refactor: add DesignIR distillation pipeline

2026-07-27 17:24:04 +08:00
parent 0adaac2b2a
commit 550b339bcc
1128 changed files with 10248 additions and 426666 deletions
@@ -0,0 +1,97 @@
+# LLM semantic review contract
+
+Read the sanitized `work/review/semantic-batch.json`, not raw STEP geometry.
+Maintain stable proposal identifiers when an existing candidate expresses the
+same method.
+
+Write `work/review/experience-draft.json` with this shape:
+
+```json
+{
+  "schema_version": "1.0",
+  "draft_kind": "llm_generalized_experience_proposals",
+  "proposals": [
+    {
+      "id": "constraint.part_family.shared_axis",
+      "kind": "constraint",
+      "scope": ["part_family"],
+      "evidence_query": {
+        "required_features": ["rotational_body", "central_passage"],
+        "required_relations": ["dominant_axis_alignment"],
+        "context_features": ["rotational_body", "central_passage"]
+      },
+      "guidance": "Preserve a shared axis across the participating feature roles.",
+      "semantic_rationale": "The relationship remains meaningful when individual dimensions change.",
+      "check": "Confirm that participating features retain their intended common axis."
+    }
+  ]
+}
+```
+
+Allowed kinds are `feature_motif`, `constraint`, `design_rule`,
+`validation_rule`, `failure_repair`, `dimensionless_distribution`, and
+`reconstruction_grammar`.
+Distribution proposals also provide `observation_name`; the publisher computes
+the distribution from private evidence and never accepts values from the draft.
+
+A `reconstruction_grammar` proposal must additionally contain symbolic,
+number-free reconstruction content:
+
+```json
+{
+  "id": "reconstruction.patterned_prismatic_part.primary",
+  "kind": "reconstruction_grammar",
+  "scope": ["patterned_prismatic_part"],
+  "evidence_query": {
+    "required_features": ["planar_dominant_body", "cylindrical_feature_network"],
+    "required_relations": [],
+    "context_features": ["planar_dominant_body"]
+  },
+  "guidance": "Build a datum-driven envelope before adding the passage network and repeated features.",
+  "semantic_rationale": "The staged method transfers across dimensional variants while preserving feature relationships.",
+  "reconstruction_grammar": {
+    "parameter_roles": ["overall_long_span", "overall_middle_span", "overall_short_span"],
+    "datum_roles": ["part_center", "primary_axis", "primary_transverse_plane"],
+    "feature_roles": ["planar_dominant_body", "cylindrical_feature_network"],
+    "relation_roles": [],
+    "canonical_stages": [
+      {
+        "id": "establish_reference_frame",
+        "operation": "define_datums",
+        "feature_roles": [],
+        "reference_roles": ["part_center", "primary_axis"]
+      },
+      {
+        "id": "construct_primary_envelope",
+        "operation": "extrude_profile",
+        "feature_roles": ["planar_dominant_body"],
+        "reference_roles": ["part_center", "primary_axis"]
+      }
+    ],
+    "cardinality_classes": ["plane_dense", "cylinder_dense"],
+    "validation_roles": ["closed_solid", "surface_mix_consistency"]
+  }
+}
+```
+
+Follow these rules:
+
+- Require an evidence query composed only of semantic feature and relation
+  tokens present in the sanitized batch.
+- Use scopes to prevent unrelated part families from supporting a proposal.
+- Use `context_features` to define the conditional population used for
+  confidence, so a valid subtype method is not diluted by unrelated variants
+  in the same broad family.
+- Explain why a proposal transfers across dimensions and variants.
+- Avoid dimensions, coordinates, source identifiers, face references, and
+  numeric literals in prose.
+- Do not infer build order or causal failure mechanisms from final geometry.
+- Treat `canonical_stages` as a recommended reconstruction order, never as a
+  claim about the source CAD history.
+- For reconstruction grammars, retain only roles and stages repeated by the
+  scoped evidence. Do not insert family knowledge that the sanitized batch
+  does not support.
+- Prefer a small set of meaningful methods over every possible feature pair.
+- Preserve earlier proposal IDs when evidence still supports the same meaning.
+- Let the publisher calculate support and confidence; never claim them in the
+  draft.