skimmer v0.1.106

Changed

• macOS binaries (universal and per-architecture) are signed with a
  Developer ID certificate and notarized by Apple, so they run without a
  Gatekeeper warning however they're downloaded, not only through Homebrew.

skimmer v0.1.105

Added

• Best practices, from hands-on runs and the evals: a README section (with
  measurements for code search, web pages, long docs, search results and
  unanswerable goals) and a short guide in skimmer --help, so agents pick
  the right flags for each kind of search. The "For agents" snippet uses them.

skimmer v0.1.104

Changed

• Homebrew installs skimmer as a cask holding the release's prebuilt binary,
  not a formula built from source. On macOS Tahoe Homebrew refuses any
  source build without the Xcode Command Line Tools, so
  brew install dorkitude/tap/skimmer failed on Macs without them. The cask
  needs neither Go nor the Command Line Tools, and it clears macOS's
  quarantine flag so the unnotarized binary runs without a warning. If you
  installed the formula, switch once:
  brew uninstall skimmer && brew install dorkitude/tap/skimmer.

Added

• A universal macOS binary in every release (skimmer_X_darwin_all.tar.gz,
  Apple Silicon and Intel in one file), which the cask installs. The
  per-architecture macOS, Linux and Windows builds are unchanged.

skimmer v0.1.103

In answer to the review in #3.

Changed

• The default model is pinned to jev-1.14.0, the model the evals measured,
  instead of the jev-latest alias (which points at jev-1.13.0). --model,
  SKIMMER_MODEL and a prompt's model: still choose another.
• The stderr summary and skim --json ("model") name the model that
  answered. The summary also counts unjudged chunks.
• The batch planner charges a chunk's text once, or twice only when a prompt
  pastes it with {{.Text}}, and measures the prompt itself without
  subtracting text it no longer contains. Batches with large --chunk-chars
  now hold up to about twice as many chunks.

Added

• --injection-check for skim and filter: a second yes/no question per
  chunk or item, in the same requests, drops ones that try to steer the
  reader (listed as flagged in --json). Measured: planted lines kept
  1–6% of the time instead of 18–20%, 0 flags in 5,439 clean judgments,
  about 25% more input tokens; in documents with a planted line it cost 3–4
  of 66 goals, by dropping answers next to the line. A filter, not a security boundary.
  The prompt is injection_check/DEFAULT.md.
• Prompt frontmatter takes criteria (true and false text), sent as the
  Noul's criteria.
• Evals: a frozen dev/holdout split, a pre-registration, near-miss negative
  goals, span-based scoring that holds across chunk sizes and shifts, one
  live runner (evalskim) and an offline report (evalreport: Wilson
  intervals, exact McNemar, a BM25 baseline, cross-draw flips, injection).
  Results in evals/skimmer/0.1.103/results.md.
• A label audit by a second model family (gpt-6-astra via codex exec,
  evals/cmd/labelaudit): it agrees with all v1 gold, distractor and
  near-miss labels and 13 of 14 confusables, but only 12 of 36 v2 labels,
  and every skimmer miss falls on a disputed v2 label. evalreport -agreed
  re-scores on agreed goals only; no decision changes.

Fixed

• Injection eval chunks now map through the planted insertion to their
  original-document spans, including chunks containing only part of the line.
  Repaired the saved spans offline without changing model probabilities:
  the check loses 3–4 complete goals of 66 per planted-line run, previously
  hidden by missing spans. Updated the results and README.
• The chunk-size sweep scored near-duplicate goals by chunk id, which names
  different text at other sizes. Re-scored by span, smaller chunks do not
  lose discrimination; the 0.1.101 claim is withdrawn.
• Eval claims corrected in place: headline counts now carry n and intervals;
  the 0.7 threshold advice is withdrawn (it loses 8 of 42 held-out goals);
  the 0.1.102 prompt change measured 14.5% fewer tokens with no measurable
  recall difference, not "19% fewer, equal-or-better at every rank".

skimmer v0.1.102

Changed

• Built-in skim and filter prompts no longer paste the goal or the chunk
  or item text into each question. Every chunk is sent once, in the request
  state, and each question points at it by id. Measured on the frozen
  goals-v2 eval (pinned jev-1.14.0, 2 draws): 19% fewer input tokens per
  goal, equal-or-better recall at every rank and better all-modes coverage
  (0.750 vs 0.722). See evals/variants/README.md. Custom prompts using
  {{.Text}} still work, but send the text twice.

Evals

This release ships the eval suite: a 12-document public corpus with
attributions (including three mutually-confusable arXiv RAG papers and the
System One paper itself), a gold-by-construction goal set authored and
blind-audited across two model families, a live harness running the real
pipeline against a pinned Jev model with every per-chunk probability
archived, and eval history foldered by skimmer version. Measured:
recall 1.000 on find-the-fact goals, distractor leakage 0.000, robust to
boundary shifts; threshold 0.5 conservative (0.7 keeps 30% fewer chunks at
identical recall); mode drops real on collect goals and fixed in part by
the state-pointer prompt. See evals/.

skimmer v0.1.100

First release of skimmer 🎉: skim pages, files, and lists with Jev and keep only what helps with your goal.

• skimmer skim: URLs, files, directories, grep/rg output, or stdin → only the helpful chunks, with file chunks cited as path:start-end (10% chunk overlap by default)
• skimmer filter: JSON arrays, JSON Lines, or plain lines → only the helpful items, in the same format
• skimmer auth: store your key once in ~/.config/skimmer; env vars, .env, and credential-injecting proxies work too
• skimmer prompts: built-in default prompts, plus your own overrides

Install: brew install dorkitude/tap/skimmer · scoop install dorkitude/skimmer · apt (see README) · go install github.com/dorkitude/skimmer@v0.1.100