Pulse Library — JanPulse

Step 01 · Coverage report

What the corpus already answers

A proposed 12-item, 800-interview study, priced against the evidence that already exists

52% of the proposed instrument is already covered by prior fieldwork, at least in part

Covered

4

Items with recent, adequately sized evidence. These do not need to be re-fielded.

Thin

5

Evidence exists but is old, small, or both. Re-field, or quote with the limitation stated.

Gap

3

Never asked. These are the only items that justify a full fieldwork budget.

Item	Statement	Prior n	Evidence age	Status

Status thresholds: covered = n ≥ 250 and < 6 months old; thin = n 90–249 or older than 6 months; gap = no prior evidence. The weighting behind the headline percentage is published alongside the dataset, not hidden in a slide.

Step 02 · Consensus and divergence

What this population agrees on, and what divides it

Two lists, computed with Laplace-smoothed proportions and a one-proportion z-test at 90% confidence. Because the two probabilities sum to at most one, a statement cannot appear in both lists — that is a proof, not a filter.

Each statement is then tested again inside every opinion group. A statement carried by one large bloc is not consensus, however large its overall majority; only statements that clear the same threshold in both groups are marked as bridging.

Agreement

Divergence

“Divergence” here means majority disagreement with the statement, not division between groups. Group-by-group splits are in step 03.

Step 03 · Opinion structure

Two opinion groups, not one average

Two-level clustering over a sparsity-weighted PCA projection, with the group count chosen by silhouette

Where the groups part company

Polarising score = |correlation| between an item’s answers and a respondent’s position on the main axis of opinion. It is a property of the statement, not of a group; near 0 means the item does not separate anyone.

Step 04 · The citation trail

Every claim, back to the response

The part most research tools leave out

What a real answer carries

Evidence references — study, item, response and turn identifiers, so a reviewer can open the transcript.
Consent scope — whether the record may be reused in pooled analysis, only in the client deliverable, or not at all.
Reliability — the inter-rater ceiling and the scorer’s measured agreement on held-out data, with sample size. If a scorer has no measured figure, any claim resting on it is marked thin, not supported.
Coverage — how many records matched, how many were excluded by consent or access, and what the corpus could not answer.

What is synthetic here

The numbers on this page are illustrative. The method is not.

Synthetic

The 360 respondents and every percentage and count derived from them.
The three quoted responses, including the Marathi and Hindi text.
The prior-wave sizes and evidence ages behind the coverage table.

Real

The statistics: Laplace-smoothed proportions, one-proportion z-tests at z₉₀ = 1.2816, silhouette-based group selection, two-level k-means on a sparsity-weighted projection.
The disjointness property of the consensus lists.
The citation structure and the reliability fields, which exist in the data model now and are populated by measured figures when real data lands.

Limits we state up front

Below roughly 30 respondents per cohort the clustering layer is noise; such cells are suppressed, not published.
The main axis has an arbitrary sign and near-ties in variance. Direction is never presented as meaning — only relative position.
Consensus describes agreement with statements. It says nothing about whether a policy is good, and we do not let a report imply otherwise.
Everything here is conditional on the instrument. A badly worded item produces a clean, significant, meaningless consensus — which is why the instrument review gate exists.

Commission a study with the library behind it How we measure