What the measurements mean, and what we claim.
This page describes what Meyppi measures, how each verdict follows from the evidence, and how you can re-check any finding. Calibration parameters and scoring thresholds stay private; everything that appears in a report is defined below.
The six verdicts
Every expected placement receives exactly one verdict. The vocabulary is fixed across every report, export and page on this site.
- Detected — Full. The creative aired in the published episode and the evidence covers its complete contracted runtime.
- Detected — Partial. The creative aired, but a measurable part of its runtime is missing (for example, a trimmed tail).
- Version Mismatch. A version of the creative aired, but the evidence attributes it to a different contracted version than the one expected for that placement.
- Not Detected. No matching evidence found in the published episode.
- Insufficient Signal. The published media's quality is too low, or the evidence too sparse, for a defensible result either way.
- Manual Review Required. The evidence is ambiguous between explanations (for example, additive noise versus a cut); an analyst examines it and the report states what was found.
How a verdict is reached
Creatives and published episodes are fingerprinted: each file is reduced to a constellation of spectral landmark pairs that survives compression, loudness processing and background noise. A true occurrence of a creative inside an episode appears as thousands of landmark pairs agreeing on one consistent time offset; interference, cross-talk and music beds do not produce that signature.
From the agreed matches, the engine reconstructs the creative's timeline inside the episode and measures four things: how much of the creative's runtime the matches cover, whether the start or the end is missing and by how much, whether there are gaps in the middle, and, when several versions of a creative exist, which version's evidence outranks the others. Structural findings are judged against the creative's own self-match profile, so a naturally quiet creative is never mistaken for a partially delivered one.
Verdicts follow from these measurements by fixed policy, calibrated once on labelled corpora. Ambiguous evidence routes to Manual Review Required rather than to an unsupported accusation, and the report states the reason in each such case.
Evidence fields
Every placement's verdict ships with the measurements behind it. The same fields appear in the PDF report, the CSV/JSON exports and the raw evidence file.
| Field | Meaning |
|---|---|
| offset_sec | Where the match starts inside the published episode. |
| covered_time_sec | How much of the creative's runtime the matches cover. |
| coverage_ratio | Coverage relative to the creative's own self-match profile. |
| missing_head_sec / missing_tail_sec | Missing runtime at the start or end of the creative, in excess of the self-match profile. |
| largest_interior_gap_sec | The largest uncovered stretch inside the matched span. |
| match_count | Number of aligned landmark pairs supporting the match. |
| confidence | Policy confidence for the verdict, from 0 to 1. It is not a probability of legal admissibility. |
| evidence_tier | How the media was obtained: Tier 1 independently captured, Tier 2 independently retrieved from the publisher's public distribution, Tier 3 customer-supplied, Tier 4 unverified. Claim strength in the report follows the tier. |
Re-checking a finding
Audits are reproducible: the same creative files and the same published media produce the same report. Every media file entering an audit is hashed and its source recorded, so any placement's evidence can be traced to a specific file from a specific source and re-processed. Raw verification events, the per-match records behind every verdict, are exported alongside the report and are never discarded after classification.
Known limits
The engine's calibration is validated on labelled synthetic corpora; real-world performance is being established with each engagement. Verification requires the published media to be independently observable: dynamically inserted placements that vary per listener often cannot be verified from a public feed, and host-read segments are outside current scope. Video podcasts are matched on their audio track; FM radio audits run as a pilot with clock-aligned recordings and time-band reconciliation. Findings are algorithmic evidence, not legal proof.