| name | sheaf-repair |
| description | Compositional harness repair via controlled quotients and discrete alignment. Diagnoses restriction obstructions, quotients hidden mediators, re-runs exact CSP. Analog of meta-operad in the sheaf-* family. Zero weight updates: discrete proposals only. Trigger on "sheaf repair", "controlled quotient", "hidden mediator", "stale representative", "align restriction maps".
|
sheaf-repair
If sheaf-glue reports ¬Φ, this skill proposes a discrete fix and asks glue
to re-check.
Allowed moves
- Align a shared field to a canonical representative (e.g. path case).
- Quotient a nuisance interior field (workspace prefix, stale handle)
out of the restriction checks — Batruin’s hidden-mediator construction.
- Propose a new discrete restriction (still a field projection).
Each move is accepted only after sheaf-glue re-runs exact CSP and Φ holds
(or the remaining pre-image is still proper and the operator stops).
Forbidden
- Treating vanishing relative H¹ as acceptance.
- Continuous / sheaf-neural learning of restriction maps (zero-weight violation).
- Overriding a critical CSP veto with a ranking score.
Honesty (do not overclaim)
Controlled experiment: quotient halved candidate budget 2 000 → 1 000 in 20/20
clusters; exact CSP matched the quotient. SWE-bench Multilingual discovery:
118 vs 116 issues, p = 0.75. Discovery gate failed; confirmatory split sealed.
This skill implements the invariance mechanism, not a proven real-world
cohomological advantage.
Aftercare
Always return to sheaf-glue. Repair that does not re-glue is not a global section.
