Repository navigation
apr: llama.cpp logit-compare verb as the PRM parity.oracle (PRM-001 v2 §4.1) #4444
Description
Operator ruling, 2026-09-25 (PRM-001 v2, item 3a): add an apr verb that compares logits against llama.cpp@d1d3c3396, to serve as the declared parity.oracle in rex-cell-admission-v1.
Today apr parity compares GPU vs CPU only. The REX-04 receipt found that.
Acceptance:
- The verb exists and emits a comparison receipt, per position, against llama.cpp logits.
- A planted-divergence fixture turns it RED.
- It can be declared as
parity.oraclein a cell.
Until it is declared, cells use apr-parity-gpu-cpu plus the H1 verdict check.
Parent: #4354 (PRM-001).
Activity
spec(prm): PRM-001 v2 — re-lock prereg (rex-prereg-v2) before any dat…
noahgift commentedon Sep 25, 2026
Branch
feat/4444-llamacpp-logit-oracle@300ec86e4, pushed only (fold owner: 59).apr parity-oracle --reference <llama.cpp APRRAWLG> --subject <apr APRRAWLG> --threshold F --threshold-basis TEXT -o receipt.json. It writes a per-position receipt (cosine, argmax ref/subject, max |diff|). With--jsonit prints therex-cell-admission-v1Parityblock:{oracle: "llama.cpp@d1d3c3396", cosine, threshold, threshold_basis, receipt_sha256}. aprender-84 confirmed the field set.1d5eaf12…/f6f79264…), the verb reproducescompare_raw_logits.pyexactly: min cosine 0.995905 at pos 4, argmax mismatches [2, 28, 36, 73]. Seeevidence/parity/oracle-4444/CROSSCHECK.md.qwen3.5-0.8b@cpu-vs-llama.cpprow inthresholds.yamlhas no measured basis. The verb requires--thresholdand--threshold-basisand supplies no default. Also, apr has no shipped producer for the subject APRRAWLG yet; the pair above came from the feat(qwen35): Qwen3.5 / Qwen3.8 hybrid GGUFs run on the CPU — apr run + apr chat (#3091) #3114-era example.feat(parity-oracle): --model produces the subject in-process over the…
evidence(parity-oracle): --model on lambda reproduces the intel subje…
noahgift commentedon Sep 25, 2026
Follow-up landed on branch
feat/4444-llamacpp-logit-oracle@03579caa9, for 59's fold. No new PR.apr parity-oracle --model <qwen3.5.gguf> [--subject-out S]now produces the subject in-process. It runs realizar's Qwen3.5 CPU forward (forward_single_qwen35, one token per call) over the reference's own token ids, so the ids agree by construction. The receipt'ssubject.producerrecords the forward and the model's sha256.Measured on lambda with the model file from REFERENCE.md (bd258782…) and reference 1d5eaf12… (llama.cpp d1d3c3396):
apr-intel-run1.bin(f6f79264…).Checks:
Still open: the
qwen3.5-0.8b@cpu-vs-llama.cpprow in thresholds.yaml is unmeasured. The 0.99 above is a crosscheck value, not a basis.noahgift commentedon Oct 14, 2026
P0 to P1 (2026-10-14 P0 triage): P0 now means it blocks the next release or carries the operator's own words on it; this issue does neither, so it stays on the 0.71.0 milestone at P1.
noahgift commentedon Oct 14, 2026
merged into #5317 (master ticket for epic #4381; triage rule 7b, 2026-10-14). It is a checklist row there; reopen this issue when work on it starts or its PR opens.