Skip to contents

Changelog

Source: NEWS.md

cat.stack 0.2.2

Python engine floor raised to cat-stack 2.0.1

  • install_cat_stack() (and SystemRequirements) now require cat-stack >= 2.0.1 — the stable 2.0 engine. This is where the fixes land that R users on the newest Anthropic generation need: creativity and thinking_budget no longer 400 on Opus 4.7+/Sonnet 5/Fable 5 (adaptive thinking + parameter gating are handled per model), thinking_budget grades consistently across providers, and description context routing is fixed in classify()/prompt_tune(). Existing installs: run install_cat_stack(upgrade = TRUE) once.

Deprecation-proof survey_question forwarding

  • classify() and prompt_tune() now forward survey_question to the engine as the canonical description parameter, so R callers never trigger the Python-level DeprecationWarning. The R signatures are unchanged; survey_question remains a documented soft-deprecated alias.

cat.stack 0.2.1

New parameter in classify()

  • multi_label — logical, default TRUE (unchanged behavior). The Python engine has always supported single-label classification, but the R wrapper did not forward the flag, so classify() could only ever request multi-label output. Setting multi_label = FALSE now reaches the engine and switches the prompt to “assign the single most appropriate category” (one 1, the rest 0), for mutually exclusive coding frames.

cat.stack 0.2.0

Tracks the Python cat-stack 1.6.0 release. The version floor in install_cat_stack() is now cat-stack >= 1.6.0 — older Python installs will hit unexpected keyword argument errors when R forwards the new parameters listed below.

New parameters in classify()

  • batch_mode, batch_poll_interval, batch_timeout — opt into the async batch APIs (OpenAI / Anthropic / Google / Mistral / xAI) for ~50% cost savings and higher rate limits. HuggingFace / Perplexity / Ollama fall back to synchronous calls.
  • json_retries — per-row retries when the LLM returns JSON that fails schema validation. On the final attempt the formatter fallback fires (if json_formatter is enabled).
  • json_formatter — three-state (TRUE / FALSE / NULL) control for the local JSON-repair model. Default NULL triggers an interactive consent prompt on the first malformed row; non-TTY contexts decline silently. Requires cat-stack[formatter].
  • two_step_classify — split classification into reasoning + JSON formatting steps for weaker models (lower-tier API + local Ollama). Auto-enables json_formatter when set.
  • embedding_tiebreaker + min_centroid_size — resolve true ensemble 50/50 ties via embedding-centroid similarity instead of the default “tie → 0”. Adds category_N_resolved_by audit column ("vote" or "centroid"). Multi-model ensemble + text input only; not yet supported in batch_mode. Requires cat-stack[embeddings].

Behavior changes

  • consensus_threshold = "majority" is now strict majority. A 50/50 tie on an even-model ensemble (2-2 of 4, 3-3 of 6, 1-1 of 2) resolves to "0", not "1". Matches sklearn’s VotingClassifier default and standard ensemble literature. Numeric thresholds (e.g. consensus_threshold = 0.5) keep >= semantics — the user picked a number, they get the literal interpretation. For 2-model ensembles, "majority" now effectively requires both models to agree on positive. Use 3+ models for a non-degenerate majority vote, or pair with embedding_tiebreaker = TRUE.
  • batch_retries default lowered from 2L to 1L to match the Python default (changed in cat-stack 1.4.1).

Documentation

  • consensus_threshold man page now describes the strict-majority semantics, the 2-model degeneracy, the numeric-input escape hatch, and the embedding_tiebreaker companion.
  • Added a new example block showing strict-majority + tiebreaker and batch_mode = TRUE.

Python-side fixes that flow through automatically

These are in Python cat-stack 1.6.0 and pass through the reticulate bridge without R-side changes — listed here so users know the behaviors they’re getting after install_cat_stack(upgrade = TRUE):

  • Google preflight 400 fixed (additionalProperties stripped before reaching responseSchema).
  • Per-model batch-job failure isolation — one model’s batch failure no longer kills the ensemble run.
  • Anthropic batch terminal-state inspection — all-errored batches raise instead of silently returning empty results.
  • system_prompt is no longer silently dropped in batch_mode.
  • PDF summary synthesis grounds on actual page text instead of the page label.
  • HuggingFace small-model strip-on-5xx (Llama-3.2-1B / response_format).
  • Image directory loading is case-insensitive; large images warn.
  • prompt_tune returns system_prompt = "" when no improvement was found (instead of returning a non-improving prompt as the default).

cat.stack 0.1.0