[faranm-stack]

Summary

Documents AIMNet2 as incomplete and experimental in the nvidia-alchemi agent, as agreed for microsoft#212, and corrects a stale AIMNet2 cache path in the agent instructions.

Type of change

• New agent (under agents/<name>/)
• Update to an existing agent
• New starter kit (under starter-kits/<name>/)
• Update to an existing starter kit
• Schema change (docs/schemas/**)
• Workflow / script change (.github/**)
• Documentation only
• Other (describe below)

Related issue / tracking

refs microsoft#212; follows #1 and #2.

Schema impact

• No schema changes
• Schema changes are backward compatible
• Schema changes are breaking — add the breaking-change label and link the rollout plan below

Validation checklist

• Documentation (README.md, schemas) is updated to match the change.
• No hand-edits to .auto-registry/** (it is generated post-merge).
• No OS / editor artefacts committed (.DS_Store, .env, *.swp, .idea/, .vs/, etc.).
• If model-weight files are added, they are Git-LFS tracked, ≤5 GB each, and in an allowed format. (No model-weight files added.)
• No secrets, customer data, or internal-only URLs are committed.
• All Markdown links resolve.

Reviewer notes (optional)

Why. The nvalchemi AIMNet2 wrapper disables the Coulomb and DFT-D3 terms (its docstring says to compose them separately), but the bundled aimnet2 checkpoint requires both: its metadata sets needs_coulomb, needs_dispersion and coulomb_mode: sr_embedded. load_model("aimnet2") therefore returns only the short-range neural-network term. On the bundled water molecule, compared with the aimnet package's reference calculator using its metadata defaults:

	load_model("aimnet2")	aimnet reference
Maximum force	9.48 eV/Å	0.68 eV/Å
O–H stretch force	+5.55 eV/Å (apart)	−0.45 eV/Å (together)
Symmetric O–H scan minimum	none below 1.10 Å	0.959 Å

The agent path matches the reference calculator with Coulomb and D3 disabled (|ΔE| ≤ 4×10⁻⁴ eV, |ΔF| ≤ 6×10⁻⁷ eV/Å), and its forces match finite differences of its own energy, so the gap is the missing terms rather than the force calculation. The existing release checks (finite values, GPU/CPU agreement, rotation invariance) cannot detect this.

Changes (6 files, text only; no numerical code paths change):

• Agent, tool and metadata descriptions: validated scope is MACE inference, relaxation and MD; AIMNet2 is incomplete.
• Agent instructions: AIMNet2 joins the workflows that require disclosure and explicit confirmation (requirement 9), and the potentials table marks it experimental. The AIMNet2 cache path now reads /opt/cache/aimnet/aimnet2_wb97m_d3_0.pt, where nvidia-alchemi: hardening fixes and H100 revalidation for #212 #1 moved it.
• list_models() description, README verification status and coverage list, and CHANGELOG.

Validation on this exact commit:

• The repository's validate_pr.py from current main, run against this branch merged with main: passed, 0 failures.
• Clean image build: pip check passed; all three checkpoints loaded after hash verification; all 43 tests passed with no skips.
• Agent instructions are 31,936 characters; the schema limit is 32,000, so future additions will need trimming elsewhere.

Not changed: the AIMNet2 wrapper behaviour itself. Restoring the Coulomb and D3 terms by default, with the reference comparison above as a regression test, is a possible follow-up for the maintainers.