Add inactive evidence-integrity evaluator

README.md

@@ -148,6 +148,30 @@ credential-like environment value, grants no authority or qualification, activat
 no profile, and performs no candidate, adapter, network, publish, deploy, or
 external-write action.
 
+## Inactive evidence-integrity evaluator
+
+`control/v1/evaluate-evidence-integrity.sh` compares caller-supplied evidence
+references with one exact public-core stage tuple. It binds the policy set,
+policy, decision, evaluator, and selected core package before returning a
+canonical `satisfied` or `violated` identity observation. Evidence and prior
+references must keep their canonical order and unique logical identities; one
+prior result digest cannot describe multiple result documents.
+
+The launcher is an explicit trusted shell boundary; it does not claim to
+self-attest bytes that Bash already loaded. The decision separately binds the
+exact marked evaluation payload. The launcher extracts those fragments from a
+private no-follow snapshot, verifies their content identity, and gives only the
+verified bytes to the worker. Scratch producers return full final descriptor
+identities over an inherited, unlinked channel; consumers bind every read to
+those identities. Public-core validation uses its accounted mode
+with a fixed budget, an anchored receipt descriptor, and a worker-owned root.
+
+The evaluator never reads proof bytes. Matching references do not prove a claim
+or qualify a workflow, and equal proof digests may belong to different logical
+references. The package stays inactive, stores nothing, grants no authority, and
+performs no candidate, credential, network, adapter, publish, deploy, or external
+write.
+
 ## The current default team
 
 You talk **only** to yshifu, in a Claude Code session. yshifu orchestrates the other roles

RESTORE.md

@@ -394,6 +394,22 @@ does not read credential material or credential-like environment values, qualify
 claim, grant authority, activate a profile, run a candidate or adapter, or perform
 a network or external-write action.
 
+Restore the five paths in the manifest's inactive evidence-integrity block, then
+run:
+
+```sh
+bash scripts/test/control-evidence-integrity.test.sh
+```
+
+This checks exact policy-set and public-core closure, stage and qualification
+identity binding, canonical evidence/prior sets, stale or aliased references, and
+deterministic observation output. It also checks the explicit trusted-launcher
+boundary, exact marked-payload identity, producer-final scratch identities,
+accounted core receipt and cleanup, and nested Bash environment isolation. The
+launcher is not self-attested. It does not read proof bytes, establish proof
+truth, qualify a workflow, store evidence, grant authority, activate a profile,
+run a candidate or adapter, or perform a network or external-write action.
+
 ---
 
 ## 5. Smoke test — prove the rebuilt team is alive

ci/required-files.txt

@@ -206,3 +206,10 @@ control/v1/credential-policy-decision.json
 control/v1/credential-policy.jq
 control/v1/evaluate-credential-policy.sh
 scripts/test/control-credential-policy.test.sh
+
+# Inactive evidence integrity policy and evaluator
+control/v1/evidence-integrity-policy.json
+control/v1/evidence-integrity-decision.json
+control/v1/evidence-integrity.jq
+control/v1/evaluate-evidence-integrity.sh
+scripts/test/control-evidence-integrity.test.sh