cookbook/environments/_22_sql_generation/TEST_LOG.md
Tested 2026-07-20 live against gpt-5.5, agno 2.7.4, using
.venvs/demo/bin/python with OPENAI_API_KEY loaded from .envrc.
Status: PASS
Description: Recursive inventory-state replay with accepted, rejected, duplicate, and unknown reservation references.
Result: 8 attempts in 48s. inventory-state landed at 7/8 (0.875), all
scored. The failed query reached the right recursive shape but used non-text
json_object() labels and SQLite rejected it. Two earlier versions were discarded
after saturating 8/8: a temporal ledger, then marginal tiers plus promotion precedence;
a recursive subscription state machine also saturated 8/8. Reserve acceptance plus
single-use release state was the change that escaped the wall of full bars.
Status: PASS
Description: Multi-table first-human-response query measured in business minutes across a holiday and overnight boundaries.
Result: 8 attempts in 18s, all scored. business-minute-sla was 7/8
(0.875). The failing query returned both organizations at 100%, exposing a business
minute boundary error. The first version used a fixed four-hour wall-clock SLA and
saturated 8/8; adding per-organization thresholds, a holiday, overnight spans, bot
and pre-open distractors, and exact minute semantics created the useful middle band.
Status: PASS
Description: Recursive inventory state retained per event, followed by windowed upward threshold crossings.
Result: 16 attempts in 85s, all scored. stateful-threshold-crossing was
7/8 (0.875); final-state-audit saturated at 8/8. The failing crossing query used
non-text JSON labels and SQLite rejected it. Three earlier grids saturated 8/8: an
effective-dated price crossing, the same stream plus returns, and marginal tiers plus
promotion precedence. Replacing arithmetic windows with reserve acceptance,
single-use releases, a recursive trajectory, and repeat up-crossings finally exposed
the middle band.