Hook Enforcement

Replays of the babel-relay drift scenarios through real agent harnesses with the sema ref-gate enforcing at the harness boundary; one experiment, multiple harness/model runs.

Exploratory pilots. Not preregistered, not confirmatory evidence.

Per-run methodology, deviations, and raw records live under experiments/babel-hook/, experiments/codex-hook/, and experiments/cursor-hook/ in the repository.

Run	Harness / model	Gate tier	Off: shipped	Warn: detection	Warn: shipped anyway	Enforce: drift halted
`20260722T120000000Z-cursor-composer-rerun`	Cursor CLI / `composer-2.5-fast`	`wrapper`	18/32	31/31	18/31	31/31
`20260721T123000000Z-cursor-kimi`	Cursor CLI / `kimi-k2.7-code`	`wrapper`	22/28	29/29	9/29	28/28
`20260721T121500000Z-cursor-flash`	Cursor CLI / `gemini-3.5-flash`	`wrapper`	20/27	29/29	1/29	29/29
`20260721T120000000Z-cursor-nano`	Cursor CLI / `gpt-5.4-nano-low`	`wrapper`	14/23	24/24	13/24	25/25
`20260720T120000000Z-cursor-run1`	Cursor CLI / `composer-2.5-fast`	`wrapper`	18/32	32/32	25/32	32/32
`20260719T120000000Z-codex-run2`	OpenAI Codex / `gpt-5.6-luna (reasoning effort low)`	`hook`	22/32	32/32	9/32	32/32
`20260718T180000000Z-claude-code-run144`	Claude Code / `claude-haiku-4-5`	`hook`	22/32	32/32	4/32	32/32
`20260718T120000000Z-claude-code-run36`	Claude Code / `claude-haiku-4-5`	`hook`	6/8	8/8	1/8	8/8

Run	Created	Harness	Gate tier	Model	Evidence claim
`20260722T120000000Z-cursor-composer-rerun`	2026-07-22T12:00:00.000Z	`Cursor CLI`	`wrapper`	`composer-2.5-fast`	144-trial babel relay replicating the 2026-07-20 composer-2.5-fast run through the same Cursor wrapper gate: warn detects 31/31 drifted relays but 18 ship anyway (58%, vs 78% original); enforce halts 31/31. Composer remains the leakiest model in the isolation set.
`20260721T123000000Z-cursor-kimi`	2026-07-21T12:30:00.000Z	`Cursor CLI`	`wrapper`	`kimi-k2.7-code`	144-trial babel relay through the same Cursor wrapper gate, model swapped to kimi-k2.7-code: warn detects 29/29 drifted relays but 6 ship anyway; enforce halts 28/28. Unaided (off) it catches just 6/28.
`20260721T121500000Z-cursor-flash`	2026-07-21T12:15:00.000Z	`Cursor CLI`	`wrapper`	`gemini-3.5-flash`	144-trial babel relay through the same Cursor wrapper gate, model swapped to gemini-3.5-flash: warn detects 29/29 drifted relays and only 1 ships anyway; enforce halts 29/29. Unaided (off) it catches just 7/27.
`20260721T120000000Z-cursor-nano`	2026-07-21T12:00:00.000Z	`Cursor CLI`	`wrapper`	`gpt-5.4-nano-low`	144-trial babel relay through the same Cursor wrapper gate, model swapped to gpt-5.4-nano-low: warn detects 24/24 drifted relays but 10 ship anyway; enforce halts 25/25. Heavy caveat: 47 hop failures and 20 malformed audits thin the denominators.
`20260720T120000000Z-cursor-run1`	2026-07-20T12:00:00.000Z	`Cursor CLI`	`wrapper`	`composer-2.5-fast`	144-trial multi-hop babel relay through Cursor CLI with a sema check wrapper gate (composer-2.5-fast): warn detects 32/32 but 25 drifted relays ship anyway; enforce halts 32/32 pre-model.
`20260719T120000000Z-codex-run2`	2026-07-19T12:00:00.000Z	`OpenAI Codex`	`hook`	`gpt-5.6-luna (reasoning effort low)`	144-trial multi-hop babel relay through OpenAI Codex with the unmodified sema Claude Code plugin (gpt-5.6-luna low): warn detects 32/32 but 9 drifted relays ship anyway; enforce halts 32/32 pre-model.
`20260718T180000000Z-claude-code-run144`	2026-07-18T18:00:00.000Z	`Claude Code`	`hook`	`claude-haiku-4-5`	144-trial multi-hop babel relay through Claude Code with the ref-gate hook (haiku): warn detects 32/32 but 4 drifted relays ship anyway; enforce halts 32/32 pre-model.
`20260718T120000000Z-claude-code-run36`	2026-07-18T12:00:00.000Z	`Claude Code`	`hook`	`claude-haiku-4-5`	36-trial multi-hop babel relay through Claude Code with the ref-gate hook (haiku): first voluntary-compliance leak observed; enforce halts all drifted relays.