ARB-APR-3: parity receipt for the decide lane's cell (Qwen3.5-9B-Q4_K_M @ lambda-labs)

[noahgift]

Closes #3574. ARB-APR-001 row 3 (aprender-owned).

What

One evidence file: evidence/parity/arb-apr-001/lambda-labs/qwen3.5-9b-q4_k_m.json. No code.

rust-agentic-harness.md §3.4 — a lane on a backend that has not passed parity is not a reviewer. The apr decide lane had no receipt for its cell. This is it.

Measured

apr parity ~/models/Qwen3.5-9B-Q4_K_M.gguf --prompt "<68-token corpus>" --json, published 0.68.2 from crates.io built --features cuda, on the agent host:

positions	68
parity	true — 68 passed, 0 failed
min cosine	0.9986588 at position 44
threshold	0.99 (evidence/parity/thresholds.yaml → models.qwen3.5-9b)
judged	scripts/check_model_parity.sh --judge … --model qwen3.5-9b → PASS
model sha256	03b74727a860a563… (recorded in the receipt)

Why now, rather than when the cell is ideal

The preferred cell — a declared apr carrying --features cuda on the agent host — is blocked on ARB-APR-2a, which is blocked on forjar's cargo provider having no features: field. Holding this row until then leaves the decide lane running unreceipted while the CLI demonstrably works (apr run: 20/20 cells across 0.8B–27B on two GPU hosts). Producing the receipt for the cell that exists, with its limits stated, beats producing nothing.

What it does not certify — in the file, not just the PR

The receipt carries a non-empty unmeasured[]:

• the lane's binary on this host is the default-features build (default = ["cli"], no CUDA), so the lane runs the CPU arm of this cell. The receipt measures both arms and finds them in agreement at 0.9987; it does not measure the lane's binary;
• apr serve cannot load qwen35 at all (apr serve cannot load Qwen3.5 at all — 'Architecture qwen35 … runs through Qwen35Model'; goal A's floor and the P0 decide-lane design both name apr serve #3571), so there is no serve-path receipt for this cell;
• the serving shape is ungated — every gate we own is single-stream (Name the Qwen model 0.69 blesses and commit its parity receipt at the SERVING shape — our gates are all single-stream #3555).

A receipt with an empty unmeasured[] on a host that could not prove everything would be a lie; that field is why this one is worth committing.

Still owed for ARB-APR-001 row 3's done_when

The infra decide config must reference this receipt, and decide must refuse apr cell unreceipted when the reference is missing. That half is infra's.

🤖 Generated with Claude Code

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=3575 head=8543674413ab9c4556369021d1f024c7ffa4c7af verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.

[noahgift]

Landing obligation: this PR must bump the committed receipt denominator from 7 to 8.

The receipts-under-contract row (#3577) commits a denominator file recording the expected focus-node count, and the extractor exits Unknown{ExtractorMiss} when its count disagrees. Measured on origin/main by the ontology lane: 113 files under evidence/parity/, 7 parity records, 0 carrying a comparator. So the committed denominator is 7, and it is correct — because this receipt is not on main yet.

When this PR lands it adds the eighth, so it must carry the bump in the same commit. If it does not, #3577's gate fires Unknown{ExtractorMiss} and names this PR — which is the falsifier working, not a defect: the acceptance for that row is literally "add a receipt without updating the denominator → ExtractorMiss".

Recording it here because the coupling is invisible from either side alone: whoever merges this will not see #3577's denominator, and #3577 cannot commit an 8 for a file that is not in the tree.

(My earlier "denominator = 8" on #3577 was wrong — I counted a receipt in my own unmerged PR as landed. The ontology lane caught it with git ls-files.)