{
 "_doc": "Sentences of the lab's stated limits for one result, copied word for word, each with its character offsets in the field and the field's own sha256. Sentences that carry a machine or estate path, or the colliding former project name, are not here.",
 "id": "inference-limits:zero-sorry-corpus",
 "source": {
  "lab": "interconnect lab",
  "commit": "9d619007982bf311b1d3b64c929b835bbceb1f97",
  "file": "top40.json",
  "sha256": "47b9a2640faab62a80449df918f744d13343ec79129c8710bb3e98a1a2cf770a",
  "bytes": 300602,
  "entry_id": "zero-sorry-corpus",
  "field": "limits",
  "field_sha256": "df870dcdf56c0c6276eb5483b33b699b0ce402f71fc8f729ba5939aa18486cf6",
  "field_chars": 3155
 },
 "stretches": [
  {
   "text": "AST-F5-02's own registry status reads 'not built in Pass A or Pass B' — the transport corpus is COUNTED, NOT COMPILED. 489 theorems (28%) are orphans of the marquee enumeration (480 before the S07 merge added nine unenumerated ones).",
   "start": 0,
   "end": 233
  },
  {
   "text": "CLM-F5-31 names four non-substitutable instruments (1,717 regex / 1,544 elaborator / 1,851 / 1,725; as CLM-F5-31 recorded them, not re-measured since; at S07 the orphan audit's regex count is 1,726) under a NO-SUM rule; they count different things. serving_limits moved 41->42 files and 408->430 theorems this week because of EssentialBits.lean.",
   "start": 234,
   "end": 579
  },
  {
   "text": "A3 2026-09-09: the registered command was a raw-line grep and it returned 4.",
   "start": 580,
   "end": 656
  },
  {
   "text": "A scan that counts its own documentation can never be made strict, so the entry published a 4 it then had to explain away. scripts/check_lean_hygiene.py strips /- -/ (nesting included) and -- before scanning, reports 0 in code, exits 1 on any hit, and carries a --selftest that plants a sorry and requires the scan to fire.",
   "start": 942,
   "end": 1265
  },
  {
   "text": "Second correction from the same pass: the declaration count is 1,852, not 1,870 -- 18 comment lines beginning with theorem or lemma were being counted, including two commented-out `theorem check : @Original = @Derived := rfl` lines in InjectivityFloorCheck.lean and TTEPreludeCheck.lean.",
   "start": 1266,
   "end": 1553
  },
  {
   "text": "STILL OPEN, per the grader: this is a textual scan of tracked sources, not a clean elaboration of every corpus; the stronger corpus-wide compilation claim needs the builds.",
   "start": 1554,
   "end": 1726
  },
  {
   "text": "The old pure-shell command hid that, because check_top40_commands.py skips commands that name no script.",
   "start": 2031,
   "end": 2135
  },
  {
   "text": "(2) The scan's KNOWN BLIND SPOT (S02) is closed: the HIT pattern was `^ *sorry|^ *axiom |by +native_decide`, which missed an inline `:= by sorry`, `(by sorry)`, an attributed or modified `axiom` and a `native_decide` not preceded by `by `.",
   "start": 2653,
   "end": 2892
  },
  {
   "text": "The pattern is now word-bounded, and --selftest plants each of those cases and requires a hit (measured: the old pattern FAILS the new selftest).",
   "start": 2893,
   "end": 3038
  },
  {
   "text": "Re-measured before the change: 0 hits under both patterns on this corpus, so the claim held; the instrument did not.",
   "start": 3039,
   "end": 3155
  }
 ]
}