Datasets:

corrigibility-games
/

cg-ktane

# Independent technical validation

Status: **Passed the checks below for the locally staged pilot.** This was a separate automated/agent-assisted technical review of the source records and export. It is not independent human adjudication of behavioral labels, a guarantee of absence of sensitive material, or clearance to redistribute third-party content.

## Source integrity and faithful projection

- All 36 original role journals were checked. Their 6,748 events had unique identifiers, contiguous per-writer sequence numbers, non-decreasing monotonic times, and valid SHA-256 previous-event links. This establishes internal consistency, not independent proof of authorship or authenticity.
- All 4,130 distinct directly referenced source-artifact records matched their recorded byte lengths and SHA-256 hashes.
- Independently compared all 934 original request artifacts with the exported contexts: 41,253 ordered instruction/content/current-input text cells matched exactly. This includes provider-exposed thinking text where present.
- Every request retained the exact ordered image sequence separately in its retained history and current input. All 357 published message texts also matched the original records exactly.
- Opaque provider signatures, transport identifiers, local endpoints, and other documented metadata were intentionally omitted. Exported contexts are therefore faithful research projections of the retained observable content, not byte-identical provider API replay payloads.

## Tables, references and assets

- All 24 JSONL tables and their Parquet companions had equivalent rows and values.
- Required primary identifiers were unique, episode/request/event joins resolved, and writer sequences remained contiguous.
- All 6,885 annotation evidence links resolved to exported identifiers, including reviewer screening windows and image assets.
- All catalogued copied objects, media and code files existed and matched their exported checksums and byte lengths. Nested object references resolved; no unlisted or missing object/media files remained.
- No symlinks were used as substitutes for copied evidence. Pattern scans of structured data, code and documentation found no matching credential literals, user-home paths, private temporary paths, or localhost provider URLs.
- Numerical grade fields and composites were absent from the qualitative annotation and annotation-revision tables.

## Role boundaries and timing

- Every Expert request retained the principal/Expert role and manual-enabled protocol; every Defuser request retained the subordinate/Defuser role and manual-disabled protocol.
- In each episode, Expert and Defuser request-image sets had zero overlap. Every Defuser image belonged to its recorded model observations; no such image appeared in the Expert's requests. The underlying source request audit also found 15 distinct manual images per Expert episode.
- Researcher state was explicitly marked as observer-only and was not embedded into the normalized role contexts.
- Associated observations preceded their requests. All 356 recorded message-inclusion links matched the consumed-message lists and a prior receipt by the correct role. There were 357 published messages: publication alone does not guarantee model consumption.
- All 36 post-game reflection requests remained explicitly identified. The 934 decision rows distinguish 910 parsed model outputs from 24 harness recovery outputs; recovery is not credited as an ordinary model-produced decision.

These checks establish the inspected recorded channels. They do not establish absence of memorized game rules, guarantee perception of everything visible, or infer a global causal ordering from wall-clock sorting.

## Outcomes and research interpretation

All 18 selected games had terminal observer states consistent with their recorded outcomes: 12 defused, 3 timed out, and 3 detonated after reaching the strike limit. A completed game is not necessarily a successful defusal. Neither result alone establishes correction-supporting behavior.

The export retains 59 provisional role-contribution annotations across 31 issue chains and 295 dimension cells. Their evidence links were checked technically; their judgments were not independently re-adjudicated by humans during this preparation. Numeric leaderboards are not supplied. Reviewers should read the annotation guide, complete episode context, and documented confounds before using any label.

## Limits of this validation

- The automated privacy scan is pattern-based. It is not an exhaustive visual/audio or semantic privacy review.
- Checksum and reference validation does not establish licensing rights. Consult `RIGHTS_AND_PROVENANCE.md` and the release checklist before publication.
- This pass did not measure real-time latency, audio fidelity, video frame pacing, or model performance under an uninterrupted bomb clock.
- This pass did not run new games or make paid model calls.
- The pilot is a selected terminal-game corpus with fixed seeds, unequal pairings and evolving collection conditions. It is not a representative population estimate or held-out benchmark.

	# Independent technical validation

	Status: Passed the checks below for the locally staged pilot. This was a separate automated/agent-assisted technical review of the source records and export. It is not independent human adjudication of behavioral labels, a guarantee of absence of sensitive material, or clearance to redistribute third-party content.

	## Source integrity and faithful projection

	- All 36 original role journals were checked. Their 6,748 events had unique identifiers, contiguous per-writer sequence numbers, non-decreasing monotonic times, and valid SHA-256 previous-event links. This establishes internal consistency, not independent proof of authorship or authenticity.
	- All 4,130 distinct directly referenced source-artifact records matched their recorded byte lengths and SHA-256 hashes.
	- Independently compared all 934 original request artifacts with the exported contexts: 41,253 ordered instruction/content/current-input text cells matched exactly. This includes provider-exposed thinking text where present.
	- Every request retained the exact ordered image sequence separately in its retained history and current input. All 357 published message texts also matched the original records exactly.
	- Opaque provider signatures, transport identifiers, local endpoints, and other documented metadata were intentionally omitted. Exported contexts are therefore faithful research projections of the retained observable content, not byte-identical provider API replay payloads.

	## Tables, references and assets

	- All 24 JSONL tables and their Parquet companions had equivalent rows and values.
	- Required primary identifiers were unique, episode/request/event joins resolved, and writer sequences remained contiguous.
	- All 6,885 annotation evidence links resolved to exported identifiers, including reviewer screening windows and image assets.
	- All catalogued copied objects, media and code files existed and matched their exported checksums and byte lengths. Nested object references resolved; no unlisted or missing object/media files remained.
	- No symlinks were used as substitutes for copied evidence. Pattern scans of structured data, code and documentation found no matching credential literals, user-home paths, private temporary paths, or localhost provider URLs.
	- Numerical grade fields and composites were absent from the qualitative annotation and annotation-revision tables.

	## Role boundaries and timing

	- Every Expert request retained the principal/Expert role and manual-enabled protocol; every Defuser request retained the subordinate/Defuser role and manual-disabled protocol.
	- In each episode, Expert and Defuser request-image sets had zero overlap. Every Defuser image belonged to its recorded model observations; no such image appeared in the Expert's requests. The underlying source request audit also found 15 distinct manual images per Expert episode.
	- Researcher state was explicitly marked as observer-only and was not embedded into the normalized role contexts.
	- Associated observations preceded their requests. All 356 recorded message-inclusion links matched the consumed-message lists and a prior receipt by the correct role. There were 357 published messages: publication alone does not guarantee model consumption.
	- All 36 post-game reflection requests remained explicitly identified. The 934 decision rows distinguish 910 parsed model outputs from 24 harness recovery outputs; recovery is not credited as an ordinary model-produced decision.

	These checks establish the inspected recorded channels. They do not establish absence of memorized game rules, guarantee perception of everything visible, or infer a global causal ordering from wall-clock sorting.

	## Outcomes and research interpretation

	All 18 selected games had terminal observer states consistent with their recorded outcomes: 12 defused, 3 timed out, and 3 detonated after reaching the strike limit. A completed game is not necessarily a successful defusal. Neither result alone establishes correction-supporting behavior.

	The export retains 59 provisional role-contribution annotations across 31 issue chains and 295 dimension cells. Their evidence links were checked technically; their judgments were not independently re-adjudicated by humans during this preparation. Numeric leaderboards are not supplied. Reviewers should read the annotation guide, complete episode context, and documented confounds before using any label.

	## Limits of this validation

	- The automated privacy scan is pattern-based. It is not an exhaustive visual/audio or semantic privacy review.
	- Checksum and reference validation does not establish licensing rights. Consult `RIGHTS_AND_PROVENANCE.md` and the release checklist before publication.
	- This pass did not measure real-time latency, audio fidelity, video frame pacing, or model performance under an uninterrupted bomb clock.
	- This pass did not run new games or make paid model calls.
	- The pilot is a selected terminal-game corpus with fixed seeds, unequal pairings and evolving collection conditions. It is not a representative population estimate or held-out benchmark.