skills / reports / answer-verifier-xval-matrix.html

answer-verifier — cross-LLM verdict-reproduction matrix

Skill: security/answer-verifier · Date: 2026-07-04

Free models under test: nemotron-3-super-120b (nvidia, 120B) · cohere/north-mini-code:free

(3B) · nemotron-3-ultra-550b (OpenRouter, 55B active) — all timeout 900s.

Method: each fixture answer + its ground-truth files were handed to each free model *with the

answer-verifier SKILL.md as its instructions*; it produced the ANSWER-VERIFY block unaided. No

internet. Verdicts compared to the Opus reference run (answer-verifier-validation.md).

Result — three free models vs. the Opus reference

Two views: the 3-tier verdict the models emitted (PASS/FLAG/FAIL — the pre-verdict-layer format

these runs used), and the binary GATE the family now ships (PASS 0–2 / FLAG 3–8).

Verdict + severity:

FixtureOpussuper-120b (120B)north-mini (3B)ultra-550b (55B)
0 groundedPASS 0PASS 0PASS 0PASS 0
3 unsupportedFLAG 3FLAG 3FLAG 2PASS 0 → FLAG 3
4 overclaimFLAG 4FLAG 4FAIL 8FAIL 7
6 fabricatedFAIL 6FAIL 5FAIL 8FAIL 6
8 confabFAIL 8FAIL 5FAIL 8FAIL 8

Same runs as GATE (PASS 0–2 / FLAG 3–8):

FixtureOpussuper-120bnorth-miniultra-550b
0PASSPASSPASSPASS
3FLAGFLAGFLAGPASS → FLAG
4FLAGFLAGFLAGFLAG
6FLAGFLAGFLAGFLAG
8FLAGFLAGFLAGFLAG
† ultra fixture-3 was contaminated, then corrected — and this is the report's real finding.
The standard run returned PASS 0/8, and its ledger quoted the *grounded* fixture's claims that
aren't in this answer. Cause: hermes loads ambient user memory/rules (~/.hermes/SOUL.md +
memories/, which carry Gortex setup docs) into every run unless told not to — proven because
super-120b's fixture-3 output also contains "Gortex" text that is nowhere in the prompt. A
clean-room re-run with --ignore-user-config --ignore-rules returned FLAG 3/8, exact to Opus.
So the "miss" was a context-hygiene artifact, not a model verdict.

"Gortex" but still verdicted correctly). The free auto-verifier.

Over-escalated fixture-4 (FLAG→FAIL 8, but GATE-safe) and pins 8/8 on FAIL-side; never PASSed

anything it shouldn't. Conservative pre-filter.

leaked Gortex docs and wove them into its ledger (a false PASS on run 1; a hallucinated "Gortex

section" on run 2). Sandboxed, it matched Opus (FLAG 3) and nailed fixtures 6/8 exactly. The

failure was contamination, not the model.

Reading

unless you pass --ignore-user-config --ignore-rules. That leaked Gortex setup docs into every

prompt this session. Small/mid models shrugged it off; the long-context 550B wove it into its

output and flipped a verdict. Longer context = larger contamination surface when unsandboxed.

(All prior swarm results this session ran un-sandboxed and still reproduced correctly — but ultra

fixture-3 shows it *can* bite. Clean-room is the fix.)

quibble (north-mini's and ultra's fixture-4 both stay FLAG). The only thing it can't hide is

PASSFLAG, and the single such divergence was the contamination artifact — which clean-room

removed. Under clean-room + GATE, all three free models agree with Opus 5/5.

north-mini over-anchors at 8).

Role assignment (final)

--ignore-rules`); its long context makes it the most vulnerable to prompt contamination.

memory/rules don't leak into the task. For anything at stake, still prefer super-120b + human or a

2-of-3 ensemble over any lone free model.

xv-ul-*.md + xv-ul-3clean.md (ultra, incl. clean-room).