Locks
Every item here is enforced in the pipeline rather than written in a policy document. For a business whose entire product is credibility, that distinction is the difference between a defence and an allegation.
The pre-registration lock is the one that answers 2024 directly. The accusation was that results were tuned to a client's preference. That is unfalsifiable if the weighting scheme and turnout model are chosen after seeing raw numbers, and trivially falsifiable if they were sealed and hashed before the first call. It is a cryptographic answer to a charge that currently has no answer at all.
Sealed, not merely embargoed.
Section 126A of the Representation of the People Act bars publication between the start of the first phase and half an hour after the last. Here the aggregate is not computable before that moment. Interviews stream in, the tabulation job does not run, and no operator account holds a key that changes it.
| Questionnaire, 17 items | 14 Aug 09:12 |
| Raking margins and sources | 14 Aug 09:12 |
| Weight trimming thresholds | 14 Aug 09:12 |
| Turnout model specification | 14 Aug 09:12 |
| Seat simulation parameters | 14 Aug 09:12 |
| Quarantine and exclusion rules | 14 Aug 09:12 |
No caste question is fielded.
No field exists in the schema. Volunteered utterances are dropped at ingest. The methodological case stands on its own: phone-frame caste cells are the worst-covered in the sample, so caste cross-tabs would be the least reliable numbers we produce.
No response joins an electoral roll.
Contact keys are sealed at ingest. Vote intention is stored against a study identifier that does not join to the contact record. There is no path from a released dataset back to a person.
No target list, no persuadability score.
The export button on Panel CRM is disabled at build time, not by a permission. This is a measurement instrument, and a research product that quietly doubles as a targeting product is one leak away from ending the research business.
Cohorts below 100 effective are suppressed.
Not shown small, not shown with a warning. Suppressed. Below that floor a cohort stops being a segment and starts describing identifiable people.
The sponsor is named on every call.
Stated in the opening line on every one of two lakh recordings. When a rival alleges the poll was a party operation, that is the artefact that answers it.
Weights cannot change after unblinding.
The pipeline will not accept a modified specification against a wave in progress. An analyst wanting a different scheme runs it as a declared, separately labelled sensitivity analysis.
The caste lock is the one he will push back on. His published Karnataka work leans on Lingayat, Vokkaliga and OBC breakdowns and his audience expects them. Two honest resolutions exist: community analysis stays on his existing field arm and never touches this pipeline, or the product is positioned as the verifiable core with community work explicitly outside it. Both are workable. Blurring them is not.