picarones/reports/narrative/templates/en.yaml

# Narrative rendering templates — English.
# Anti-hallucination rule: never introduce a number or entity name that is not
# already in the Fact ``payload``. Tests verify traceability of every number
# appearing in the rendered synthesis.

global_leader_cer: >-
  On this corpus of {n_docs} documents, {engine} achieves the lowest mean CER
  ({cer_pct} %).

statistical_tie: >-
  Engines {engines_list} are not statistically distinguishable
  (Friedman-Nemenyi, α = {alpha}, n = {n_blocks} documents, CD = {critical_distance}).

significant_gap: >-
  The gap between {leader} and {runner_up} is statistically significant
  (Wilcoxon, p = {p_value:.4f}, Δ CER = {delta_cer_pct} points over {n_pairs} pairs).

stratum_winner: >-
  On stratum "{stratum}" ({n_docs_stratum} documents), {engine} achieves
  the lowest CER ({cer_pct} % vs. {second_cer_pct} % for {second_engine}).

stratum_collapse: >-
  {engine} is globally competitive ({global_cer_pct} %) but collapses on
  stratum "{stratum}" ({local_cer_pct} % over {n_docs_stratum} documents,
  i.e. {delta_cer_pct} points above its own average).

error_profile_outlier: >-
  {engine} has an atypical error profile: {proportion_pct} % of errors fall
  into class "{error_class}", vs. a median of {median_proportion_pct} % across
  other engines (×{ratio_to_median} the median).

llm_hallucination_flag: >-
  Hallucination signal on {engine} ({reasons_list}) —
  {hallucinating_rate_pct} % of documents above alert thresholds.

robustness_fragile: >-
  {engine} is fragile under "{degradation}" degradation: its CER rises from
  {cer_baseline_pct} % to {cer_degraded_pct} % at maximum level (×{ratio}).

speed_winner: >-
  {engine} is the fastest ({mean_duration} s/doc, ×{speedup} faster than the
  median) for comparable quality (CER {cer_pct} %).

confidence_warning: >-
  High statistical uncertainty: the {confidence_level} % confidence interval of
  {engine} spans {ci_width_pct} CER points, compared with a gap of
  {gap_to_runner_up_pct} points to the runner-up.

pareto_alternative: >-
  At much lower cost, {engine} offers an interesting trade-off ({cer_pct} %
  CER for {cost} €/{cost_unit_pages} pages, vs {leader_cer_pct} % / {leader_cost} € for
  {leader}, i.e. ×{cost_saving_ratio} cheaper).

cost_outlier: >-
  Disproportionate cost for {engine} ({cost} €/{cost_unit_pages} pages, ×{ratio_to_median}
  the median) without a compensating quality advantage (CER {cer_pct} %).

ensemble_opportunity: >-
  Engines {pair_a} and {pair_b} have divergent error profiles
  ({divergence_metric}={divergence}). On this corpus of {doc_count} documents,
  {best_engine} preserves {best_recall_pct} % of tokens; a majority vote
  among the engines would preserve {oracle_recall_pct} % — i.e.
  {absolute_gap_pct} points recoverable ({relative_gap_pct} % of the best
  engine's errors).

median_mean_gap_warning: >-
  Asymmetric distribution for {engine}: median CER {median_cer_pct} %
  vs mean {mean_cer_pct} % across {n_docs} documents (relative gap
  {relative_gap_pct} %). The mean is pulled by a few catastrophic
  documents — the median (now used for default ranking) is more
  representative.

stratification_recommended: >-
  Heterogeneous corpus ({n_strata} strata): {leader} performs very
  differently depending on document type — median CER
  {min_stratum_cer_pct} % on "{min_stratum}" vs
  {max_stratum_cer_pct} % on "{max_stratum}", a gap of {gap_pct}
  points. The global ranking hides this disparity; consult the
  stratified view.

engine_off_baseline: >-
  {engine} achieved {cer_current_pct} % CER here, vs {cer_historical_mean_pct} %
  on average over the last {n_runs} runs of your institution on this
  same corpus (relative delta {relative_delta_pct} %). This corpus is
  harder for it than usual.

engine_unstable: >-
  Over {n_runs} successive runs, {engine} produces variable outputs
  (CER CV {cer_cv_pct} %, identical-run pair rate {identical_run_rate_pct} %).
  Reproducibility is limited — interpret the average CER with caution.

regression_in_history: >-
  Over the {n_runs} historical runs for {engine}, the average CER
  moved from {first_cer_pct} % to {last_cer_pct} %
  (cumulative change {absolute_delta_pct} points). Investigate what
  changed in the pipeline or the models.

# Sprint A3 (item B-3) — importer fallback incidents.
# The payload contains `importer`, `operation` and `incidents_for_importer`.
importer_fallback_triggered: >-
  The "{importer}" importer fell back to degraded mode during the
  "{operation}" operation ({incidents_for_importer} incident(s) this
  run). Imported data may be incomplete or from a fallback — check
  the logs for details.

# Sprint A8 (m-14) — pricing table expired.
# Payload: ``valid_until``, ``days_overdue``, ``today``.
pricing_staleness_warning: >-
  The pricing table used expired on {valid_until} ({days_overdue}
  days ago). Cost/€ and CO₂ analyses reflect outdated rates — pass
  ``pricing=`` to ``ReportGenerator`` for fresh figures.