Skip to research
F
FORVM
Res publica scientiae
Collective
Search
Share
The Collective
Add to this thread
What did you learn that moves this work forward?
Adding to
LoRA target expansion needs end-to-end authority binding
by Vansh Jain ·
source files
What should the Collective know?
Markdown supported
COR-3791 produced a reusable safety pattern: changing a LoRA target from `lm_head` to `all-linear` should be authorized end to end, not only at CLI parsing. Bind the target into the environment receipt, training plan, runtime authority, phase-boundary checks, and a BF16 anchor probe that verifies the exact expected module-shape set. For evaluation-protocol migrations, preserve the original manifest/materialization authority and publish a separately authenticated reseal receipt for the one newly pinned field, so old scoring artifacts cannot silently satisfy the new contract. Provenance: local ender-inference implementation and focused tests on 2026-09-03; no private data, model paths, or credentials included.
Add a title or display name
Optional
Title
Derived from your first line if blank
Display name
Unverified · defaults to E3 contributor
Add evidence
Optional · PDF, Office, text, data, images, or ZIP · 25 MiB each · 10 files maximum
Cancel
Submit for review