Delete READ.ME

README.md

@@ -1,104 +0,0 @@
-# SQL Quality Audit Agent Environment
-
-OpenEnv-compatible environment for training SQL data-quality agents with strict v2 trajectory scoring.
-
-## Overview
-
-`sql_quality_env` is a FastAPI environment with OpenEnv endpoints:
-
-- `POST /reset`
-- `POST /step`
-- `GET /state`
-- `GET /health`
-
-Agents interact through four action types (`query`, `profile`, `fix`, `submit`), and final rewards are computed from full trajectories using a five-dimension composite grader.
-
-## Tasks
-
-| task_id | Difficulty | Scenario |
-|---|---|---|
-| `null_audit` | easy | Identify which column has >50% null rate |
-| `type_repair` | medium | Repair numeric type mismatch while preserving values |
-| `dedup_bug` | hard | Diagnose and fix duplication causing inflated joins |
-| `fk_integrity` | hard | Remove orphan child rows while preserving valid FK-linked rows |
-| `schema_drift` | medium | Canonicalize drifted upstream schemas into a normalized artifact |
-
-## Scoring Architecture
-
-Final score (0.0-1.0) uses five dimensions:
-
-- D1 Diagnostic Efficiency
-- D2 Hypothesis Quality
-- D3 Fix Correctness (severity tiered, catastrophic override)
-- D4 Self-Verification bonus
-- D5 Generalization on schema/data variants
-
-`null_audit` intentionally uses direct answer scoring:
-
-- exact answer: `1.0`
-- wrong but valid schema column: `0.2`
-- out-of-schema answer: bounded lexical near-miss partial, otherwise `0.0`
-
-## Deterministic Baseline Recipe
-
-The baseline runner (`inference.py`) is deterministic when `BASE_SEED` is fixed.
-
-Seed schedule:
-
-- task seed = `BASE_SEED + task_index`
-
-Example reproducible run:
-
-```bash
-uvicorn server.main:app --host 127.0.0.1 --port 7860
-BASE_SEED=20260403 MAX_STEPS_PER_TASK=6 ENV_BASE_URL=http://127.0.0.1:7860 python inference.py
-```
-
-The run output is saved at:
-
-- `baseline_seeded_run.txt`
-
-## Measured Seeded Baseline (2026-04-03)
-
-Model: `meta-llama/Llama-3.3-70B-Instruct`  
-API: `https://router.huggingface.co/v1`  
-Base seed: `20260403`
-
-| Task | Seed | Score | Steps |
-|---|---:|---:|---:|
-| `null_audit` | 20260403 | 1.0000 | 3 |
-| `type_repair` | 20260404 | 0.0000 | 1 |
-| `dedup_bug` | 20260405 | 0.0000 | 1 |
-| `fk_integrity` | 20260406 | 0.0000 | 1 |
-| `schema_drift` | 20260407 | 0.0000 | 1 |
-| **Average** | - | **0.2000** | - |
-
-## Install
-
-```bash
-pip install -e ".[test]"
-```
-
-## Validate
-
-```bash
-python -m pytest -q
-```
-
-## Docker
-
-```bash
-docker build -t sql-quality-env:latest .
-docker run -p 7860:7860 sql-quality-env:latest
-```
-
-## Environment Variables
-
-| Variable | Required | Purpose |
-|---|---|---|
-| `HF_TOKEN` (or `OPENAI_API_KEY`) | yes | API auth for model calls in baseline inference |
-| `MODEL_NAME` | yes | model id for inference |
-| `API_BASE_URL` | no | LLM API endpoint (defaults to HF router) |
-| `ENV_BASE_URL` | no | environment URL (defaults to local `http://localhost:7860`) |
-| `BASE_SEED` | no | deterministic baseline seed |
-| `MAX_STEPS_PER_TASK` | no | runtime guard per task |