST NOT be generated, derived, inferred, pre-filled, defaulted, ranked, or suggested by any automated system. Facial analysis, voice and speech analysis, tone, prosody, word choice, sentiment classification, affect or emotion detection, biometric signals, video, and audio are prohibited as inputs to a response, whether used to produce one or to recommend one to a worker.
The prohibition covers recommendation as well as generation, because a worker confirming a suggested answer produces an inferred result recorded as a human judgment. A system that displays a proposed response, orders the options by predicted likelihood, or marks one as expected is producing the response rather than collecting it.
CONSTITUTIONAL
A response, and any record derived from one, MUST NOT be used to train, fine-tune, evaluate, or calibrate a model that infers emotion, affect, personality, or conduct about any person. This restriction travels with the data and binds every downstream recipient.
7.3
Worker protection
These requirements exist because the instrument is answered by someone whose employer controls their schedule and pay.
CONSTITUTIONAL
Answering is never mandatory. No business may require a response, make one a condition of the role, or treat declining as a performance matter. A worker may abstain or not respond without penalty, explanation, lost work, reduced pay, adverse scheduling, or performance effect.
CONSTITUTIONAL
Capture, required training, review participation, and appeal participation are work, and MUST be compensated as work in every jurisdiction, not only where local law compels it. A business MUST NOT rely on its own judgment of what compensation is appropriate.
CONSTITUTIONAL
No manager and no workforce system may receive individual response values, completion records, customer outcomes, or rater-level distributions. This applies to any function inside the business, however it is named, including analytics, insight, and customer-experience teams.
CONSTITUTIONAL
NCS MUST NOT be used to score employee performance, set compensation, or impose rating quotas, and MUST NOT replace any emergency, safety, fraud, harassment, or incident-reporting process.
Whether a worker can meaningfully decline a task their employer scheduled is not settled by any of the above. It is an open question for labor review, and it gates a pilot.
7.4
Sampling
Selective prompting can distort a result more than the scoring model does.
CANDIDATE
A business MUST use a declared, auditable sampling policy and MUST NOT select customers for rating based on status, complaint activity, protected traits, or expected response.
Eligibility, selection, prompt delivery, nonresponse, and abstention MUST remain distinguishable states. A delivered prompt is not equivalent to a completed observation, and missingness has to be modeled rather than eliminated by pressure.
REQUIRED FOR COHERENCE
Multiple responses from one interaction form one statistical cluster. Their combined weight MUST be capped, and they MUST NOT increase interaction count, issuer diversity, or independence as though they were separate events.
7.5
Aggregation and output
The provisional baseline calculation rescales the weighted mean of response points to a −100 to +100 range.
PROVISIONAL RESEARCH FAMILY
Raw NCS = 10 × Σ(wᵢ × pᵢ) ÷ Σwᵢ where pᵢ ∈ {+10, +3, −3, −10}
A history composed entirely of Definitely yes produces +100. Entirely Probably yes produces +30, entirely Probably not −30, and entirely Definitely not −100. No coefficient here is approved for production.
A single average is not sufficient output. The component MUST also supply the rate of each response type, the unscored abstention rate, eligible response count and effective weight, a confidence or uncertainty measure, recency distribution, context distribution, issuer and interaction concentration, active dispute and invalidation rates, and the schema and calculation versions in use.
CONSTITUTIONAL
NCS MUST report insufficient history until a validated minimum evidence threshold is met. A new participant does not begin at a neutral zero that could be mistaken for an established record.
UNRESOLVED
That threshold does not exist. No number of responses has been established as sufficient, and no reliability estimate, inter-rater agreement figure, or standard error has been produced for the 10:3 model or any alternative. Reliability cannot be estimated from a single item the way it can from a scale, so producing one requires a different study design rather than more of the same data.
UNRESOLVED
NCS ranges from −100 to +100. Conduct synthesis takes its inputs on an internal 0 to 100 scale, which VCI already uses. No transform between the two has been published, and the choice of transform changes every resulting score. Until it is specified, the worked examples in the Conduct Quotient specification are illustrative of structure rather than computable.
Time decay, cross-context weighting, pattern recognition, business-failure handling, and the final balance against VCI belong to conduct synthesis rather than to NCS. This component preserves timestamps and context instead of resolving them.
7.6
Consumer visibility and challenge
Both negative responses MUST carry one additional structured conduct category, selected from an approved set, so that eligibility review, bias analysis, and appeal have something to examine. The category does not change numerical severity.
CONSTITUTIONAL
The response set is the whole instrument. A rater MUST NOT be given a free-text field describing the customer, and no narrative account of a person may be captured, stored, transmitted, or scored. Where an approved category requires a short factual note, that note MUST NOT enter any calculation, MUST NOT reach a receiving business, and is available only through protected review.
This requirement exists because the durable harm in existing systems is usually the sentence rather than the score. A permanent free-text note asserting what kind of person someone is cannot be corrected by aggregation, ages out of nothing, and follows the person indefinitely. A structured category can be counted, audited for bias, and appealed against a defined set.
CONSTITUTIONAL
A consumer MUST be able to see what affected them, correct a source fact, add context, report bias or selective sampling, and reach independent review, without learning the identity of an individual worker or being placed in direct dispute with them.
Disclosure timing has to protect the worker as well. Where a location has few eligible responders, or a customer visited once, delay and grouping alone do not prevent identification. In a three-person shift, abstaining is itself identifying.
UNRESOLVED
Capture requires a minimum number of eligible responders at a location. Setting and validating that minimum is a precondition of any capture anywhere, and it has not been done. No location can be shown to clear a threshold that does not exist, so no capture is permitted under this standard today.
7.7
Open questions
Does the question measure willingness to serve again for the intended reason? Construct validity for a single item has not been established, and the wording could be read as a judgment of character rather than of conduct in one interaction.
Do the four labels communicate strength consistently across cultures, languages, and roles? The 10:3 ratio assumes “definitely” carries a stable multiple of “probably.” That assumption is untested.
Is the abstention understood and safe to use? If abstaining is read as a soft negative, or is itself identifying in a small team, the option fails.
What minimum evidence is required before any summary is available? Unresolved, and it determines whether early results are stable enough to act on.
What minimum number of eligible responders must a location have before capture may occur there at all? Currently unspecified.
Can a worker decline in practice? Requires labor review rather than a design answer.
CHANGES
Release history
0.5.1 — 23 August 2026. Added three constitutional requirements to section 7.2. A response must be a deliberate act by a person who participated; no automated system may generate, infer, pre-fill, rank, or suggest one; facial, voice, tone, prosody, sentiment, affect, and biometric signals are prohibited as inputs; and responses may not train a model that infers emotion, affect, personality, or conduct. Recorded as D-590 through D-592 in the decision register.
The prohibition was previously stated only on the explanatory page, which does not govern. Adding it here is a structural amendment and carries a new release identity.