Sentiment Decision Surface

This is the first implementation slice for the selected sentiment-analysis candidate. It is an interaction contract demonstration, not a publication of the underlying model or dataset.

The question

When should a three-class sentiment label be trusted, and when should the system route the case for review? The instrument exposes two explicit gates: normalized top-class confidence and the margin over the runner-up.

Flagship candidate / C-01

Sentiment decision surface

Evidence-gated · illustrative fixture

A first bounded instrument for the selected sentiment-analysis candidate. It makes a review policy inspectable without presenting synthetic scores as a trained-model result.

L2 / decision surface

When should a label be trusted?

Preparing deterministic worker

Adjust an illustrative three-class score vector and the two declared gates. The instrument makes ambiguity visible; it does not run BERT inference or report empirical performance.

Operator inputs

Score vector

Illustrative sentiment evidence scores

Positive evidence

0.72

0.00 — 1.00 evidence weightNeutral evidence

0.18

0.00 — 1.00 evidence weightNegative evidence

0.10

0.00 — 1.00 evidence weight

Decision policy

Review gates

Confidence and margin review gates

Confidence gate

50%

Minimum normalized top-class share · 34% — 95%Margin gate

10%

Minimum lead over the runner-up · 0% — 50%

Restore representative state

Deterministic output

Current reading

Decision clears gates

Top labelPositive72% normalized share

Confidence gate

cleared

Margin gate

cleared

Top / runner-up gap

54%

Positive

72%

Neutral

18%

Negative

10%

Positive clears both declared gates: 72% normalized confidence and a 54-point lead.

Static reference

The representative state is Positive 0.72 / Neutral 0.18 / Negative 0.10 with a 0.50 confidence gate and 0.10 margin gate. It clears both gates in the deterministic policy; these values are synthetic and are not a model result.

Contract: bounded inputs · fixed units · deterministic version `sentiment-decision/v1` · no visitor data or external fetches.

Evidence boundary

The current values are a deterministic synthetic fixture chosen to make the decision rule inspectable. They are not observations from a trained BERT model, not a performance result, and not a claim about the source repository.

The working source candidate is the public BAA sentiment-analysis repository, observed at a pinned commit during the T-4.3A intake. Its dataset provenance, privacy handling, license, model identity, validation, and reproduction package remain open review items. The repository is linked as an observation pointer; no raw dataset, model weight, or notebook output is copied into NAOBI.

What must happen before publication

• confirm that this source repository is the intended Sentiment Classification with BERT work;
• complete the research/experiment dossier and independent validation;
• clear dataset, code, and model rights and any review-text privacy boundary;
• replace the illustrative fixture only with approved, versioned evidence; and
• pass the ADR-0006 scorecard and Raihan’s explicit publication approval.