Document WCAB 0.3.0 evaluation protocol

Add the tool-neutral observation contract, coverage-aware scoring rationale, and v0.3.0 validation record while retaining the schema-v2 fixture snapshot.

README.md

@@ -62,4 +62,5 @@ wcab validate --fixtures fixtures
 ```
 
 See the [source repository](https://github.com/SybilGambleyyu/workbook-change-benchmark)
-for the schema, validation contract, and releases. The dataset is MIT licensed.
+for the truth schema, tool-neutral observation protocol, validation contract,
+and releases. The dataset is MIT licensed.

RESEARCH.md

@@ -52,3 +52,16 @@ change-review case that ordinary formula text comparison misses: the stored
 formula can remain unchanged while the Table's stored range grows. WCAB schema
 version 2 therefore adds a generated Table-expansion case that checks those
 observable facts without evaluating a formula or asserting a result value.
+
+## Evidence scoring without a false-positive fiction
+
+The [NIST static-analysis methodology](https://www.nist.gov/system/files/documents/2021/03/24/CAS%202012%20Static%20Analysis%20Tool%20Study%20Methodology.pdf)
+defines precision as reported true findings divided by all reported findings,
+and recall as reported true findings divided by all known findings. WCAB's
+truth is intentionally not a complete list of every difference in a workbook:
+it declares review-relevant facts plus coverage boundaries. Calling every
+unlisted tool observation a false positive would therefore misrepresent the
+oracle. Version 0.3 adds a normalized adapter protocol that measures recall of
+declared facts, analyzed coverage, and reference-policy agreement, while
+preserving unrecognized observations for review instead of forcing a precision
+claim.

docs/observations.md

@@ -0,0 +1,116 @@
+# WCAB normalized observation protocol, version 1
+
+WCAB evaluates tools that inspect a baseline/candidate workbook change. Native
+tool reports differ, so an adapter should translate its result into this small
+JSON protocol before invoking `wcab score`. The protocol is deliberately
+separate from the fixture truth schema: a tool may keep its own report format,
+but its benchmark adapter has one portable target.
+
+## Start with a template
+
+```bash
+wcab observation-template --fixtures fixtures --output observations.json
+# Run an adapter that replaces unsupported entries with observed evidence.
+wcab score --fixtures fixtures --observations observations.json \
+  --output score.json
+```
+
+The template is valid and marks every case `unsupported`. That is intentional:
+unsupported coverage is visible in the score rather than being treated as a
+pass.
+
+## Document shape
+
+```json
+{
+  "schema_version": 1,
+  "benchmark": {"fixture_schema_versions": [2]},
+  "tool": {"name": "example-tool", "version": "1.4.0"},
+  "cases": [
+    {
+      "id": "finance.formula_to_value",
+      "status": "analyzed",
+      "facts": [
+        {
+          "fact": {
+            "kind": "formula_to_value",
+            "sheet": "Revenue",
+            "cell": "C8"
+          },
+          "evidence": {"native_rule": "hardcode-replacement"}
+        }
+      ],
+      "review": "block",
+      "coverage": []
+    }
+  ]
+}
+```
+
+`schema_version` is the observation protocol version. `fixture_schema_versions`
+must name the truth-schema versions the adapter analyzed. `tool.name` is
+required; `tool.version` is optional.
+
+Each `facts` item wraps an exact WCAB truth fact in `fact`. The optional
+`evidence` object is adapter-owned metadata and is never scored; it can record a
+native rule ID, report location, or another reproducible pointer without making
+one tool's report schema normative.
+
+## Case status and review
+
+| `status` | Required behavior | Score treatment |
+| --- | --- | --- |
+| `analyzed` | May report facts and one optional `review` disposition. | Its recognized facts can contribute to expected-fact recall. |
+| `unsupported` | Must not report facts or a review. | Explicitly counted as unsupported, never as a pass. |
+| `error` | Must include a non-empty `error` string and must not report facts or a review. | Counted separately from unsupported coverage. |
+
+`review` is `allow`, `review`, `block`, or `null`. A non-null disposition is
+compared with WCAB's reference convention and reported as **reference-policy
+agreement**. It is not a claim that every organization must use the same
+approval policy.
+
+Omitting a known case is permitted for partial runs, but it is reported as
+`not_reported` and cannot contribute to coverage or agreement.
+
+## What the score means
+
+The scorer returns deterministic per-case details plus these aggregate metrics:
+
+| Metric | Meaning |
+| --- | --- |
+| `fact_recall` | Matched declared facts divided by all declared benchmark facts. |
+| `analysis_coverage` | Cases marked `analyzed` divided by all WCAB cases. |
+| `review_agreement` | Cases whose analyzed review matches the reference convention divided by all WCAB cases. |
+| `complete_case_count` | Analyzed cases with every declared fact matched, no unrecognized fact, and matching reference review. |
+| `unrecognized_fact_count` | Reported normalized facts that are not one of the case's declared facts. |
+
+WCAB intentionally declares targeted observable facts and explicit boundaries;
+it does **not** claim an exhaustive inventory of every workbook difference.
+For that reason, the scorer does not call unrecognized facts false positives
+and does not publish a precision metric. They remain reviewable evidence rather
+than a silent score adjustment. This follows the basic distinction between
+precision and recall used in the [NIST static-analysis study methodology](https://www.nist.gov/system/files/documents/2021/03/24/CAS%202012%20Static%20Analysis%20Tool%20Study%20Methodology.pdf),
+while respecting WCAB's deliberately partial oracle.
+
+Use `--strict` when a complete reference-policy run is required. It exits
+nonzero unless every case is complete and no unrecognized normalized fact was
+reported.
+
+## Optional FormulaFence reference adapter
+
+FormulaFence remains optional and is not a package dependency. If a trusted
+local executable is installed, WCAB can emit a normalized report from its
+existing adapter mapping:
+
+```bash
+wcab formulafence-observations --fixtures fixtures \
+  --executable formulafence --output formulafence-observations.json
+wcab score --fixtures fixtures --observations formulafence-observations.json
+```
+
+The adapter copies a WCAB fact into the normalized report only after it matches
+the documented native FormulaFence change/rule category, which is preserved in
+the optional `evidence` object. It leaves `review` as `null`: FormulaFence
+surfaces evidence but does not impose WCAB's reference policy as a universal
+approval decision. Any intentionally unmapped WCAB fact becomes adapter
+coverage text and remains missing from expected-fact recall.

docs/validation.md

@@ -1,6 +1,6 @@
 # Validation record
 
-This record describes the WCAB 0.2.0 / schema-version-2 validation run on
+This record describes the WCAB 0.3.0 / schema-version-2 validation run on
 2026-08-02. It is reproducible from this repository; no network service or
 private workbook is required.
 
@@ -19,6 +19,11 @@ Commands:
 python -m pip install -e '.[dev]'
 wcab build --output fixtures
 wcab manifest --fixtures fixtures
+wcab observation-template --fixtures fixtures --output /tmp/observations.json
+wcab score --fixtures fixtures --observations /tmp/observations.json
+wcab formulafence-observations --fixtures fixtures \
+  --executable formulafence --output /tmp/formulafence-observations.json
+wcab score --fixtures fixtures --observations /tmp/formulafence-observations.json
 ruff check .
 ruff format --check .
 pytest
@@ -31,27 +36,32 @@ Results:
 - 19 observable truth facts: 12 `block` and five `review` dispositions.
 - 53 workbook and truth-manifest fixture files, plus one generated JSONL case
   catalogue, all generated from source.
-- Eleven unit tests passed under both Python versions, including independent
+- Eighteen unit tests passed under both Python versions, including independent
   regeneration and byte-for-byte fixture-tree equality.
 - The fixture validator accepted all 17 cases.
 
 ## Distribution supplement
 
-The 0.2.0 release builds a one-row-per-case `manifest.jsonl` catalogue. Each
-row retains the schema-version-2 truth contract and includes byte counts and
-SHA-256 digests for the workbooks it names.
+The 0.3.0 release builds a one-row-per-case `manifest.jsonl` catalogue and a
+tool-neutral observation protocol. Each catalogue row retains the
+schema-version-2 truth contract and includes byte counts and SHA-256 digests
+for the workbooks it names.
 
 Commands:
 
 ```bash
 python -m build
-twine check dist/workbook_change_benchmark-0.2.0*
+twine check dist/workbook_change_benchmark-0.3.0*
 python -m venv /tmp/wcab-wheel-test
 /tmp/wcab-wheel-test/bin/python -m pip install \
-  dist/workbook_change_benchmark-0.2.0-py3-none-any.whl
+  dist/workbook_change_benchmark-0.3.0-py3-none-any.whl
 /tmp/wcab-wheel-test/bin/wcab validate --fixtures fixtures
 /tmp/wcab-wheel-test/bin/wcab manifest --fixtures fixtures --output /tmp/manifest.jsonl
 cmp fixtures/manifest.jsonl /tmp/manifest.jsonl
+/tmp/wcab-wheel-test/bin/wcab observation-template --fixtures fixtures \
+  --output /tmp/observations.json
+/tmp/wcab-wheel-test/bin/wcab score --fixtures fixtures \
+  --observations /tmp/observations.json
 ```
 
 Results:
@@ -59,8 +69,12 @@ Results:
 - The source distribution and universal wheel passed `twine check`.
 - A fresh Python 3.12 wheel installation validated all 17 fixtures and emitted
   byte-identical JSONL output.
-- The full eleven-test suite, lint, and format checks passed under the supported
+- The full eighteen-test suite, lint, and format checks passed under the supported
   local Python versions (3.12 and 3.13).
+- The generated unsupported template scored as zero analyzed coverage and zero
+  expected-fact recall, confirming that unsupported cases cannot become a pass.
+- The FormulaFence normalizer emitted 18 matched facts, one intentionally
+  unmapped fact, and no invented review disposition.
 
 ## FormulaFence reference adapter