> Public sanitized report copy. Local infrastructure paths and opaque contract identifiers are redacted where present, and nonessential internal process/authorization commentary is omitted; financial values and scientific limitations are retained. Historical pending/open wording below records the report-time state, not current program status. The scoped successor program subsequently completed; commissioning-delay and recontracting extensions were deferred, not executed. Historical predecessor holds remain unchanged. This is report-level evidence, not a raw-data reproduction bundle. Referenced checker, CSV, database and provider artifacts are not included. No new experiment or audit was performed for this export.

# SUMMARY: disclosed settlement order control (read-only reconstruction)

**Answer.** With the settlement order stated in the brief, Lab A pre-borrowed 5 at 2028-H1 (800 bps), before conditional credit was withdrawn at 2028-H2, in **6 of 6** episodes. The undisclosed comparator did so in **1 of 6**. A went dark in 2028-H2 in **0 of 6**, against **5 of 6** in the comparator. It opened the final turn with 22.5 rather than 17.5, paid the full 20 minimum, served 2 GW and left no unpaid claim.

| | early borrow | dark final turn | A CLI | B CLI | Y |
|---|:-:|:-:|---:|---:|---:|
| fixed rival, control r1–r3 (identical) | 3/3 | 0/3 | −19.173684210520 | −20.526315789480 | −39.7 |
| fixed rival, comparator r1, r2 / r3 | 0/2, 1/1 | 2/2, 0/1 | −27.5 / −19.17 | −10.0 / −20.53 | −37.5 / −39.7 |
| liquidity rival, control r1–r3 (identical) | 3/3 | 0/3 | −19.7 | −19.6875 | −39.3875 |
| liquidity rival, comparator r1–r3 (identical) | 0/3 | 3/3 | −27.5 | −9.6875 | −37.1875 |

- **Lab A in control:** cash close 10.83 (fixed rival) or 10.3 (liquidity rival), principal 5, unpaid 0. Terminal net is 5.83 or 5.3.
- **Lab A in comparator dark episodes:** cash 0, principal 0, 2.5 unpaid to the project.
- **Combined Y is lower in control.** Both labs now operate into 1 unit of final demand, and A pays coupon on the new note. Lower Y is not worse finance: the comparator's higher Y came from A going dark.
- **Lab B** loses about 10 against the comparator because it now shares the final sales. Its script is unchanged and was recomputed independently.
- **Rationales** are reported intent, not cognition. In all 24 control decisions the plan and commitments restate the order ("sales arrive last", "do not use same-turn sales to fund earlier settlement steps"). That count is 0/24 in the comparator. The short 2028-H1 rationale still gives withdrawal risk as the reason ("protects continuity if another revenue failure suspends next-opening credit").

**Verification.**
- `check_saved.py` exits 0 with 5,213 assertions.
- 24 accepted attempts and 24 reservations. 30 PROCESS receipts, all exit 0. 0 STOP files.
- 24 provider receipts: HTTP 200, completed, `gpt-6-astra-900k` → `gpt-6-astra`. Ledger, request and response hashes join.
- 6 Stores pass integrity and FK checks and match the saved JSON. 6 cold receipts pass (inspected, not re-executed).
- 1,780 source pins match `6952b6d0…`.
- Cash is conserved: 544.4 in every episode.
- The only model-input delta from the comparator is the brief paragraph, plus a namespace-bound `packet_digest`.
- The run dir has 1,251 files. They are hash-identical before and after (`sha256sum -c`). SQLite was opened `immutable=1` and no side files were created.

**Limits.**
- Before/after design, not randomized. Same model identity (`gpt-6-astra-900k` → `gpt-6-astra`), different run time.
- Three trajectories per cell are not independent samples. No significance tests, rates or welfare claims.
- The checker comes from the same model family and method lineage as the comparator analyst.
[Omitted: internal rejected-authorization field discrepancy; no spending occurred.]
