From question to guarded SQL.

Inspect real evidence, one pipeline decision at a time.

Recorded examples only.
No generation or SQL execution during replay.

Live AI / full-schema context

Private testing only. The project owner must approve access and participant eligibility before sharing a private access code.

No online embedding or pruning. Gemini compatibility and accuracy are not yet measured. The figures below are historical Qwen results.

Free services may sleep or run out of quota. Five outbound attempts per day are shared globally; retries consume allowance. Up to two submissions per observed address per UTC day; visitors may share that allowance. No paid fallback.

Recorded sessionWaiting for recording

Inspect the pipeline

Select a curated example to see its question, schema, SQL, validation and result.

01Schema selection
02LLM generation
03AST validation
04Safe execution
05Audit

Stages describe the original evidence. Replay reads a static recording; it does not execute this pipeline or create an audit entry.

Selected schema

Candidate SQL

No recording loaded.

Final safe SQL

No recording loaded.

Validation

The deterministic gate is independent of the model.

Recorded result

No result loaded.

Recording provenance

Measured, including the tradeoffs.

Custom synthetic benchmark. One local model, three repeats.
Not Spider/BIRD. Not a public leaderboard.

Step 10 / 100 questions, including 64 complex questions / qwen3.8:latest
Configuration	Execution accuracy	Median latency	p95 latency
Full / no retry	94.67%	20.02 s	29.84 s
Pruned / no retry	77.33%	9.48 s	37.91 s
Pruned / retry	77.00%	9.44 s	38.14 s

Pruning effect

-17.33 pp

Smaller context and lower median latency came with substantially lower accuracy and worse p95 latency.

Retry uplift

-0.33 pp

Bounded correction preserves validation, but measured no accuracy improvement. Zero successful corrections in Step 10.

Deterministic enforcement

96/96 adversarial safe outcomes

6/6 unsafe candidates blocked.
0 unsafe SQL executed.
104/104 SQL attacks blocked; 0 unexpected accepts.
48/48 supported-valid queries accepted; 0 false rejects.

Curated successes illustrate behavior, not average performance or pruning superiority. Safety results describe tested inputs under this gate policy, not a guarantee against unknown vulnerabilities. Three repeats do not establish statistical significance.