Every result here names the system it was measured on.

Because the claims register holds a percentage change for those rows rather than two published endpoints, and indexing to 100 at baseline is the accurate way to draw a change whose absolute values are not published. Inventing plausible absolute numbers to make an axis look concrete would be fabrication. The one row that does hold both endpoints, deployment success at Arlo Technologies, is drawn as real percentages instead.

Because a ceiling is not an achievement and must not be drawn as one. The 2% is the grounding failure threshold above which the CourtNetra build fails, so those cells are the whole budget rather than a shortfall: they are drawn solid inside an outlined field of a hundred, while every other row marks the gap between its measured rate and 100. Drawing the ceiling the same way as a level would put 98 marked cells on a row whose real figure is 2.

It cannot be compared directly, and this page does not try. The corpora, the query sets and the scoring rubrics all differ, so a rate measured on one commercial tool over its own content and a gate measured on CourtNetra over Indian judgments are two different measurements that happen to share a unit. What MetaMinds publishes is the gate on its own system: an automated hallucination evaluation runs weekly in CourtNetra's continuous integration and the build fails above 2% grounding failure. Independently published evaluations of commercial legal research tools do exist, and this page carried one until 16 September 2026. It came down because it has no row in this site's claims register yet, and a number without a row does not ship here even when somebody else measured it.

Because a benchmark scored by the vendor who commissioned it proves nothing about the vendor: every vendor that has published one has won it. MetaMinds publishes a gate on its own system instead, where a bad week fails a build rather than a slide: CourtNetra fails its build above 2% grounding failure in weekly CI. An independently published evaluation is the right thing to reach for instead, and it belongs in its own section, on its own axis, with its citation, its dates and its limits printed beside it, never drawn on one axis with our gate. This page carries none today, because the one it carried has no row in this site's claims register yet.

The register is a file in the repository this site is built from, docs/05-CLAIMS-REGISTER.md. It holds one row for every number and every named framework the site publishes, the source each was checked against, and the date it was checked, and every figure on this page prints the register section it traces to in the source line beneath it. It is why this page publishes one case study rather than a portfolio, why no client is named anywhere on the site, and why a chart came off this page rather than staying up with a footnote. Prices are gated automatically: a dollar figure rendered on any route with no row here fails the build. The rest are checked by review before a page ships, and the row behind any figure on this page can be sent to you if you ask through the contact form.

That the judgment is in the CourtNetra production database and can be found by a search, not that a model has read it. Retrieval is hybrid: dense similarity in self-hosted PostgreSQL with pgvector and tsvector BM25 keyword search, reconciled by reciprocal rank fusion, with embeddings from Qwen3-Embedding-4B on a self-hosted CUDA GPU pipeline. The vector index covers an embedded subset of the corpus rather than all 18,863,754 judgments. Which slice that subset is, is not published here, because the claims register does not yet state it and naming it without that would be a guess dressed as a specification.

Because the quantities span nearly six orders of magnitude, from 50 government web assets to 18,863,754 judgments, and on a linear axis every value except the largest renders as an invisible sliver, which reads as roughly nothing when the real figure is tens of thousands. It is drawn as positions rather than as lengths for a related reason: a length taken from a logarithm invites comparison by ratio, and that comparison is false. Each system is one dot of the same size on one axis whose decades are ticked and labeled from 10 to 10 million, so what you read off it is which decade a value falls in and how far past it, which is exactly what the number means.

Here is what is established by construction, and what is not. Every figure is server rendered SVG, so there is nothing to fetch and nothing to hydrate; each figure's measured geometry is its resting geometry; and the only motion on them declares a start frame and no end frame, so a browser with no scroll driven animation support, a reader who asked for reduced motion, a printed page and a crawler each receive the finished picture rather than an empty one. Every value, label and unit is real HTML text beside the picture rather than inside it, so the numbers survive with CSS off as well. What is not claimed here is a clean bill of health from a browser: an earlier version of this page shipped charts whose start frame was the absence of their own data, they rendered blank at one width while passing every automated gate, and a screenshot is what caught it. Those charts were rebuilt and that construction is not on this page.

One, in full technical detail: CourtNetra, MetaMinds' own system, retrieval over 18,863,754 Indian judgments. MetaMinds names no client, because none has given written permission. SaveLIFE Foundation, INNEFU Labs and Arlo Technologies, named in the figures on this page, are Aniruddh Atrey's prior employers, not MetaMinds clients.

Article 50 transparency obligations came into force on 2 August 2026. The Article 50(2) grace period for machine readable marking of synthetic content runs to 2 December 2026. Annex III high risk obligations were deferred to 2 December 2027 and Annex I to 2 August 2028 by Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force since 27 July 2026. The MetaMinds EU AI Act page carries the full timeline and what each date reaches.

A digital product between $99 and $890, bought on its own page in a single payment, with the file delivered on purchase. The next rung is a fixed-scope engagement from $2,500, where a deposit reserves a named start week and is refunded in full if MetaMinds declines the work or the scope turns out materially different from what was described.