Methodology

UnderTone identifies a reproducible workplace request in a message, then computes two independent scores for it on a 0–10 scale.

Surface strength

How strongly the request itself is phrased — its directness, modal wording, hedges, and optionality formulas like “no rush” or “whenever you get a chance.”

Communicative force

How strongly observable, non-surface evidence in the surrounding language makes action expected — deadlines, consequences, dependencies, accountability, and repetition.

Pragmatic gap

The difference between the two: force − surface. A large positive gap means the surrounding context carries more pressure than the request’s phrasing makes immediately visible. A large negative gap means the phrasing carries more force than the surrounding context independently supports.

Confidence

A separate reliability estimate for the analysis itself — how internally consistent and unambiguous the detected evidence is. It describes the rule-based analysis, not how certain the sender was or how likely the recipient is to comply.

Every score traces back to exact character spans in the original message. Surface and communicative force share only one early step: a first pass identifies the request itself. From there the two scores are computed independently. Surface strength is scored from the request’s own visible wording. Before force scoring begins, every surface-visible span in the message is masked out — the force scorer never receives the surface score, the request strategy, the mood or modal wording, or any surface softening. It sees only the masked message and masked prior context, so its evidence can never be surface material in disguise.

The engine is deterministic and rule-based — no language model, no network calls, nothing outside the function’s own arguments. The same message and configuration always produce the same scores and the same evidence trace.

• Infer a sender’s intent — what they meant, wanted, knew, or believed.
• Predict how a recipient will actually behave.
• Detect deception, personality, or emotion.
• Claim calibrated accuracy against human perception. The 0–10 scales are an engineering convenience, not a validated psychometric instrument.

UnderTone’s internal release suite currently reports:

These are specification tests measuring implementation fidelity and internal consistency — whether the code behaves the way the specification says it should — not external human validation. They confirm, for example, that softening a request’s wording moves surface without moving force, and that adding an independent deadline moves force without changing the request’s wording. They do not confirm that a human reader would assign the same numbers.

The one PARTIAL result concerns CCSARP request-strategy detection: levels 1–7 of the nine-level directness scale are fully implemented and tested, while levels 8 and 9 (conventionally indirect hints, which require reasoning about prior thread context) are not yet detected and are excluded from scoring rather than guessed at.