70 lines
3.2 KiB
Markdown
70 lines
3.2 KiB
Markdown
# Selector
|
|
|
|
This folder contains rule-based selector configuration and examples.
|
|
|
|
The v0.2 selector should not call an LLM.
|
|
|
|
It should use simple matching rules:
|
|
|
|
- Trigger keywords
|
|
- Input types
|
|
- Negative triggers
|
|
- Model-level hard exclusion triggers
|
|
- Pipeline position
|
|
- Selection priority
|
|
- Negative trigger first
|
|
- No-call threshold
|
|
- QPI-before-IA gate
|
|
- QPI positive-signal gate
|
|
- Direct-execution no-call signals
|
|
- Calibration smoke checks
|
|
|
|
Round 03.2 selector rule:
|
|
|
|
- Explicit refusal of a specific model, such as `不要思想考古` or `只做 QPI`, is a model-level hard exclusion for that model.
|
|
- A hard exclusion rejects only the named model; it does not globally no-call the request.
|
|
- Generic bare QPI gate words such as `如何` and `判断` should not be used alone. Use compound problem-definition phrases instead.
|
|
|
|
Round 03.2a selector rule:
|
|
|
|
- Depth-limiting phrases such as `不要展开` and `不要深入分析` remain global no-call signals by default.
|
|
- If a depth-limiting phrase is paired with explicit QPI-limited intent such as `只做 QPI`, `只做问题定性`, or `只判断主导稀缺`, the selector should continue scoring, allow QPI, and still reject IA.
|
|
|
|
Round 04.1 selector rule:
|
|
|
|
- Translation / direct transformation instructions should no-call even when the payload contains complexity words such as `模型`.
|
|
- Task instruction and payload content should be separated for no-call routing when a direct transformation phrase appears before a colon.
|
|
- IA refusal variants such as `不要进入思想考古` and `不做思想考古` are model-level hard exclusions for IA.
|
|
- Depth-limiting phrases paired with `主导稀缺` or `缺信息 / 缺路径 / 缺共识` intent should allow QPI and reject IA.
|
|
|
|
Current files:
|
|
|
|
- `selector_rules.json`
|
|
- `selector_examples.json`
|
|
- `selector_calibration_inputs.json`
|
|
- `round04_blind_inputs.json`
|
|
- `qpi_case_digests.json`
|
|
|
|
The selector does not call an LLM.
|
|
|
|
The executable demo is `scripts/run_selector_demo.py`.
|
|
|
|
Selector regression is `scripts/run_selector_regression.py`.
|
|
|
|
Selector calibration smoke check is `scripts/run_selector_calibration_smoke.py`.
|
|
|
|
Round 04 blind routing evaluation is `scripts/run_round04_blind_routing.py`.
|
|
|
|
Round 04.1 post-patch routing verification is `scripts/run_round04_post_patch_verification.py`.
|
|
|
|
`round04_blind_inputs.json` is a frozen blind input pool for first-run routing evaluation. It intentionally contains no expected selector behavior and should not be merged into calibration inputs before owner / GPT / CCRA review.
|
|
|
|
`qpi_case_digests.json` stores owner-reviewed QPI case digests that can feed selector calibration or later regression selection. It is not a model status upgrade and should not include unreviewed draft cases.
|
|
|
|
Digest field notes:
|
|
|
|
- `misclassification_risk` is the canonical digest field matching QPI's structured output contract.
|
|
- `qpi_complexity_pattern` records judgment complexity: `not_mixed`, `intra_frame_mixed`, or `inter_viewpoint_divergence`.
|
|
- `qpi_complexity_pattern=intra_frame_mixed` does not imply `classification=mixed`; `classification` is the final routing class, while `qpi_complexity_pattern` records the structure of the judgment.
|
|
- Multi-perspective / inter-viewpoint cases require `classification_by_viewpoint` or `viewpoint_summary`.
|