54 lines
1.6 KiB
Markdown
54 lines
1.6 KiB
Markdown
|
# Hybrid Pipeline Contract
|
||
|
|
|
||
|
|
This editor supports two entry paths:
|
||
|
|
|
||
|
|
1. `language-first`: operate from source text in a target language.
|
||
|
|
2. `ast-first`: operate directly on canonical AST/IR structures.
|
||
|
|
|
||
|
|
Both paths must converge on shared canonical semantics before projection to a target language.
|
||
|
|
|
||
|
|
## Contract Rules
|
||
|
|
|
||
|
|
1. Canonical-first convergence
|
||
|
|
- Every sprint result should include canonical signals (for example `AST`, `IR`, `Schema`, `Canonical`, `Semantic`, `Model`) in integration artifacts.
|
||
|
|
|
||
|
|
2. Projection neutrality
|
||
|
|
- Every sprint should show evidence that behavior is not locked to C++ only.
|
||
|
|
- Acceptable evidence includes non-C++ language/projection signals (for example `Rust`, `Go`, `Java`, `Python`, `TypeScript`, `Wasm`, SQL-family adapters).
|
||
|
|
|
||
|
|
3. C++ as backend, not source of truth
|
||
|
|
- C++ implementation artifacts are valid runtime host/projection work.
|
||
|
|
- They are not sufficient alone to satisfy hybrid pipeline quality.
|
||
|
|
|
||
|
|
## Validation Script
|
||
|
|
|
||
|
|
Use:
|
||
|
|
|
||
|
|
```bash
|
||
|
|
./tools/mcp/validate_hybrid_pipeline_contract.sh --start 50 --end 90
|
||
|
|
```
|
||
|
|
|
||
|
|
Strict enforcement (fail CI/run if checks fail):
|
||
|
|
|
||
|
|
```bash
|
||
|
|
./tools/mcp/validate_hybrid_pipeline_contract.sh \
|
||
|
|
--start 50 --end 90 --strict --enforce-non-cpp
|
||
|
|
```
|
||
|
|
|
||
|
|
Write machine-readable report:
|
||
|
|
|
||
|
|
```bash
|
||
|
|
./tools/mcp/validate_hybrid_pipeline_contract.sh \
|
||
|
|
--start 50 --end 90 --json-out logs/taskitem_runs/hybrid_contract_50_90.json
|
||
|
|
```
|
||
|
|
|
||
|
|
## Suggested Workflow Hook
|
||
|
|
|
||
|
|
After finishing a sprint batch:
|
||
|
|
|
||
|
|
1. Run taskitem pipeline.
|
||
|
|
2. Export LoRA capture.
|
||
|
|
3. Run hybrid contract validator.
|
||
|
|
4. Treat failures as blocking when preparing training-ready data.
|
||
|
|
|