[Jibbscript]

Summary

The real-model A/B on eval/fixtures/t1-energy-audit (eval/results/t1_ab_report.md) found two arithmetic defects and one research miss the skills should have prevented. Each fix is a short reasoned paragraph in the skill's existing Fable 5.1 register (no CAPS, no checklists, no step scripts); frontmatter, contracts/, eval/, and LICENSE are untouched.

• market-sizing — the model built "62% of homes pre-2000" from NAHB buckets by mislabelling a 2000–2009 bucket (correct figure ~72%). New paragraph in Working the Numbers: a derived share is recomputed in the artifact from the cited buckets, each boundary quoted as the source labels it, with the addition written out; the closing self-check re-adds any bucket-derived share from the quoted boundaries.
• pricing-wtp — the generated value_ladder ran $0, $29, $79, $60, and ltv_cac_ratio was 3.6 while its own inline math gave 3.8. The ladder is checked for increasing price and increasing value before it is written (a rung that breaks the order is usually a different product wearing a tier label), and LTV:CAC is arithmetic over the file's own fields, stated once in ltv_cac_ratio with its formula beside it; the closing self-check covers both.
• competitive-landscape — both pack runs missed HomeBoost, a funded direct competitor one TechCrunch search away. One search per landscape now goes to funded startups in the space the brief's solution hypothesis actually describes, plus two query templates.

Benchmark

python3 eval/fable_prompt_bench.py -v: primary 95.25 → 95.37, zero integrity failures, boilerplate_dups 0. Per-skill: competitive-landscape 93.7 → 94.6, market-sizing 96.6 → 96.7, pricing-wtp 92.2 → 93.0.

🤖 Generated with Claude Code