Record live Modeling authority proof

This commit is contained in:
Meta Vens 2026-08-08 04:07:43 +01:00
commit dae006a01c
3 changed files with 53 additions and 5 deletions

View file

@ -1317,6 +1317,20 @@ authority fields whose exact identities Mind validates. This moves the first
line of defense from a prose warning plus downstream rejection into the model's
declared output contract; Mind remains the final admission owner.
Live v42 proves that contract end to end: ordinary resident Modeling received
read-only source tools, grounded its map in inspected repository paths, emitted
only generic node/edge evolution, and reached Mind RepoModel admission. The next
split is downstream. CRRC still derives resumability from the thread state's
durable investigation checkpoint, while Modeling acceptance commits RepoModel
admission plus observation/evidence state and does not establish that
checkpoint. Consequently accepted Modeling can leave `prepareCheckpoint` true
and trigger another generic Modeling launch. Neither generic `statePatch` nor
the worker may become a second checkpoint owner. A typed proposal now asks the
frontier-planning circuit to choose one authority: either Mind admission derives
the continuity checkpoint deterministically from accepted RepoModel evidence,
or CRRC consumes canonical RepoModel admission directly. Canonical Imagination
and dedicated Mind must adopt the choice before Hands changes the seam.
## Exact release build cache ownership
Release construction has two serialized cache authorities. The graph cache is
@ -1363,3 +1377,9 @@ Cargo in 29.58s. The current iteration bottleneck is the broad recompilation
and relink fan-out of the monolithic core crate. Incremental release reuse,
linker choice, or a deliberate authority-based crate split must be benchmarked
as separate release-contract changes rather than mixed into a cognition fix.
Exact `ecfff489` confirmed the slow case at 7m04s of Cargo work. The active
release rustc command had no incremental directory, and the root bundle built
all 21 binaries. Release incremental compilation and `rust-lld` therefore have
separate measurable hypotheses; an authority-based core crate split remains the
larger structural option if those two changes do not materially reduce edits to
the current monolith.