diff --git a/config/tasks.json b/config/tasks.json
index 444a2e8..96ae008 100644
--- a/config/tasks.json
+++ b/config/tasks.json
@@ -1,4 +1,7 @@
 {
   "phase0-private-repo-smoke-test": "tasks/phase0-private-repo-smoke-test.md",
-  "reconcile-repository-state": "tasks/reconcile-repository-state.md"
+  "reconcile-repository-state": "tasks/reconcile-repository-state.md",
+  "qualify-exact-head": "tasks/qualify-exact-head.md",
+  "advance-cloud-safe-work": "tasks/advance-cloud-safe-work.md",
+  "reduce-experiment-evidence": "tasks/reduce-experiment-evidence.md"
 }
diff --git a/tasks/advance-cloud-safe-work.md b/tasks/advance-cloud-safe-work.md
new file mode 100644
index 0000000..0f9d3ce
--- /dev/null
+++ b/tasks/advance-cloud-safe-work.md
@@ -0,0 +1,35 @@
+# Advance cloud-safe work
+
+## Objective
+
+Advance the next bounded unit of work that the target repository's durable authority explicitly identifies as safe and useful in the current hosted execution environment.
+
+## Responsibilities
+
+- Read repository-local instructions and the active issue/PR/workset before acting.
+- Recover the current accepted state, UNKNOWNs, next bounded action, and reserved authority from repository-native durable state.
+- Confirm the hosted environment satisfies the hard capability requirements before mutation.
+- Prefer existing native, standard, maintained open-source, or already-local capabilities over custom machinery.
+- Implement the smallest change that advances the authorized cloud-safe frontier.
+- Run applicable repository-provided validation.
+- Preserve explicit boundaries for work that requires another executor, privilege level, secret, device, interactive session, paid resource, or human acceptance.
+
+## Constraints
+
+- Do not infer authority from authentication or tool availability.
+- Do not merge, release, publish, deploy to production, rotate/change secrets, change visibility, spend money, or perform destructive host/repository actions.
+- Do not claim native-Windows or other executor-specific acceptance unless the target authority explicitly identifies the current executor as suitable for that evidence class.
+- Do not invent a scheduler, queue, database, registry, router, callback service, orchestration layer, or broad abstraction to avoid an ordinary boundary.
+- Do not weaken validation to obtain a pass.
+- Do not broaden into unrelated refactors or cleanup.
+
+## Completion criteria
+
+- One bounded cloud-safe increment is implemented or a concrete capability/authority blocker is established.
+- Existing relevant validation passes, or failures are classified.
+- The resulting change is reviewable and reversible through Git.
+- Any remaining native/local/human boundary is explicit.
+
+## Result contract
+
+Return one narrowly scoped pull request when mutation is warranted. In the PR describe the authoritative work item, the bounded increment advanced, validation performed, evidence class produced, and work deliberately left for another executor/authority. Do not duplicate routine evidence already expressed by authoritative CI/result state.
diff --git a/tasks/qualify-exact-head.md b/tasks/qualify-exact-head.md
new file mode 100644
index 0000000..6bf4de6
--- /dev/null
+++ b/tasks/qualify-exact-head.md
@@ -0,0 +1,36 @@
+# Exact-head qualification
+
+## Objective
+
+Qualify the target repository's current bounded candidate at the exact Git commit/PR head identified by the repository's own durable authority.
+
+## Responsibilities
+
+- Read repository-local agent/contributor guidance and the active issue/PR/workset first.
+- Resolve the exact candidate SHA from authoritative repository state; do not reuse evidence from an older head.
+- Run only repository-provided deterministic checks, tests, linters, parsers, contract validators, or build steps that are applicable in the hosted environment.
+- Distinguish product failure, harness/test defect, unavailable capability, transient infrastructure failure, and evidence that requires a different executor.
+- Prefer existing native/project validation commands over new helpers.
+- If a small cloud-visible defect prevents the existing qualification path from running, repair only that defect when the target authority permits ordinary repository mutation and validate the repair.
+- Keep acceptance claims within the evidence class actually produced.
+
+## Constraints
+
+- Do not merge, release, publish, deploy, change secrets, change repository visibility, spend money, or perform destructive cleanup.
+- Do not claim native-Windows, interactive-desktop, driver/device, hardware, or other executor-specific acceptance from a hosted environment that cannot produce that evidence.
+- Do not create a new test framework, scheduler, registry, orchestration service, or broad compatibility layer merely to obtain a result.
+- Do not broaden scope into adjacent cleanup.
+- Do not treat agent/provider "completed" state as qualification evidence.
+- If the exact head changes during the task, stop and classify prior results as stale rather than silently applying them to the new head.
+
+## Completion criteria
+
+- The exact candidate head is identified.
+- Applicable deterministic qualification has run against that exact head.
+- Any failure is classified with the smallest useful evidence.
+- Any repair is minimal and revalidated.
+- Remaining executor/authority boundaries are explicit.
+
+## Result contract
+
+If repository changes are required, return one narrowly scoped pull request. Otherwise return the exact head, commands/checks run, result classification, and the next unresolved evidence boundary through the target repository's normal durable result surface when one exists. Record only evidence that changes a decision or is required to establish exact-head qualification.
diff --git a/tasks/reduce-experiment-evidence.md b/tasks/reduce-experiment-evidence.md
new file mode 100644
index 0000000..3706e6d
--- /dev/null
+++ b/tasks/reduce-experiment-evidence.md
@@ -0,0 +1,34 @@
+# Reduce experiment evidence
+
+## Objective
+
+Reduce existing experiment, CI, Actions, test, and repository evidence into the smallest decision-relevant update to the target repository's current disposition.
+
+## Responsibilities
+
+- Read repository-local guidance and the active authority first.
+- Use existing receipts, logs, artifacts, commits, PRs, issues, and tests; do not rerun experiments merely to create more data unless the authority explicitly requires it.
+- Separate advertised, installed, callable, exercised, oracle-satisfied, and workload-relevant evidence where those distinctions matter.
+- Identify only evidence that changes a preserve/replace/retire/defer decision, closes/reopens a gate, invalidates stale evidence, or establishes a meaningful blocker.
+- Preserve exact provenance when the consequence requires it.
+- Prefer editing the repository's existing authority/evidence surface over creating a new ledger or report.
+
+## Constraints
+
+- Review/evidence reduction only unless a tiny documentation/evidence correction is explicitly part of the target authority.
+- Do not change product/runtime behavior, workflows, dependencies, secrets, release state, or infrastructure.
+- Do not infer success from a green wrapper job when the required inner oracle did not execute.
+- Do not infer native-Windows or other executor-specific acceptance from weaker evidence.
+- Do not create dashboards, databases, registries, or recurring reporting machinery.
+- Do not duplicate routine successful execution evidence.
+
+## Completion criteria
+
+- Material evidence is reconciled against current exact state.
+- Stale or weaker evidence is clearly bounded.
+- Any disposition or gate change is explicit and supported.
+- If nothing decision-relevant changed, no new durable record is created.
+
+## Result contract
+
+When a decision-relevant change exists, update the target repository's existing governing issue/PR/document with the minimum supporting evidence. If mutation requires a PR, return one narrowly scoped PR. If no decision-relevant change exists, make no repository change.