SollanSystems · SollanSystems · Jul 10, 2026 · Jul 10, 2026
diff --git a/docs/superpowers/plans/2026-07-08-v0.8.0-composes-the-field.md b/docs/superpowers/plans/2026-07-08-v0.8.0-composes-the-field.md
@@ -14,7 +14,7 @@
 
 - **Two PRs:** PR-A = ST3 (Tasks 1–4), PR-B = ST4 + release cut (Tasks 5–10). PR-B branches from main after PR-A merges. The 0.8.0 release commit is the **last** commit of PR-B.
 - **Paths are the shipped reality, not the 2026-06-30 spec's:** helper = `loop/integrations.py` (NOT `loop_engineer.integrations`); docs = `docs/integrations/<engine>.md` (NOT `reference/integrations/`).
-- **All disk writes go through `loop.emit`** (`open_contract` / `append_iteration` / `append_receipt` / `terminate`). No second write path for contract artifacts. (Recipes may write their own verify-bundle/verdict JSONs under `.loop/artifacts/` — those are evidence artifacts, not contract objects.) *Adjudicated carve-out (PR-A whole-branch review, 2026-07-08):* both recipes reset `<ws>/RUNLOG.md` to emit's own fresh header immediately after `open_contract`, because the scaffold seeds an unfilled `{{ITERATION_OUTCOME}}` placeholder that `loop metrics` flags as an unrecognized outcome token and the root fix (templates/scaffold/emit) is outside PR-A's file discipline. The reset strips a placeholder, never fabricates state. Root-fix affordance (e.g. `emit.open_contract(seed_runlog=False)`) is tracked as a PR-B contributor issue.
+- **All disk writes go through `loop.emit`** (`open_contract` / `append_iteration` / `append_receipt` / `terminate`). No second write path for contract artifacts. (Recipes may write their own verify-bundle/verdict JSONs under `.loop/artifacts/` — those are evidence artifacts, not contract objects.) *Adjudicated carve-out (PR-A whole-branch review, 2026-07-08):* both recipes reset `<ws>/RUNLOG.md` to emit's own fresh header immediately after `open_contract`, because the scaffold seeds an unfilled `{{ITERATION_OUTCOME}}` placeholder that `loop metrics` flags as an unrecognized outcome token and the root fix (templates/scaffold/emit) is outside PR-A's file discipline. The reset strips a placeholder, never fabricates state. Root-fix affordance (e.g. `emit.open_contract(seed_runlog=False)`) is tracked as a PR-B contributor issue. **[Retired 2026-07-09 by the #40 root fix — `templates/RUNLOG.md.tmpl` seeds no `## Iteration` block, so a fresh scaffold is metrics-clean by construction; both recipe resets are removed and no `seed_runlog` kwarg was needed.]**
 - **PR-A file discipline:** `loop/` gains ONLY `integrations.py`. No modification to `scripts/holdout_gate.py`, `scripts/anticheat_scan.py`, `scripts/inspect_loop.py`, `scripts/metrics.py`, `scripts/self_eval.py`, `scripts/validate_frontmatter.py`, or anything under `schemas/`, `templates/`, `evals/`. `scripts/test_langgraph_recipe.py` stays **byte-identical** (new assertions go in a NEW file `scripts/test_langgraph_recipe_st3.py`). New test files under `scripts/` are additions, allowed.
 - **PR-B scorer discipline:** `scripts/inspect_loop.py` may change ONLY (a) the path-resolution seam (3 `resolve_loop_paths(...)` call sites → `_resolve_paths(...)`) and (b) additive report labeling (`foreign_layout`, `advisory`). Zero changes to `_CHECKS` weights, credit tiers, signal regexes, the score cap, or `_verdict`. A regression test pins that the fixture earns no invoked credit.
 - **Fixed precedence** in `to_terminal_state`: safety → human → blocked → budget → spec-gap → gate verdict. `FailedSafety` and `AbortedByHuman` beat a green gate.

diff --git a/examples/langgraph-emit/graph_example.py b/examples/langgraph-emit/graph_example.py
@@ -99,12 +99,6 @@ def certify(state: State) -> dict:
 
 def main(workspace: str, sabotage: bool) -> int:
     emit.open_contract(workspace)
-    # The fresh scaffold's RUNLOG carries a placeholder example iteration (its
-    # unfilled outcome renders as "REPLACE"); start the demo's run history clean
-    # so `loop metrics` scores only the graph's own iterations.
-    (Path(workspace) / "RUNLOG.md").write_text(
-        f"# RUNLOG.md — {Path(workspace).name}\n", encoding="utf-8"
-    )
     graph = (
         StateGraph(State)
         .add_node(do_work)

diff --git a/examples/temporal-certify/workflow_example.py b/examples/temporal-certify/workflow_example.py
@@ -153,12 +153,6 @@ def certify_workflow_failure(workspace: str, cause: BaseException | None) -> dic
 
 async def run_and_certify(client: Client, workspace: str, *, sabotage: bool, wf_id: str) -> dict:
     emit.open_contract(workspace)
-    # The fresh scaffold's RUNLOG carries a placeholder example iteration (its
-    # unfilled outcome renders as "REPLACE"); start the demo's run history clean
-    # so `loop metrics` scores only the workflow's own iterations.
-    (Path(workspace) / "RUNLOG.md").write_text(
-        f"# RUNLOG.md — {Path(workspace).name}\n", encoding="utf-8"
-    )
     try:
         return await client.execute_workflow(
             CertifiedGoalWorkflow.run, WorkArgs(workspace=workspace, sabotage=sabotage),

diff --git a/scripts/test_emit.py b/scripts/test_emit.py
@@ -8,6 +8,7 @@
 
 import pytest
 
+import metrics
 from loop import emit
 from loop.contract import validate_contract
 
@@ -24,6 +25,14 @@ def test_open_contract_is_doctor_clean(workspace):
     assert validate_contract(workspace)["ok"] is True
 
 
+def test_open_contract_is_metrics_clean(workspace):
+    # The seeded RUNLOG is reference-only prose — no `## Iteration` block, so a
+    # fresh scaffold must score zero iterations and no outcome tokens.
+    scorecard = metrics.compute_metrics(workspace)
+    assert scorecard["provenance"]["unrecognized_outcomes"] == []
+    assert scorecard["iterations_claiming_success"] == 0
+
+
 def test_append_iteration_writes_parseable_runlog_and_updates_state(workspace):
     runlog = emit.append_iteration(
         workspace, iteration_id=1, outcome="task_passed", task_id="T1",

diff --git a/scripts/test_template_roundtrip.py b/scripts/test_template_roundtrip.py
@@ -90,21 +90,8 @@ def _fill_values() -> dict[str, str]:
         "ARTIFACT_PATH": ".loop/artifacts/verify-T1.json",
         "FALSE_COMPLETION": "false",     # boolean
         "TERMINAL_REASON": "fast gate passed with evidence",
-        # --- RUNLOG.md prose ---
-        "ITERATION_DATE": "2026-06-30",
-        "STATE_NAME": "running",
-        "ACTION_1": "ran the fast gate",
-        "ACTION_2": "recorded evidence",
-        "VERIFY_GATE": "verify-fast",
-        "VERIFY_OUTCOME": "PASS",
-        "VERIFY_CMD": "scripts/verify-fast",
-        "SCORE_BEFORE": "0.0",
-        "SCORE_AFTER": "1.0",
-        "ITERATION_OUTCOME": "task_passed",
-        "REPAIR_ACTION": "n/a",
-        "REPAIR_ATTEMPT": "0",
-        "IMPROVEMENT_FLAG": "n/a",
-        "ITERATION_NOTES": "clean first pass",
+        # RUNLOG.md seeds a reference-only preamble that declares only
+        # {{PROJECT_NAME}}, so it needs no per-iteration prose fill.
         # --- SPEC / WORKFLOW / AGENTS / EVALS prose ---
         "NON_GOAL_1": "no production writes",
         "NON_GOAL_2": "no schema changes",

diff --git a/templates/RUNLOG.md.tmpl b/templates/RUNLOG.md.tmpl
@@ -2,41 +2,30 @@
 
 > Human-readable iteration history. Append-only — one entry per loop iteration.
 > Machine state lives in `.loop/state.json`; this is the audit trail.
+>
+> This file starts with no iterations. `loop metrics` reads only `## Iteration`
+> blocks, so a fresh RUNLOG scores clean by construction. The entry format below
+> is a reference, not an iteration — `loop.emit.append_iteration()` writes real
+> blocks for you (or copy the shape by hand) beneath the marker at the bottom.
 
 ---
 
-## Iteration {{ITERATION_ID}} — {{ITERATION_DATE}}
+## Entry format (reference — not an iteration, not scored)
 
-**State entered:** `{{STATE_NAME}}`
-**Active task:** `{{ACTIVE_TASK_ID}}` — {{TASK_TITLE}}
-**Plan version:** {{PLAN_VERSION}}
+Each iteration is a section headed `## Iteration <id> — <date>` carrying:
 
-### Actions taken
-
-- {{ACTION_1}}
-- {{ACTION_2}}
-
-### Verification result
-
-- **Gate:** `{{VERIFY_GATE}}` — {{VERIFY_OUTCOME}} (`{{VERIFY_CMD}}`)
-- **Score before:** {{SCORE_BEFORE}}
-- **Score after:** {{SCORE_AFTER}}
-
-### Outcome
-
-`{{ITERATION_OUTCOME}}` — one of: `task_passed`, `task_failed`, `repair_triggered`, `approval_requested`, `replanned`, `terminal`
-
-### Repair record (if repair triggered)
-
-- **Failure mode:** {{FAILURE_MODE}}
-- **Repair action:** {{REPAIR_ACTION}}
-- **Attempt:** {{REPAIR_ATTEMPT}} of {{REPAIR_CAP}}
-- **Measurable improvement:** {{IMPROVEMENT_FLAG}}
-
-### Notes
-
-{{ITERATION_NOTES}}
+- **State entered / Active task / Plan version** — the loop state, the task
+  worked, and the plan version in force.
+- **Actions taken** — a bullet list of what the agent did.
+- **Verification result** — the gate (`scripts/verify-*`), its outcome, and the
+  score before/after.
+- **Outcome** — exactly one backticked token, one of: `task_passed`,
+  `task_failed`, `repair_triggered`, `approval_requested`, `replanned`,
+  `terminal`.
+- **Repair record** — failure mode, repair action, attempt N of the cap, and the
+  measurable-improvement flag (only when a repair was triggered).
+- **Notes** — free-form context.
 
 ---
 
-<!-- Add new iterations above this line. Do not edit past entries. -->
+<!-- Append new `## Iteration <id>` blocks below this line. Do not edit past entries. -->