Card: seed probes are five proverbs; attestation stated per probe

README.md

@@ -73,7 +73,7 @@ Fluency is not alignment: a model can achieve native-level fluency in a language
 
 ## Limitations (v0.1)
 
-Seed probe set of five attested proverbs (expanded set with dialect metadata in v0.2). `runs_v0.jsonl` holds the July and August 2026 evidence: the three chat-interface runs on the flagship probe (ilu-001), hand-scored on every rubric axis, and the first API runs over the seed set. `runs_v0.1.5.jsonl` holds a full 5 model × 5 probe matrix on current frontier models (September 2026), elicited with the open runner. On the API rows, dimension 1 is machine-scored and axes 2–3 are machine-preliminary on the live leaderboard (LLM judge, labelled as such); axes 4–5 are marked pending until native-speaker judges score them, which ships as v1. The finding was first observed and documented May 2026 in our internal experiment logs; this release establishes the public protocol.
+Seed probe set of five proverbs (expanded set with dialect metadata in v0.2). Attestation is tracked per probe on the live IlùBench page, and as of 2026-09-15 none of the five has cleared the native-speaker panel quorum: ilu-001 is documented in two independent written sources, ilu-002 and ilu-003 in one each, and ilu-004 and ilu-005 in none yet. The `status` values `attested_seed` and `attested_v0_demo_pair` in `probe_set_v0.jsonl` are v0 release labels for the seed set and the demo pair, kept unchanged so existing loaders keep working; they are not an attestation verdict. `runs_v0.jsonl` holds the July and August 2026 evidence: the three chat-interface runs on the flagship probe (ilu-001), hand-scored on every rubric axis, and the first API runs over the seed set. `runs_v0.1.5.jsonl` holds a full 5 model × 5 probe matrix on current frontier models (September 2026), elicited with the open runner. On the API rows, dimension 1 is machine-scored and axes 2–3 are machine-preliminary on the live leaderboard (LLM judge, labelled as such); axes 4–5 are marked pending until native-speaker judges score them, which ships as v1. The finding was first observed and documented May 2026 in our internal experiment logs; this release establishes the public protocol.
 
 ## Versioning