v0.70.1

aprender 0.70.1 — release notes

How this release was gated

0.70.1 ships under a recorded operator emergency scope, crux-smoke, as 0.69.1 did.
The operator's words: "0.70.1 ships on CRUX smoke on lambda and gx10 GPU. Everything bigger is nightly."
The record is the crux-smoke entry for release "0.70.1" under emergency_scopes in
contracts/model-capability-ladder-v1.yaml.

The gate. The CRUX smoke ran on the lambda GPU (RTX 4090) and the gx10 GPU (GB10), with nothing else on either GPU. It
covered every certified model (Qwen3.5-2B Q4_K_M, Qwen3.5-4B Q4_K_M and Qwen3.5-4B UD-Q4_K_XL) and every admitted
thinking mode, on the control prompts. apr was checked against llama.cpp (pinned d1d3c3396), Hugging Face transformers
and vLLM, on the verbs run, chat, serve and code.

host	apr	cells	result
lambda	apr 0.70.1 (cc4463f7a6)	134	134 GREEN, 0 RED
gx10	apr 0.70.1 (cc4463f7a)	134	134 GREEN, 0 RED

These rows are the smoke of the binary built at cc4463f, and the judge passed them under the scope (rc 0).
The publish path judges the release commit itself, so both GPUs were smoked again with the binary built at the
release commit, whose source is identical to cc4463f. A commit cannot hold results produced from itself, so those
receipts are published with the GitHub release for v0.70.1.

On lambda, the two 4B models ran with APR_QWEN35_PREFILL_GEMM=f32. This skips apr's fp16 prefill prewarm (6.8 GiB,
#4313), so the Hugging Face and vLLM reference engines fit beside apr on the 24 GB card. On the default path, every
4B serve cell had no reference answer because hf and vLLM could not start beside apr. apr agreed with llama.cpp on
all 80 of those cells. The 2B ran on the default path.

Printed as evidence, not the gate for this release. The full model matrix (R7) and the release-readiness grade
(R8) still run at publish, and their verdicts are printed. They stay red where they were red; see Known failures.

Unchanged. Clean tree, tag on HEAD, release branch, dogfood GO for this commit, probes 14/14, plants 22/22,
Lean T5, and the clean-room green on the tagged commit.

Coverage: the floor held, so its waiver was not used. The operator had waived the 89% line-coverage floor for
0.70.1 only ("coverage floor doesn't gate this release"). The dogfood at 7d2dc28, whose code is identical to the
release commit, measured 91.36% of lines, so the floor held and the waiver was never applied. The dogfood at the
release commit ran with the floor in force (no override); its coverage row: PASS, "91%, floor met".
P0 for 0.70.2 is unchanged: measure coverage after the release, restore the floor, and move coverage to the nightly run.

Next (P0 for 0.70.2). No check may be enforced at publish until its inputs have been produced green on main,
nightly, three times. The full CRUX lanes and the readiness check move to nightly, and CRUX gets the GPU to itself.

Fixed

• apr code now gets an input budget on small-context models such as tinyllama. It failed with "context overflow:
  required 12 tokens, available 0" (#4599 regression, fixed in cc4463f).

Build note

The release commit differs from cc4463f only in scripts/check_publish_preflight.sh, one contract entry and
evidence/. crates/, src/, Cargo.toml and Cargo.lock are identical. crates/aprender-contracts/build.rs reads
contracts/*.yaml, so the contract entry recompiles aprender-contracts and the crates that depend on it. build.rs
does not read emergency_scopes, so the code is unchanged. evidence/release/surface/0.70.1.json records the
binary built at cc4463f. A file inside the release commit cannot name that commit's own hash.

Why the release commit is 04332f8, not 7d2dc28

The dogfood at 7d2dc28 was NO-GO on one row, bashrs. That commit added one self-test line
(scripts/check_publish_preflight.sh:821) whose skip comment named SEC014, while bashrs reports SEC010 on that line.
04332f8 names both rules in that comment and adds its review record under evidence/. The bashrs gate itself did
not change, and 7d2dc28 was never tagged. crates/, src/, Cargo.toml and Cargo.lock are identical in both commits.

Known failures (shipped, tracked as P0 for 0.70.2)

Each fails the same way on v0.69.5-rc.1, so none is new in 0.70.1. Both results are saved under
evidence/release/0.70.1/known-red/.

model	host(s)	check	failure in 0.70.1	same on v0.69.5-rc.1	ticket
qwen35-0.8b-q4km	lambda, gx10	golden_output thinking_on	think block unclosed within 2048 tokens	yes	#4666 (continues #4030 (closed))
Qwen2.5-0.5B-Instruct-f16.gguf	lambda	cuda serve /api/chat	gibberish output	yes	#4661
Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf	lambda, gx10	cuda serve /api/chat	gibberish output	yes	#4662
Qwen3.5-0.8B-IQ4_XS.gguf	lambda	golden_output thinking_on	no measured thinking budget (named refusal; the GPU ran). The old binary failed earlier, with a CPU fallback.	fails on both, differently	#4663
Qwen3.5-0.8B-UD-IQ2_XXS.gguf	lambda, gx10	golden_output	GPU forward fell back to CPU (chat rc=14); wrong answer	yes, it fell back on v0.69.5-rc.1 too (the old binary has no rc=14)	#4664
Qwen3.5-35B-A3B-UD-IQ4_XS.gguf	lambda	capability_match	no CUDA forward for qwen35moe (named refusal)	yes	#4665

None of these is marked known red or green for this release. Each stays red in the printed evidence and in its ticket.

Refused (does not fit)

The memory-fit check refuses these rows on lambda, the same as on v0.69.5-rc.1. They are not counted as failures.

• qwen2.5-coder-0.5b-instruct.apr
• qwen2.5-coder-1.5b-instruct-q4k-v2.apr
• qwen2.5-coder-1.5b-instruct-q4k.apr
• qwen2.5-coder-1.5b-instruct-st.apr
• qwen2.5-coder-1.5b-q4k.apr

On gx10, the memory-fit check refuses these rows (llama-fit-params exit 1). They are not counted as failures.
v0.69.5-rc.1 was not measured on gx10 for these rows, so there is no old result to compare.

• qwen2.5-coder-1.5b-instruct-fp16.apr
• qwen2.5-coder-1.5b-instruct-q4_k_m.apr
• qwen2.5-coder-1.5b-instruct-q4k.apr

The fit check stops apr before it runs on these .apr rows, so they have no CRUX result (#4668).

Not measured for 0.70.1

• Capability cells (model_ladder.sh --cells) were not measured on any host. At the measured ~45 min per model for
  26 models, they take 15-20 GPU-hours per host. This is a gap, not a pass: #4667
• The full CRUX sweep (beyond the smoke above) and CPU lanes: nightly, per the scope.

The smoke at the release commit (attached to this release)

Both GPUs were smoked again with the binary built at the release commit 04332f8 (source identical to
cc4463f), with nothing else on either GPU. Every cell was compared with the receipts smoked at 7d2dc28 on its
whole key (host, verb, thinking, rung, prompt, model sha256, and mode and route for serve cells).

host	apr	cells	result	vs 7d2dc28
lambda	apr 0.70.1 (04332f838c)	134	134 GREEN, 0 RED (PASS)	MATCH 134/134 on apr's verdict
gx10	apr 0.70.1 (04332f838)	134	134 GREEN, 0 RED (PASS)	MATCH 134/134 on apr's verdict

The receipts are the files attached to this release: lambda-gpu.json and gx10-gpu.json (crux-inference-receipt/v1).

The smoke at the release commit is the release evidence; one older gx10 cell is superseded. In the cc4463f
smoke vLLM did not start for one gx10 cell (run, thinking off, easy rung, prompt ctl-code-add, model
aaf42c8b7c3c…): "AssertionError: Error in memory profiling. Initial free memory 99.12 GiB, current free memory
102.63 GiB". Hugging Face and llama.cpp agreed with apr there. In the smoke at 7d2dc28 every reference engine
started for that cell and agreed with apr, and apr's answer (def add(a, b): return a + b) and its GREEN verdict were
the same in both runs. A cell differs only if apr's verdict differs, so a reference engine that started only in the
newer run is not a difference (operator ruling C286).