Skip to content

docs(material): document the two-head sigmoid learned-lam reroute (.cvcnav v2) - #107

Open
transfix wants to merge 1 commit into
mainfrom
docs/material-nav-two-head-lam
Open

transfix wants to merge 1 commit into
mainfrom
docs/material-nav-two-head-lam

Conversation

@transfix

Copy link
Copy Markdown
Contributor

docs/MATERIAL_NAV.md documented only fixed MaterialParams lam_soft/lam_hard. The shipping learned two-head sigmoid reroute (#105 — .cvcnav format v2) was undocumented. Docs-only.

Changes

A new "Learned lam — the two-head sigmoid reroute" subsection under Using it:

  • a trained coefficient net can predict lam_soft/lam_hard per-agent (the source method's sigmoid-bounded λ heads) instead of using fixed params;
  • single head (4th output overrides lam_soft, softplus) vs two heads (4th/5th = lam_soft/lam_hard, each lam_max·σ(raw));
  • why sigmoid bounding (a head can cleanly suppress its channel, λ→0), the per-head ceilings stored in the .cvcnav v2 trailer, the pre-v2-host hard-fail, and that lam_hard stays ungated (witness gate multiplies the predicted lam_soft only);
  • orthogonal to the mu-in-features widen; fixed params remain fully supported.

Parallels the corresponding libcvc NAV_MATERIAL.md update for the C++ twin.

…vcnav v2)

MATERIAL_NAV.md documented only fixed MaterialParams lam_soft/lam_hard. Document
that a trained coefficient net can instead predict them per-agent — the source
method's sigmoid-bounded lam heads — landed as the .cvcnav format-v2 two-head
reroute:

- single head (4th output overrides lam_soft, softplus) vs two heads (4th/5th =
  lam_soft/lam_hard, each lam_max*sigmoid);
- why sigmoid bounding (clean channel suppression), the per-head ceilings in the
  v2 trailer, the pre-v2 host hard-fail, and that lam_hard stays ungated;
- orthogonal to the mu-in-features widen; fixed params remain supported.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant