让数字对得上：修抽取端的三处失真，加评测台

CLAUDE.md

@@ -9,7 +9,8 @@ LoreGraph extracts knowledge graphs from **closed-world fictional texts** (novel
 - Python 3.11+ only. `from __future__ import annotations` at the top of every module.
 - All async by default. Synchronous wrappers only at CLI entry points.
 - **All LLM calls go through `src/loregraph/llm/client.py`.** Never instantiate a provider SDK (`anthropic.Anthropic()`, `openai.OpenAI()`) directly elsewhere.
-- **Every extracted claim must carry an `evidence_span`** — a literal substring of the source chunk. Pass-7 enforces ≥ 95% literal match rate as a hard gate.
+- **Every extracted claim must carry an `evidence_span`** — a literal substring of the source chunk. Pass-7 enforces two gates: `supported_rate` (does the span entail the claim?) against `LOREGRAPH_COVE_SUPPORTED_FLOOR`, and a ≥ 95% literal-match tripwire that only fires on a span-handling bug.
+- **Edges carry both a free-form `predicate` and a closed `predicate_class`** (`models/predicates.py`). The verb is for reading, the class is for querying. Never query on the verb.
 - Pydantic models live in `src/loregraph/models/`. Database schema in `src/loregraph/db/schema.py`. Keep them in lockstep via integration tests (`tests/integration/test_db.py`).
 - Lint: `uv run --extra dev ruff check && uv run --extra dev ruff format`. Type check: `uv run --extra dev python -m mypy src`.
 
@@ -37,6 +38,25 @@ Touch points (in order):
 
 Every LLM call records token usage to `pass_runs.stats`. A per-book budget ceiling (`LOREGRAPH_COST_CEILING_USD`, default $20; `0` disables) is **enforced** between passes by `pipeline/orchestrator.py`: it estimates spend from token counts × configurable prices (`LOREGRAPH_PRICE_INPUT_PER_MTOK` / `…_OUTPUT_PER_MTOK`, default DeepSeek V4 Pro) and aborts (resumable with `--from`) when exceeded. Always design prompts so the system prompt + shared ontology block stays **stable across chunks** — this lets prompt caching (Anthropic, or `anthropic/*` via OpenRouter) kick in and saves 80%+ on input tokens.
 
+## Evaluating a change
+
+Counts are not quality. Before claiming an extraction change helped, run
+`uv run loregraph eval all --book <id>` and quote the numbers that moved.
+
+- `graph` and `gaps` need no model or credentials — they always run.
+- `entailment`, `contamination` and `perturbation` need a provider; without
+  one they print a dry preview of exactly what would be sent.
+- Every result carries a `skipped` list. Read it: an eval that scored the
+  subset it could reach is not coverage.
+
+Two figures are routinely confused and must not be:
+`literal_match_rate` is an invariant tripwire (Pass-2/5/6 drop non-literal
+spans, so it is 1.0 by construction and cannot measure anything);
+`supported_rate` is the quality number. Quote the second.
+
+New evals go in `src/loregraph/evals/`, one module per eval, returning an
+`EvalResult`, registered in `cli/main.py`'s `eval` command.
+
 ## Testing tiers
 
 | Tier | Marker | Speed | Deps |