Release 0.3.0 under compatibility contract evidence-v3.2

CHANGELOG.md

@@ -0,0 +1,41 @@
+# Changelog
+
+Each entry names the verifier `VERSION` and its compatibility contract. The
+contract is what decides composition: records carrying a contract other than the
+verifier's own are refused (`docs/design/evidence-v3.md` §J). A release that
+changes what admissible means increments it; a patch that does not, keeps it.
+
+## 0.3.0 — contract `evidence-v3.2`
+
+The same evidence/3 record shape as `evidence-v3.1`, judged under stricter
+rules. Records written under `evidence-v3.1` no longer compose with this
+verifier: re-run `run`, `attest` and `import-ci` to re-establish them.
+
+What changed what admissible means:
+
+- **Execution binds to the state it measured** (#3). `run` captures the target
+  before the gates execute and records any dimension that moved while they ran.
+  A drifted record, or one whose own verdict was BLOCKED or FAIL, is refused at
+  composition rather than rehabilitated by the passing rows inside it.
+- **CI evidence binds the whole provenance chain** (#4). Repository, run,
+  attempt, workflow, job and commit are checked against expectations the policy
+  declares in `[ci]`, never against values read from the payload under
+  examination. `import-ci` refuses a chain it cannot bind before writing any
+  record, and composition checks the chain again.
+- **The adopted contract is committed and fails closed** (#5). A project adopts
+  evidence/3 with `evidence = "verification.ladder.evidence/3"` in its policy
+  *and* an activation record in the checkout. A damaged activation record, an
+  unrecognized contract, or an activation the policy does not carry is refused
+  rather than read as evidence/2. Qualification is scoped to the predicates the
+  consumer's own policy puts in play.
+
+Installation (#2): `scripts/install-ladder.sh` and `scripts/check-install.sh`
+install and check a pinned checkout. **Their pin still names `760977f`**, which
+predates everything above and still reports `evidence-v3.1`. Moving the pin to
+this release is a separate change, because the installer refuses a pin that is
+not already on `main`.
+
+## 0.2.0 — contract `evidence-v3.1`
+
+evidence/3 authority becomes available behind qualification and activation
+(`docs/design/evidence-v3.md` §M4).

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "verification-ladder"
-version = "0.2.0"
+version = "0.3.0"
 description = "A portable verification procedure and evidence machinery for coding agents"
 readme = "README.md"
 requires-python = ">=3.11"

skills/verification-ladder/bin/verify.py

@@ -60,8 +60,14 @@
 # Provenance says which build produced a record; the contract says which records
 # may compose. They are separate: a verifier can keep the v3 record shape and
 # change what admissible means, and that must invalidate earlier evidence.
-COMPATIBILITY = "evidence-v3.1"
-VERSION = "0.2.0"  # tracks pyproject's version; tests/test_schema_v3.py holds them equal
+#
+# evidence-v3.2 (0.3.0): drifted and refused records are rejected at composition,
+# CI evidence binds repository -> run -> attempt -> workflow -> job -> commit
+# against policy-declared expectations, and the adopted contract is committed
+# and fails closed. evidence-v3.1 was the same record shape under weaker rules,
+# so its records must not compose with these. CHANGELOG.md says what moved.
+COMPATIBILITY = "evidence-v3.2"
+VERSION = "0.3.0"  # tracks pyproject's version; tests/test_schema_v3.py holds them equal
 # A gate declares the target dimensions it depends on. Evidence must carry every
 # dimension its gate declares, and dimensions it does not declare are never
 # compared - so a moved runtime expires a behavioural gate while a lint gate,

tests/test_v3_adoption.py

@@ -466,6 +466,26 @@ def test_check_still_binds_a_record_to_the_runtime_it_names(tmp_path):
         "only the runtime moved"
 
 
+def test_composition_refuses_a_record_from_the_pre_hardening_contract(repo):
+    """Through `compose`: a record stamped `evidence-v3.1` does not compose here.
+
+    That is the contract the previously pinned verifier writes. Nothing else
+    about this record is wrong - it is this checkout's own run, re-stamped - so
+    the refusal can only come from the contract.
+    """
+    activate(repo)
+    local = repo / ".verification" / "v3.json"
+    assert cli(repo, "run", "--output", str(local))[0] == 0
+    record = json.loads(local.read_text())
+    record["verifier"]["compatibility"] = "evidence-v3.1"
+    local.write_text(json.dumps(record))
+
+    code, output = cli(repo, "compose", str(local))
+    assert code != 0, output
+    assert "produced under contract 'evidence-v3.1'" in output, output
+    assert "READY FOR HUMAN GATE     TRUE" not in output
+
+
 def test_check_refuses_a_shadow_record(repo):
     """A shadow record binds to nothing, so there is nothing to re-bind."""
     pre_activation_evidence(repo)

tests/test_verifier_contract.py

@@ -48,6 +48,38 @@ def test_a_contract_this_verifier_does_not_apply_is_refused():
     assert "the rules in force" in reason
 
 
+def test_records_from_the_pre_hardening_contract_are_refused():
+    """evidence-v3.1 is the contract the pre-hardening verifier still stamps.
+
+    Its records have the same shape as these and were judged under weaker rules:
+    no drift rejection, no bound CI chain, an adoption that fell back to
+    evidence/2 on a parse failure. The contract was incremented so that they
+    stop composing here, not merely so that the number moved.
+    """
+    assert verify.COMPATIBILITY != "evidence-v3.1"
+    status, reason = verify.verifier_status([_record(compatibility="evidence-v3.1")])
+    assert status == verify.INADMISSIBLE
+    assert "'evidence-v3.1'" in reason and "the rules in force" in reason
+
+    status, reason = verify.verifier_status([_record(), _record(compatibility="evidence-v3.1")])
+    assert status == verify.INADMISSIBLE
+    assert "different verifier contracts" in reason
+
+
+def test_the_release_notes_name_the_contract_this_verifier_applies():
+    """A release that changes the contract says so in its notes (§J).
+
+    Held mechanically, like VERSION against pyproject: the newest CHANGELOG entry
+    must name this verifier's VERSION and COMPATIBILITY, so neither can move
+    without the notes moving with it.
+    """
+    changelog = (_P(__file__).resolve().parents[1] / "CHANGELOG.md").read_text()
+    newest = re.search(r"^## (\S+) \S+ contract `([^`]+)`", changelog, re.MULTILINE)
+    assert newest, "CHANGELOG.md has no '## <version> - contract `<contract>`' entry"
+    assert newest.group(1) == verify.VERSION
+    assert newest.group(2) == verify.COMPATIBILITY
+
+
 def test_a_record_without_a_contract_establishes_nothing():
     status, reason = verify.verifier_status([{"verifier": {"commit": "a" * 40}}])
     assert status == verify.INADMISSIBLE