Prompt version 1.2.0: removes the hard score floors and cross-dimension double counting that 1.1.0 introduced, adds a structural calibration rule so the 8 anchor is not a default for ambiguity, and sharpens the procedures for decision flips, recomputable indicators, policy feedback and solution adjudication. The anchor rubrics and direction statement are unchanged. Scores from 1.0.0, 1.1.0 and 1.2.0 are not poolable.
Prompt version 1.1.0: every dimension template now carries a scope guard (the evidence that belongs to the dimension and the evidence it must not absorb from another dimension) and an ordered scoring procedure (quote the evidence, match the closest anchor, apply the dimension’s partial-credit rules, interpolate, and reserve the extremes for structural cases). The anchor rubrics at 2, 5 and 8 and the direction statement are unchanged. Scores produced under 1.0.0 and 1.1.0 are not poolable; the prompt version is recorded in every result and every benchmark row.
Add the methodology vignette, which documents the evolution from a naive single-shot judge to the current design and measures the effect of 1.1.0 and 1.2.0 against the frozen benchmarks.
Benchmark caches now record prompt_version per row. The 1.0.0 and 1.1.0 runs are frozen in inst/extdata/model-comparison-results-v1.0.0.parquet and inst/extdata/model-comparison-results-v1.1.0.parquet; the live cache always holds the current version.
liminal2 0.2.0
Add assess_liminal_space(): scores a business ask (and optional situation) through seven dimension-specific prompts with an LLM and aggregates the scores into the size of the liminal space. The seven calls run concurrently by default; failed dimensions are retried sequentially with backoff.
Add liminal_chat() for the OpenAI-compatible Panama endpoint (model glm-5.3 by default, temperature 0), configured entirely through the PANAMA_API_KEY, PANAMA_BASE_URL and LIMINAL2_MODEL environment variables. ellmer builds the client; the scoring transport itself runs on httr2, because the gateway labels its JSON responses text/plain, which ellmer’s strict content-type check rejects.
Add liminal_prompts() and liminal_dimensions(): the seven versioned prompt templates ship in inst/prompts/ with fixed anchor rubrics that are independent of the test scenarios, keeping scores comparable across business cases.
Implement measure_liminal_space() and default_weights() as specified by the test suite: L = 1 - prod(scores ^ weights) with full input validation and informative error classes.
Add the model-comparison vignette, which benchmarks every model available on the endpoint over the six test scenarios and caches the raw scores, token usage and cost in inst/extdata/model-comparison-results.parquet.