eval: compare st2-managed and harness-native delegation across 12 matched cells

.gitattributes

@@ -0,0 +1,3 @@
+# A unified diff legitimately carries a bare space on blank context lines; the corpus whitespace gate must
+# not read the reviewed proposal fixture as an authoring defect.
+cells/**/review/*.patch -whitespace

CATALOG.md

@@ -1,8 +1,8 @@
 <!-- Generated by bin/generate-catalog.sh from cell KDL and evidence/*.tsv. Do not edit. -->
 # Current eval catalog and evidence
 
-This is the sole current corpus surface: **59 included cells** (19 model-free,
-40 model-backed) and **4 retired exclusions**. Inventory fields are derived from the active
+This is the sole current corpus surface: **71 included cells** (19 model-free,
+52 model-backed) and **4 retired exclusions**. Inventory fields are derived from the active
 KDL and executable judge scripts. Latest accepted PASS evidence remains distinct from the append-only run
 history: a failed last run is visible without being advertised as accepted evidence, and cells with no
 structured run row say so explicitly.
@@ -29,6 +29,18 @@ harness-native loader plus canonical hook file.
 | `assignment-contract-hot-resources` | Codex | `gpt-5.6-sol` / medium | 2 | medium | `1200s` | 4 | — | **NO STRUCTURED RUN** |
 | `context-resource-continuity` | model-free | — | 0 | none | `90s` | 4 | — | **NO STRUCTURED RUN** |
 | `crash-ding` | mixed | `claude-sonnet-5+gpt-5.6-sol` / medium | 4 | high | `180s` | 5 | — | **NO STRUCTURED RUN** |
+| `delegation-implement-claude-native` | Claude | `claude-sonnet-5` / medium | 1 | low | `900s` | 7 | — | **NO STRUCTURED RUN** |
+| `delegation-implement-claude-st2` | Claude | `claude-sonnet-5` / medium | 2 | medium | `900s` | 7 | — | **NO STRUCTURED RUN** |
+| `delegation-implement-codex-native` | Codex | `gpt-5.6-sol` / medium | 1 | low | `900s` | 7 | — | **NO STRUCTURED RUN** |
+| `delegation-implement-codex-st2` | Codex | `gpt-5.6-sol` / medium | 2 | medium | `900s` | 7 | — | **NO STRUCTURED RUN** |
+| `delegation-review-claude-native` | Claude | `claude-sonnet-5` / medium | 1 | low | `900s` | 5 | — | **NO STRUCTURED RUN** |
+| `delegation-review-claude-st2` | Claude | `claude-sonnet-5` / medium | 3 | high | `900s` | 5 | — | **NO STRUCTURED RUN** |
+| `delegation-review-codex-native` | Codex | `gpt-5.6-sol` / medium | 1 | low | `900s` | 5 | — | **NO STRUCTURED RUN** |
+| `delegation-review-codex-st2` | Codex | `gpt-5.6-sol` / medium | 3 | high | `900s` | 5 | — | **NO STRUCTURED RUN** |
+| `delegation-sweep-claude-native` | Claude | `claude-sonnet-5` / medium | 1 | low | `900s` | 5 | — | **NO STRUCTURED RUN** |
+| `delegation-sweep-claude-st2` | Claude | `claude-sonnet-5` / medium | 3 | high | `900s` | 5 | — | **NO STRUCTURED RUN** |
+| `delegation-sweep-codex-native` | Codex | `gpt-5.6-sol` / medium | 1 | low | `900s` | 5 | — | **NO STRUCTURED RUN** |
+| `delegation-sweep-codex-st2` | Codex | `gpt-5.6-sol` / medium | 3 | high | `900s` | 5 | — | **NO STRUCTURED RUN** |
 | `ding-mode` | Claude | `claude-sonnet-5` / medium | 2 | medium | `1200s` | 4 | — | **NO STRUCTURED RUN** |
 | `ding-reply` | Claude | `claude-sonnet-5` / medium | 1 | low | `420s` | 2 | — | **NO STRUCTURED RUN** |
 | `docs` | Claude | `claude-sonnet-5` / medium | 3 | high | `1200s` | 5 | — | **NO STRUCTURED RUN** |

README.md

@@ -159,6 +159,41 @@ and include model-free regression tests for those oracle boundaries. These pre-p
 evidence, not accepted corpus receipts; `CATALOG.md` remains authoritative and will show no accepted PASS until
 a committed cell is rerun.
 
+## Delegation-parity tournament
+
+The twelve `delegation-*` cells form one matched tournament asking whether **st2-managed sub-agents are worse
+than harness-native sub-agents** on real delegation work. It is a regression check on an existing fleet
+decision, not a benchmark: the preregistered question, metrics, replication plan, falsification rule, and
+limitations live in
+[`evidence/delegation-parity-design-20260812.md`](evidence/delegation-parity-design-20260812.md).
+
+| Task | `claude-st2` / `codex-st2` | `claude-native` | `codex-native` |
+| --- | --- | --- | --- |
+| Broad multi-file search (`sweep`, fan-out 2) | peer seats over the native bus | one seat, `Agent` sub-agents | one seat, `spawn_agent` sub-agents |
+| Independent review of a diff (`review`, fan-out 2) | peer seats over the native bus | one seat, `Agent` sub-agents | one seat, `spawn_agent` sub-agents |
+| Scoped implementation (`implement`, fan-out 1) | peer seats over the native bus | one seat, `Agent` sub-agents | one seat, `spawn_agent` sub-agents |
+
+Within one task the four arms are byte-identical in task, held-out graders, frozen product repository,
+deliverable contract, and timeout; only the delegation layer differs, and the outcome graders never learn the
+arm. Each native arm has a same-harness managed counterpart, because the decision-relevant comparison is
+within a harness family. Deliverables land in a shared `findings/` directory with per-delegate attribution, so
+one mechanical grader reads a bus delegation and a native fan-out identically.
+
+Two honest boundaries: the native arms necessarily run **unmanaged** — a folder eval launches the harness
+directly, so the fleet's native-spawn refusal never fires — and native fan-out is **self-attested**. The only
+mechanical guarantee there is negative: a native cell declares exactly one bus seat and its delegation judge
+fails if a second mailbox ever appears. Timing is measured as bus latency between the kickoff receipt and the
+coordinator's confirmation, because a single-seat folder eval never signals completion and therefore always
+consumes its whole `max-timeout`.
+
+Matching, grader discrimination, and the tasks' premises are proven for free, without starting a model:
+
+```sh
+bin/check-delegation-parity.sh
+```
+
+No arm has been run. `CATALOG.md` will show no accepted PASS for these cells until one is.
+
 ## Add or change a cell
 
 1. Create exactly one `cells/<name>/<name>.kdl`.

bin/check-corpus.sh

@@ -40,6 +40,7 @@ bin/check-fixture-reset-terminal.sh
 bin/check-harness-contract.sh
 bin/check-vrs-scope-drift.sh
 bin/check-vrs-variations.sh
+bin/check-delegation-parity.sh
 bin/check-weird-git-setup.sh
 bin/check-preflight-closed-set-mutations.sh
 bin/check-overnight-policy.sh