# Reference-class calibration results

Frozen candidate: `postgresql_regional_failover_reconciliation_1`

Result: four observed GPT-5.5 public-only attempts produced zero perfect
scores. The two-low plus two-medium score gate in `CALIBRATION_PROTOCOL.md` is
complete, with the retention limitations documented below. This is not an
official ALE difficulty label, and four attempts do not establish a
statistical pass rate below one percent.

## Frozen bindings

- Scorer SHA-256:
  `2d713076657b69898804537e4ee169e829ef225df8c470189ac3226b300a8072`
- Builder SHA-256:
  `ded97428ff81f35f696d1f51b86c7f9a6a074f8b5c35fe1789f1eea9a48a12fe`
- Task-card SHA-256:
  `050649eccc15fbda64c9363df52ef458985435b287df5ec127ad85ce0f50e3dc`
- Input manifest SHA-256:
  `510c92e7c0485ada35cf0c3d5c4f7f286d0b80fd2b24a6af4f0de762da82b046`
- Hidden truth SHA-256:
  `1836389e84d7c1692913c933cef9e68cb81951d75dfb453e609139f2c2c61edb`
- Solve image: `ale-postgresql-pitr-pilot:17.11`, Linux AMD64, ordinary
  user, image ID
  `sha256:6fcf1efdc02edc303118f58abaa7918784ca09fa9c81687b838c1c636d166656`

## Isolation and scoring procedure

Each attempt used a fresh task-specific model turn with `fork_turns=none`,
the frozen public input and software trees, an empty output directory, no host
workspace mount, no task-container network route, and a 4-CPU/6-GiB Linux
AMD64 solve container. Plain or encrypted reference data, scorer code, private
builder code, and the private seed were absent from the solve container.

Codex Desktop `0.148.0-alpha.15` rollouts confirm the `gpt-5.5` model and
low/medium effort settings. Every recorded tool action targeted only its
designated solve container. Candidate files were copied out only after the model stopped. The solve
container was then deleted. A separate no-network evaluator container received
the candidate, frozen scorer, PostgreSQL 17.11, and hidden reference. The four
scores below were all recomputed against the scorer binding above.

## Results

| Attempt | Model effort | Model duration | Files | Score | Cap or terminal cause | Evaluator wall time |
| --- | --- | ---: | ---: | ---: | --- | ---: |
| low-1 | GPT-5.5 low | 630.728 s | 7 | `0.55` | `required_event_state_below_95_percent` | 20.312 s |
| low-2 | GPT-5.5 low | 546.286 s | 4 | `0.0` | Model stopped without the three required dumps | 11.450 s |
| medium-1 | GPT-5.5 medium | 944.576 s | 7 | `0.972191932834` | No cap; shard A, event graph, and sidecars remained imperfect | 22.679 s |
| medium-2 | GPT-5.5 medium | 767.556 s | 7 | `0.507854304311` | `required_event_state_below_95_percent` | 19.380 s |

Exact component scores:

| Attempt | Archive | Coordinator | Shard A | Shard B | Branch | Event graph | Cross-cluster | Sidecars |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| low-1 | 1 | 1 | 0.194826880060 | 0.198603078285 | 1 | 0.104170589901 | 1 | 0.526668884453 |
| low-2 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| medium-1 | 1 | 1 | 0.918559483837 | 1 | 1 | 0.953010561182 | 1 | 0.928661092457 |
| medium-2 | 1 | 1 | 0.088766399678 | 0.091901073907 | 1 | 0.050329018944 | 1 | 0.516949598631 |

The retained rollout metadata records:

| Attempt | Input tokens | Cached input | Output tokens | Reasoning output | Total tokens |
| --- | ---: | ---: | ---: | ---: | ---: |
| low-1 | 1,929,066 | 1,849,600 | 27,730 | 1,862 | 1,956,796 |
| low-2 | 1,901,661 | 1,822,208 | 22,447 | 2,801 | 1,924,108 |
| medium-1 | 2,848,902 | 2,738,176 | 40,179 | 5,598 | 2,889,081 |
| medium-2 | 2,562,326 | 2,473,728 | 34,172 | 4,974 | 2,596,498 |

Rollout source paths, bytes, hashes, session IDs, timestamps, model settings,
and tool-isolation counts are retained in
`trajectories/TRAJECTORY_MANIFEST.json`. Raw rollouts remain in the private
Codex session directory and are intentionally not duplicated into this
workspace. The cross-agent task message is encrypted in the rollout, so its
plaintext is not independently retained. Exact billed API or subscription cost
was not exposed and is not inferred. A final sequential re-score, including
start/end timestamps and complete scorer reports, is retained in
`RESCORE_TIMING_EVIDENCE.json` (SHA-256
`396bbca83e49780945c0efb00d2dc0eedd21ac5ba6e2fa491673306273f34bbc`).

## Candidate artifact hashes

Each line is `filename`, bytes, and SHA-256.

### GPT-5.5 low attempt 1

```text
branch_salvage.csv         4148  6cd328af5809bf634c60ae8c233ba2612ad8d3301bc211e5be6ca526c28ac5d0
coordinator.dump          89183  a02b298ed077dc1edabeef28e431cac5f49cde2a8cf6901e08388233e45e3f64
event_reconciliation.csv 2695489 e8003cc19df0b470ada736ea8ef282fd4c874a83fe7ab40734be85b533df32c5
incident_report.json       1195  a1fd9ee9dc542d82ad864432e3337d21834d9ffc332350b53fd966320af3d905
recovery_manifest.json     2908  c2daa26ba66a18bef033c1d3b658d85c17b7a404686ce45f55f4fbefaf352578
shard_a.dump             505238  cc36f6af912cdb72e9789159f70849c181e7330315148471c0ab5d15540524a0
shard_b.dump             506780  3899567169bfea4a742d525fbc0ab2b77f7ca3fc855646f58a29d8c019a53e54
```

### GPT-5.5 low attempt 2

```text
branch_salvage.csv         4148  6cd328af5809bf634c60ae8c233ba2612ad8d3301bc211e5be6ca526c28ac5d0
event_reconciliation.csv 3079769 642122aa90cc3ddd7e1b9914fdb7f0fdb663a0b437421451ebad5230b416cbd2
incident_report.json        950  8f1dd1cd059be16909b227ef51b77f9343637cbd12908fe4e3cd90dc0250a405
recovery_manifest.json     1710  20096913c3bc346b087d73217d35615cb7570022c8d62ee5200fe8bad5777011
```

### GPT-5.5 medium attempt 1

```text
branch_salvage.csv         4148  6cd328af5809bf634c60ae8c233ba2612ad8d3301bc211e5be6ca526c28ac5d0
coordinator.dump          89181  1164a8be402dab8246190238634507f5cefe149e9f4f89342fe485d534a39770
event_reconciliation.csv 3082841 ed15424728317b9749e8fdfb738a15edc61e1aab05cb344237f1182e269f040a
incident_report.json        943  816625b16fc39d1de577cf901315cf255435d44806e5fd9dc175bd8e924ae948
recovery_manifest.json     2384  3cd1b01659025c059a881d9e4c8167db20f0a7fceb7bec9e80a7b12cc7d4d331
shard_a.dump            1867281  a4caefdebd08e08b5fbcd590e92715271766b98f7c072464acc4ae79936101a4
shard_b.dump            1872006  57ff17de0f15a74ec69909ff17272b4f81847e2b91ac3255d4ea4ce13f5ec08d
```

### GPT-5.5 medium attempt 2

```text
branch_salvage.csv         4148  6cd328af5809bf634c60ae8c233ba2612ad8d3301bc211e5be6ca526c28ac5d0
coordinator.dump          89182  1f390178c7fd4007c708825c982309f6d5a66a1d6ef1fb57f8fa4a04181cb051
event_reconciliation.csv 2741409 21862b1d4f0bf6a3f9da3abe2570010b78f20c21c79b44d7a9335aa59084a223
incident_report.json        914  f1d9e5fb0a9d0f086cee7ef905be556354787d84c1546990168e266aeca82be3
recovery_manifest.json     2382  02335bc34fc8b7386b51460300e69ea5fd5b1c797b0db2473ef62f78a7e2b043
shard_a.dump             507131  6b1c8b8ac52e569456ab080451467478b1da445e85daee1ac476f3cc6b922275
shard_b.dump             507585  280a65babafb19ad440a0346f1e169e0fdc6292dc0af704bfdfaadd9a5a1aa38
```

## Cross-vendor availability

No Sonnet or Gemini run was launched for this regional candidate. A prior
direct Sonnet 4.6 launch on the earlier single-cluster prototype showed that
the Anthropic account lacked usable credit; that is only an account-availability
preflight and is not a result for this task. No Gemini credential was
available. A valid cross-vendor run remains recommended when credentials
permit, while the contributor-side minimum matrix was two GPT-5.5 low and two
GPT-5.5 medium attempts.