| MiMo r4, Oracle 7 |
Two planning outputs used a visually similar non-ASCII arrow where the earlier prompt did not require literal ASCII ->. A separate structured-edit output wrapped otherwise-valid JSON in a Markdown fence. |
The arrow cases are instrument ambiguity, retained without rescoring. The fenced JSON is a genuine format defect. The r4 result remains diagnostic rather than a promotion result; its independent SWE result also fails the campaign gate. |
| Oracle 8 instrument |
The updated prompt explicitly requires literal ASCII ->; the scorer is unchanged. |
This is a prompt clarification for a new oracle revision, not a repair or reinterpretation of Oracle 7 observations. Future native artifacts must identify the oracle revision. |
| MiMo r5 sampling |
The r5 recipe separately set server generation defaults to temperature 0.0 and top-p 0.95. Pinned runtime source describes greedy normalization for zero temperature. |
Keep this configuration change distinct from the Oracle 8 prompt correction. It supports the recorded r5 configuration only and does not revise r4 results. |
| MiMo r5 SWE |
The native resolve rate is 1.0 only across the two graded submissions. Three assigned tasks lack submissions; the official grader is incomplete and the required Pylint sentinel failed. |
Report campaign outcome as 2/5 assigned tasks, not 2/2 graded-only. The denominator boundary rejects ranking or promotion from the native rate. |
| Mistral r2 |
The protocol run passed, but its measured memory policy exceeded the campaign ceiling. |
The r2 configuration is rejected on resource policy, regardless of protocol success. |
| Mistral r3 |
Reducing GPU memory utilization to 0.94 created bounded headroom, but the frozen nine-scenario, two-repetition Oracle 8 population had one parallel-tools failure: 17/18 in round one rather than the required 18/18 per round and 54/54 aggregate. |
The 0.94 change is a headroom remedy, not a reliability remedy. The configuration remains rejected; the retained native 17/18 pass flag must not be presented as campaign acceptance. |
| DeepSeek build r1 |
The cached compiler base reported Torch 2.14.0+cu130 where the build gate required Torch 2.13.0+cu130. |
Stopped before compilation; this is a toolchain-identity failure, not model quality evidence. |
| DeepSeek build r2 |
The exact toolchain gate passed, then compilation exceeded the managed 8 GiB memory cgroup. |
Resource bound raised to 16 GiB with no swap and one build job. This does not establish an SM120 source or kernel incompatibility. |
| DeepSeek build r3 |
Native compiles and CPU/import tests completed, but the final inspection helper imported the extension before Torch and failed to resolve a Torch shared library. |
The r4 build path imports Torch first for that lookup, retaining source pins and numerical assertions. CPU build correction is not GPU correctness evidence. |
| DeepSeek build r4 |
The bounded CPU build passed its recorded CPU/import and architecture gates. |
The separately retained Stage 2 r6 numerical gate passed. The CPU build result alone establishes no model quality. |
| DeepSeek GPU identity |
Runtime CUDA UUID text omitted the conventional GPU- prefix even though it represented the requested device. |
The harness now compares only a full canonical UUID with an optional prefix, preserving mismatch refusal. Later attempts verified the canonical identity before executing numerical groups; that closes the representation guard without claiming the full numerical gate passed. |
| DeepSeek JIT tmpfs |
The first numerical group passed; the next JIT-built shared object failed to map from the bounded /tmp tmpfs. Primary Moby source defaults to noexec, and the recorded mount request omitted exec; the old effective mount flags were not retained. |
The harness requests executable /tmp while retaining the 4 GiB cap, read-only root, nosuid, nodev, network isolation, limits, and cleanup. The r3 and r4 receipts proved effective executable tmpfs and completed the basic Engram shared-object checks. The complete numerical gate remained blocked by later, separate failures. |
| DeepSeek PID limit |
The r3 spin job exhausted its fixed 512-task cgroup (pids.current=512, pids.events.max=1) and reached its inner deadline. Frozen source requires two 256-worker tables plus runtime tasks; the captured 473 child threads are a post-exhaustion snapshot, not a baseline or exact peak. |
The reviewed fixed 1,024-task envelope retains before/after cgroup evidence. r4 completed the spin checks with zero task-limit events. The original wait site remains unlocalized; no partial r3 result transferred to r4. |
| DeepSeek Triton cache |
In r4, the next Engram/vLLM group tried to create /root/.triton under the read-only root and raised OSError: [Errno 30]. Seven failure markers were printed before the bounded output cap stopped the command; no completed JUnit count was retained for that group. |
The minimum correction sets Triton's supported cache directory inside the existing /tmp tmpfs and checks launch readback. Independent review verified the supported setting against retained pinned upstream source; the exact installed Triton revision remains unknown. The fresh r5 run passed the Engram/vLLM group (7/7) and registration (7/7), verifying the override on the pinned image. The final group then failed the separate strict skip gate; the complete numerical gate remained incomplete. This is a harness write-path failure, not demonstrated model or kernel numerical error. |
| DeepSeek optional benchmark selection |
r5 completed all six groups. The final pytest invocation reported 136 passed and six skipped, so the zero-skip guard rejected it. Four predeclared optional benchmark functions expanded to six nodes; absolute exclusion prefixes did not match collected relative node IDs. |
Independent source and environment review identified those six benchmark nodes. A new spec corrects only the four exclusion prefixes to cwd-relative IDs; source assertions, resources, and the zero-skip gate stay unchanged. r5 remains failed. The fresh r6 run selected exactly the prior 142-node set minus those six optional nodes and passed all 136, with zero skips or failures. Independent review accepted all six numerical groups; the result remains single-GPU kernel evidence only. Per-case JUnit identities were not retained in r5. |
| Engram microtimings |
Standalone source tests print host-node/spin timing ratios. Their fixture files are in the job's memory-backed temporary filesystem. |
Retain the raw output for reconstruction; it is not NVMe latency, serving throughput, end-to-end model speed, or comparative capacity evidence. |
| DeepSeek Stage 3 r7-r12 |
r7 stopped at host-memory admission; r8-a1 found its required metadata snapshot absent; r8-a2 reached dummy readiness but later proved an incomplete EXL3 residency proxy. Corrected r9 failed CUDA OOM in dense-mrow startup, r10 failed custom-all-reduce IPC initialization after capture progress, and r11 reached health but missed the frozen 4-GiB idle reserve. |
The narrow remedies were a bounded metadata pull, restoring original EXL3 controls, disabling dense-mrow, disabling custom all-reduce, and preserving the reserve gate. r12 then passed the bounded dummy-path integrity, containment, reserve, health, and one text-probe graph evidence gate. It remains dummy evidence only: no main-pack load, 327K, C4, image8, quality, throughput, or promotion claim transfers. |
| DeepSeek Stage 4 |
All six managed main-pack pull groups completed. The corrected r8 loader verified all 178 integrity pairs before engine start and passed the 64K/C1/image0 resource and basic 2/2 correctness gates. |
This is bounded real-weight evidence only. Full-role context, C4, vision, native quality, and promotion remain separate gates. |
| Pi selected-route acceptance |
An interim Pi fixture failed before its prompt because get_available_models emitted one catalog JSONL record larger than the client's 262,144-byte bound. get_state had already identified the selected provider/model and idle state. |
For the narrow selected-route read-turn acceptance, omit only the full-catalog request and its count; retain selected-state checks before/after, the exact read tool turn, agent_end, and settlement proof. Full-catalog propagation acceptance remains a separate contract and must keep its catalog digest check. This is a framing limit, not a model, provider, or prompt failure. |
| Router readmit invocation |
A direct .command attempt supplied only the command stem where the managed readmit surface requires explicit arguments. |
Treat this as a command-shape error, not an authentication failure. The subsequent managed preview and confirmed readmit receipt supply the operational evidence; no credential conclusion follows from the stem-only attempt. |