# Intent-contrast experiment integration

Workbench PR25 head `ea45b6846ebb25dba05cda81e008860f75cea88c` passed
PR CI `37748967184` and push CI `37748960557`, then merged as
`145922bf9dfd55f8bb05fb85938a4bb0ff4b9d6d`. The merged tree equals the tested tree.

Downloaded artifact `11536783457` ZIP SHA-256:
`176bcc4a25787940b0cc71397eb6f8712a58783beda63c422c13d901a757479e`.
Only the two relevant feature assessments were extracted; the ZIP was removed.

Both CI input hashes, every full-float-array group and label distribution,
policy source hash and actual Gooo decision equal the local observations.
Full input: 576 rows, 576 arrays, compatible maximum 576.
Intent ablation: 576 rows, 24 arrays, compatible maximum 72.
All source families remain explicit. Model calls are zero.

CI reruns array projection and the Gooo assessment using SDK v0.2.26 and its
pinned compiler 0.6.8 (`86b182da`). It rechecks the local native observations via
independent Go comparisons; it does not recollect graph contexts with that older
compiler or freshly execute all 24 native variants in Linux. Those raw native
results remain explicitly local observations of compiler `e0d04650`.

`merge.json` and both raw assessment records are retained. The experiment source,
compressed corpus and native output archive are in the public workbench PR.