v1.1.0 — the dataset lane can no longer mislabel silently

1.1.0 — 2026-08-20

Dogfood swarm: health pass (bugs, proactive, humanization) plus a pre-Phase-9
wave. 39 tests → 88. Two CRITICALs closed, both of the same shape — output a
downstream consumer would trust, wrong, with no signal.

Fixed — the dataset lane could silently mislabel training data

• Sidecar collisions are refused, not merged. Sidecar paths derived from the
  image stem alone, so img.png and img.jpg in one folder both claimed
  img.txt. The second was reported as skip (exists) and the run exited 0 — a
  green run in which one image carried a description of a different image. It
  needed no flags. Colliding batches are now refused before the model loads,
  naming the offenders. Sidecars are never renamed to dodge a clash: trainers
  pair by exact stem, so a rename would orphan the caption.
• OCR can no longer present invented text as extracted text. Florence-2
  emits a decoded string for every image, including images with no text — a
  photograph returns '2', lexically indistinguishable from a correct reading.
  Results now always carry absence_of_text_unreliable (MCP) or an
  [OCR_CAVEAT] line (CLI). The text is never suppressed, emptied, or
  length-thresholded, because a short reading may be genuine.
• Sidecar writes are atomic — temp file plus os.replace, so an interrupt
  cannot leave a partial caption at the final path. An existing but empty
  sidecar is treated as unfinished and re-captioned.
• A mid-batch failure no longer aborts the run. The batch loop caught only
  FileNotFoundError/ValueError, so a CUDA OOM killed the job and swallowed
  the JSON summary. It now records the failure and continues; exit 3 on partial.

Fixed — broken promises

• PLAIN_SIGHT_MODEL_ID was documented in nine READMEs and honoured by the MCP
  server, while the CLI silently ignored it. Resolved once, at module scope.
• PLAIN_SIGHT_LOG_LEVEL was named in the CLI's own error hint and had no
  effect there — DEBUG was unreachable. Shared configure_logging() now serves
  both surfaces.
• PLAIN_SIGHT_EAGER_LOAD is honoured on the MCP surface again, and a failure
  during eager load no longer kills the server import: it surfaces via
  sight_status and as a ToolError on first use.
• The CLI emitted a raw traceback at exit 1 when engine construction failed —
  the construction sat outside the error boundary. Now a structured line, exit 2.

Added — provenance

• The model revision is pinned to 4271c66b88cdbc05735372ec13b2360108de5317.
  Unpinned, HuggingFace resolves to whatever the default branch points at, so a
  silent retag would change captions under unchanged inputs.
  PLAIN_SIGHT_MODEL_REVISION overrides.
• Every output payload names the weights — model_id and the resolved
  revision on describe_image, read_text, describe_batch, sight_selftest,
  the CLI --json modes and the batch summary. sight_status reports requested
  and resolved separately, so a mismatch is visible.
• --manifest PATH writes an opt-in run record: versions, model, both
  revisions, device, dtype, tier, prefix/suffix, per-image results. Never
  inferred; a path colliding with a sidecar is refused.

Added — it now says what it is doing

• The load is announced before work begins, at default verbosity, with the
  count of images that will actually be captioned — so the pause never appears
  mid-run after a stretch of skips.
• Progress heartbeat every 25 items or 30s: written / skipped / failed,
  rate, ETA. Per-skip lines are gone; a re-run over a finished set is quiet.
  Failures stay one line each.
• --dry-run prints the whole plan — model, revision, counts, collisions —
  loading nothing and writing nothing.
• --help carries the exit-code table, the stderr/stdout split, and the
  first-load cost. Every flag on every subcommand has help text.

Changed

• CI and verify.sh select tests by marker (-m "not dogfood") rather than by
  filename, so a new CI-safe test file is covered without touching CI.
• --help and CLI error output are ASCII; the em-dash separator mojibaked when
  stderr was a cp1252 pipe on Windows.
• pythonpath = ["."] so the console script and python -m pytest agree.
• status() reports device-scoped VRAM, omitted on CPU.

Notes

Severities throughout this release were assigned by a cross-family panel of
pinned non-Claude model seats with authorship stripped, not by the authors of
the findings. Several moved in both directions.

plain-sight v1.0.0

An AI says what it sees. First shipped release: Florence-2 describer (MCP server + CLI), three detail tiers, OCR, LoRA-dataset caption sidecars with exact basename pairing and deterministic decoding. Model pinned to florence-community/Florence-2-large (MIT, native transformers — trust_remote_code never used). Shipcheck audit 100%; landing page + handbook: https://mcp-tool-shop-org.github.io/plain-sight/ — full details in CHANGELOG.md.