v1.1.1

A patch to the citation check of 1.1.0, whose README promised that figures in
words are checked, and whose parser read some of them wrong.

Fixed

• A figure in words is read whole. 1.1.0 read one point five billion dollars as five billion, so a passage printing $5 billion backed it; it did
  not read two and a half billion at all, and split two thousand five hundred dollars into two figures. Words now compose (hundred, scale words
  stepping down or multiplying, point, a half), and words that do not
  compose are a figure with no value, which backs nothing and nothing backs.
• The no_network test fixture also refuses socket.connect_ex.

Added

• Two positive controls in docs/eval/report-v1.1-citations.md, not
  pre-registered. A reference number put in place of a cited figure is
  accepted by v1.0.0 in 88/88 and withheld by the M1 rules in 55/88, so the
  M1 zero bounds support that comes only from a reference number. Figures
  rewritten in words are counted by M2 in 85/85.
• A "Deviations from the protocol" section in the same report: the
  percentage rule and the 3-digit rescale floor came from issue #10, not from
  the protocol's scope, and each alone changes 0 of 139 outcomes.

Changed

• The rescale rule is named for what it counts: 3 digits, trailing zeros
  included, not 3 significant digits. The README says so, and lists the
  round-count and (in percent) cases among the limitations.
• The report's arm tables collapse arms that admit the same cases.

The counts of M1, M2 and M3 are unchanged.

Full report: docs/eval/report-v1.1-citations.md.

v1.1.0

v1.1 changes the citation check, not the gate. The headline numbers of
report-v1.md were measured with v1.0.0 and stand.

Fixed

• The citation check no longer counts the item label, or the passage's
  references to its own notes, pages, items, exhibits and sections, as support
  for a figure. A percentage is backed only by a percentage, and a number the
  passage prints without a scale word is rescaled only when it has at least 3
  significant digits (#10).
• A figure in words that names a scale, percent or dollars now needs a
  marker and is held to the cited passage like one in digits
  (#10).
• An unsupported_claim abstention names the sentence and its markers, never
  the figure it withheld (#11).

Added

• docs/eval/protocol-v1.1.md, committed before the measurement, and
  docs/eval/report-v1.1-citations.md, rebuilt by make eval: the v1.0.0 and
  v1.1 checks replayed over the 139 v1 generations with text, offline. No
  outcome changes in any arm (0/117 answered cases under arm A, exploratory,
  Wilson 95% upper bound 3.2%); a positive control that changes one digit per
  answer is withheld in 102/105
  (#13).

Changed

• The no_network test fixture also refuses httpx's async transport and
  socket.connect to an internet address.
• astral-sh/uv image 0.12.23, hypothesis 6.168.4.

Full report: docs/eval/report-v1.1-citations.md, measured under docs/eval/protocol-v1.1.md, committed first.

v1.0.0

First public release.

Added

• A FastAPI service that answers questions about one SEC 10-K at a time,
  cites the passages it used, and abstains with a closed set of reasons when
  the filing does not support an answer.
• Exact NumPy search over a sharded on-disk index, Ollama for embeddings and
  generation, a nonce-delimited prompt that treats filing text as untrusted,
  and a citation check on every answer.
• An evaluation harness that runs the production pipeline: a golden set built
  from XBRL, a pre-registered protocol, six gate arms as masks over one
  generation, and a CI gate that replays the dev split from a recorded tape.
• The pre-registered headline round (qwen3:32b, 20 companies): H1 not
  attainable, H2 not met; the brier relevance model was removed and the
  default gate is none (ADR-0014).