From a8288810433800be15c0e410355c4f9149510cd5 Mon Sep 17 00:00:00 2001 From: Denis_Drobyshev Date: Fri, 18 Sep 2026 23:44:45 +0300 Subject: [PATCH] Correct the counts, and stop calling a GPU measurement a guarantee "Thirty-one algorithms and twenty-two environments" is the fifth place today carrying that stale pair, and the only one written in words, which is how it survived every correction made to the digits. The retrieval sentence said quality is "pinned by an eval". The offline column of that table is pinned, by a test that re-runs it; recall@5 0.92 and MRR 0.94 are measured on a GPU that CI does not have. Both are true and they are not the same claim, and this organisation's whole argument is that the difference matters. --- profile/README.md | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/profile/README.md b/profile/README.md index 280c7ff..9be5bf4 100644 --- a/profile/README.md +++ b/profile/README.md @@ -40,8 +40,10 @@ are transcriptions from Wikisource, marked in the corpus as pending a check against the official publication at pravo.gov.ru; the parser for that source is how an edition gets confirmed rather than how a code gets added. Judicial practice is the next pipeline rather than something already shipped. -Retrieval quality is pinned by an eval on a golden set: recall@5 0.92, MRR 0.94 -on the full corpus, tracked run over run. +Retrieval quality is measured by an eval on an eighteen-question golden set. The +offline figures are pinned by a test that re-runs it; recall@5 0.92 and MRR 0.94 +on the full corpus are measured on a GPU that CI does not have, and are stated +here as measurements rather than as guarantees. ```bash docker compose up app # → http://localhost:8077, no keys required @@ -93,7 +95,7 @@ LangGraph. If you want a small, transparent loop you fully understand, use this. Pricing, inventory, energy, queues and supply chains: the decisions a business makes thousands of times a day, where a small policy improvement compounds. -Thirty-one algorithms and twenty-two environments — nine of them applied — typed +Thirty-two algorithms and twenty-four environments — nine of them applied — typed and tested, aimed at problems that have a cost function rather than a leaderboard.