MM-01Quality-constrained quantizationModel efficiency · Preparation blocked
The question
Find smallest same-model Q4/Q5/Q6/Q8/high-precision format within locked-task noninferiority margin; report quality, RAM, latency, CPU cost.
Next step in the dated plan
close the stopped preparation honestly; inventory available references/licenses without loading models; revise the proposed comparison to fit actual formats and envelope.
Completion and acceptance criteria
identical model/settings, independent paired evaluator, multiplicity-adjusted interval and margin frozen before run; disclose missing reference formats and budget mismatch. Any placement measurement may be an optional covariate only if independently warranted within an admitted scope; it is not a new prerequisite.
Evidence and limits
Preparation was blocked. No scientific outcome is reported; the comparison design and available formats need reconciliation.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-02Layer-sensitive mixed precisionModel efficiency · Research design
The question
Sensitivity-guided per-layer/per-tensor precision (attention/FFN, first/last layers) versus uniform at equal bytes; select on development, confirm sealed.
Next step in the dated plan
specify sensitivity score, deterministic packing and equal-byte rule; validate map on tiny fixture.
Completion and acceptance criteria
exact byte equality, reproducible layer map, sealed noninferiority and worst-task results.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-03Calibration under domain shiftModel efficiency · Research design
The question
Generic/task-matched/random/shifted calibration for importance-matrix quantization; assess domain shift and reserved tasks.
Next step in the dated plan
define disjoint calibration/development/sealed domains, sampling and dedup audit before calibration.
Completion and acceptance criteria
provenance/overlap report, same model and budget, shifted-domain uncertainty, reserved tasks remain sealed.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-04Structured pruning and layer removalModel efficiency · Research design
The question
Compare block/channel/layer and unstructured removal; distinguish CPU pruning feasibility from execution.
Next step in the dated plan
freeze mask schema and baseline; check mask and byte counts on tiny fixture before runtime comparisons.
Completion and acceptance criteria
qualified CPU timing includes conversion/index overhead; matched quality and noninferiority; zero count alone is not speed.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-05Sparse CPU kernel break-evenModel efficiency · Research design
The question
Test sparse layouts, activation sparsity, bit-slice reuse and zero-skipping versus dense after index/dequantization costs.
Next step in the dated plan
lock numerical microbenchmark/oracle; model overhead before small shape grid.
Completion and acceptance criteria
oracle agreement, overhead-inclusive repeat timings and workload-specific break-even; no one-shape generalization.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-06Native ternary models on CPUModel efficiency · Research design
The question
Reproduce compatible released ternary/BitNet checkpoint inference with specialized lookup CPU kernels; compare usefulness per byte/CPU-second.
Next step in the dated plan
verify primary release, license, tokenizer, format and kernel compatibility via metadata.
Completion and acceptance criteria
exact artifact identity, supported CPU path, correct output contract, task/resource-normalized result.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-07MoE versus dense CPU economicsModel efficiency · Research design
The question
Matched-usefulness MoE/dense comparison including resident weights, active bytes, routing cost and throughput.
Next step in the dated plan
specify utility matching and account total resident bytes including inactive experts/runtime buffers.
Completion and acceptance criteria
matched utility, full RSS/routed bytes, routing overhead and throughput; active params do not stand in for memory.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-08Expert hot-set caching and prefetchModel efficiency · Research design
The question
Replay traces for locality, ideal-cache bounds, LRU/selective residency/prefetch before runtime hooks.
Next step in the dated plan
define trace provenance; build simulator and ideal-cache bound first.
Completion and acceptance criteria
misses, bytes, memory and delay across same traces; stop if upper bound misses useful threshold.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-09DRAM and Optane executionModel efficiency · Research design
The question
Verify physical memory population and Optane mode, characterize working-set sizes/access patterns, and distinguish Memory Mode's DRAM-cache behavior from separately addressable App Direct placement; no BIOS changes. Current scope: Optane testing is deferred pending replacements. Do read-only prior-record and method review only; no fresh hardware measurement, admission request, probe, or hardware route is proposed.
Next step in the dated plan
inventory existing safe records/method definitions and record exactly what is unknown; leave physical state unasserted.
Completion and acceptance criteria
source-bound method note clearly separates historical observation from current state; no test or hardware claim. Any future measurement plan requires Jim to revisit the scope and fresh host qualification before a separate admission.
Evidence and limits
Optane measurements are deferred. The source leaves the physical memory configuration unverified.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-10Memory mapping and layer streamingModel efficiency · Research design
The question
Cold/warm loads, page cache, SSD/SATA, out-of-core streaming/prefetch under caps; preserve models and recovery storage.
Next step in the dated plan
define read-only inventory and synthetic file-layout fixture; cold-cache protocol must not disrupt services.
Completion and acceptance criteria
RSS/cache/bytes/startup/per-layer latency/I/O contention and correctness; no source or recovery-store writes.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-11KV-cache quantizationContext & memory · Research design
The question
K/V precision and context length versus RAM/speed/retrieval/reasoning/code accuracy, against higher precision on identical inputs.
Next step in the dated plan
verify CPU runtime format support; freeze lengths and long-context cases; test tensor round-trip.
Completion and acceptance criteria
same tokens/model, K/V separately recorded, verified CPU path and task/memory/latency strata.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-12Evidence-aware KV retentionContext & memory · Research design
The question
Sink/heavy-hitter/salience/recency/hierarchical/semantic retention at fixed memory versus full KV and truncation; CPU ports conditional.
Next step in the dated plan
portability review, fixed-byte policy and synthetic attention/retrieval fixture.
Completion and acceptance criteria
equal memory ceiling, both baselines, retrieval-position strata, document CPU feasibility.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-13Persistent prefix and KV librariesContext & memory · Research design
The question
Exact prefix reuse, COW branching, save/restore/rehydration; complete cache key and invalidation.
Next step in the dated plan
schema includes model/tokenizer/runtime/settings/prefix hashes; adversarial mismatch fixture.
Completion and acceptance criteria
identical outputs on reuse; reject stale/incompatible caches; isolation and resource cost; no sealed answers cached.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-14CPU paged-KV allocationContext & memory · Research design
The question
CPU page allocation/fragmentation under sessions; GPU PagedAttention is reference only.
Next step in the dated plan
synthetic allocator fixture over small page/lifetime grid before inference.
Completion and acceptance criteria
fragmentation/tail latency/RAM waste/quality versus contiguous allocator; label simulated vs real runtime.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-15Context selection and summary policiesContext & memory · Research design
The question
Raw, truncation, retrieval, structured/evidence summaries, ordering, lost-middle and TTL compaction at equal tokens on long tasks.
Next step in the dated plan
freeze task provenance and token counter before summaries.
Completion and acceptance criteria
equal budget, later-task correctness, source traceability, position strata, omission/compression errors.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-16NUMA and CPU placementCPU systems · Engineering preparation
The question
Socket/local/interleaved, first touch, replica, affinity, thread and hugepage comparison on verified topology while protecting production allocations.
Next step in the dated plan
preserve seven-file preparation and 31/31 fixture; resolve active-time uncertainty and 14-call retry reserve capacity; obtain fresh topology/allocation.
Completion and acceptance criteria
independent fixture replay, current reservation, no production interference, retry plan fits batch.
Evidence and limits
Preparation includes a reported 31/31 engineering fixture. The science comparison remains NOT_TESTED; this is not a NUMA performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-17Instruction sets and packed CPU kernelsCPU systems · Research design
The question
Supported AVX2/AVX-512/VNNI, packed kernels/codebooks, W8A8 vs weight-only; no assumed AMX/BF16.
Next step in the dated plan
read-only CPU/runtime feature audit and independent numerical oracle; fixture only verified paths.
Completion and acceptance criteria
runtime feature check, oracle agreement, unsupported variants excluded, same-shape/settings timing.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-18Separate prefill and decode budgetsCPU systems · Research design
The question
Separate prompt-processing/generation bottlenecks and tune within fixed total CPU resources, including multi-prompt/end-to-end.
Next step in the dated plan
fix total CPU budget and prompt lengths; define phase/queue/end-to-end accounting.
Completion and acceptance criteria
no borrowing beyond total, tails and quality by prompt/output length; phase speed not substituted for user-visible gain.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-19Deadline-aware batchingCPU systems · Research design
The question
Serial/micro/continuous batch on mixed deadlines; throughput, tails, peak KV and correct completion.
Next step in the dated plan
deterministic arrival/deadline replay; simulate policies before concurrent run.
Completion and acceptance criteria
correct tasks by deadline, p50/p95/p99, peak RSS/KV, expired work and fairness; not throughput alone.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-20Activation workspace reuseCPU systems · Research design
The question
Buffer reuse/allocation removal/recompute tradeoffs for activation memory on small fixtures.
Next step in the dated plan
freeze tiny operator inputs, oracle tolerances, and allocation accounting.
Completion and acceptance criteria
oracle tolerance, live bytes/allocations, no request-state leak, recompute overhead included.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-21Bandwidth-guided autotuningCPU systems · Research design
The question
Measured roofline; compare grid, bandwidth rules, tiny predictor with independent timing/correctness.
Next step in the dated plan
stable bandwidth probes and bounded grid; independent heldout shapes; predictor later.
Completion and acceptance criteria
oracle, tune overhead, noise bounds, unseen shapes, no out-of-regime claims.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-22Energy-aware schedulingCPU systems · Research design
The question
Measured or explicitly estimated energy per successful task, energy-delay, thermal drift and scheduling; no power-limit changes.
Next step in the dated plan
list permitted read-only sensors/units; test scheduling on simulated traces.
Completion and acceptance criteria
sensor provenance/accuracy, sampling overhead, task denominator, thermal/quality results; estimates labeled and no BIOS/power writes.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-23Exact speculative decodingInference & reasoning · Research design
The question
Draft/target CPU pairs, draft lengths, selection, acceptance and placement vs target-only with exact contract.
Next step in the dated plan
freeze greedy/sampling equivalence fixture and tokenizer compatibility.
Completion and acceptance criteria
exact output/distribution contract, rejected work/acceptance, end-to-end quality/latency and contention.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-24Prompt-lookup and n-gram draftingInference & reasoning · Research design
The question
Lookup, n-gram, suffix and copy-aware drafting vs target decoding.
Next step in the dated plan
deterministic repetition fixture, cache identity/reset, simulated match rates.
Completion and acceptance criteria
exact target contract, bookkeeping cost, acceptance, memory, task quality; synthetic repetition separated from real workload.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-25Self-speculation and early exitsInference & reasoning · Research design
The question
Released compatible layer-skip/draft-head and confidence depth; training/architecture changes are separate.
Next step in the dated plan
checkpoint/license/runtime inventory and intermediate-state support; synthetic policy fixture.
Completion and acceptance criteria
compatibility proven, numerical and locked-task quality, calibrated early-exit failures; no training absent admission.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-26Small-to-large local model routingInference & reasoning · Research design
The question
Always-large/rules/confidence/specialist cascades at matched utility/resources.
Next step in the dated plan
freeze labels/cost, replay fixed rules, then fit confidence on development only.
Completion and acceptance criteria
escalation and total cost, worst-group accuracy, false-safe routing, independent heldout gate.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-27Adaptive reasoning stopsInference & reasoning · Research design
The question
Fixed vs difficulty-aware reasoning budgets and step-verified stops on exact-checker tasks.
Next step in the dated plan
define budget, stop signal and premature-stop categories; simulate on development.
Completion and acceptance criteria
correctness, early-stop failures, tokens/time, checker independence by difficulty; explanation length is not correctness.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-28Verified candidate searchInference & reasoning · Research design
The question
Fixed/adaptive self-consistency, verifier-gated search, diversity and stopping at equal budget.
Next step in the dated plan
cap total candidate tokens/CPU; fixture known false positives; freeze stop policy.
Completion and acceptance criteria
every candidate cost, diversity, verifier false accepts, correctness under single-candidate budget match.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-29Grammar-constrained generationInference & reasoning · Research design
The question
Prompt vs grammar/schema constraints including lexicon/output masks; separate valid format from useful correctness.
Next step in the dated plan
freeze schemas, valid/invalid cases, parser oracle and runtime support.
Completion and acceptance criteria
schema validity, extraction correctness, usefulness and overhead distinct; valid format alone is not success.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-30Verifier-filtered local distillationInference & reasoning · Research design
The question
Locally generated, independently verified examples for student, rejection sampling, compressed reasoning or drafter; CPU training feasibility first.
Next step in the dated plan
cost worksheet, teacher/student provenance, synthetic checker examples only.
Completion and acceptance criteria
feasible cap, dedup/no leakage, independent filtering, heldout student vs unchanged/larger model; stop if infeasible.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-31Small-model CPU LoRALearning & adaptation · Research design
The question
CPU LoRA time, optimizer/gradient/activation RAM and tiny heldout gains vs unchanged/larger models.
Next step in the dated plan
derive cost from architecture; tiny fixture and stop cap before training.
Completion and acceptance criteria
peak RSS/CPU time/adapter bytes, heldout gain and forgetting, baseline parity, contamination audit.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-32Memory updates versus weight updatesLearning & adaptation · Research design
The question
Retrieval/adapters/RAG fine-tuning/verified bank on new facts/procedures with stale-fact/forgetting checks.
Next step in the dated plan
freeze fact add/update/delete timeline and create provenance-aware fixture; weight updates wait for training gate.
Completion and acceptance criteria
freshness, provenance, supersession/deletion, retention and resource cost; no adaptation/evaluation leakage.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-33Tiny recurrent and state-space memoryLearning & adaptation · Research design
The question
Tiny CPU recurrent/state-space probes vs matched transformer/external memory on synthetic checkable sequence tasks.
Next step in the dated plan
exact sequence oracle and fixed time/RAM feasibility.
Completion and acceptance criteria
synthetic labeling, matched size/compute, scaling/recall failures, fixed stop; no broad reasoning extrapolation.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-34Independent task parallelismDistributed systems · Research design
The question
Task replicas, pipelines, heterogeneous assignment/work stealing/distributed farming vs serial and model sharding.
Next step in the dated plan
simulate synthetic queues; verify per-node allocation interface before distributed trial.
Completion and acceptance criteria
correct tasks per total CPU-second, transfer/scheduling overhead, fairness, recovery and node receipts.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-35CPU layer and pipeline shardingDistributed systems · Research design
The question
CPU layer/RPC splits on explicitly approved verified nodes vs one-node/constrained baseline.
Next step in the dated plan
derive layer/activation bytes; simulate split/failure locally; require exact host approval before RPC.
Completion and acceptance criteria
per-node RSS, serialization, sync/tail latency, output equivalence and node-loss recovery.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-36Tensor-parallel communication feasibilityDistributed systems · Research design
The question
Derive/measure bytes per token, link/sync cost before implementation; compare whole-task and pipeline.
Next step in the dated plan
analytic model from verified shapes plus synthetic latency sensitivity.
Completion and acceptance criteria
independently checked bytes/collectives, measured link separately labeled, overlap and decision threshold.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-37Specialist committees and verificationDistributed systems · Research design
The question
Cross-family error correlation, voting, draft/verify, independent chains/ensembles at equal compute; count correlated false accepts.
Next step in the dated plan
checker-backed tasks and all-in compute; simulate correlated error fixtures.
Completion and acceptance criteria
joint errors/false accepts with intervals, latency/correct-task cost, single-model baseline, verifier independence.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-38CPU hybrid retrieval and rerankingRetrieval & memory · Research design
The question
BM25/vector/hybrid/rerank/graph/adaptive depth at fixed corpus/query/context, including storage indexes and parse cache.
Next step in the dated plan
freeze corpus/version/labels/parser and cache invalidation fixture.
Completion and acceptance criteria
recall/nDCG plus answer/citations, equal context, build/query cost, stale/delete tests; labels independent.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-39Temporal source-linked evidence graphsRetrieval & memory · Research design
The question
Flat retrieval/tables/graphs with source links, dates, supersession, reliability, contradictions, citation checks; prove graph adds beyond database.
Next step in the dated plan
claim schema and validator; create contradictory/superseding fixture; table baseline.
Completion and acceptance criteria
temporal/citation/contradiction correctness and maintenance cost, demonstrated graph benefit.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-40Hierarchical episodic and procedural memoryRetrieval & memory · Research design
The question
Hierarchical summaries, episodic/semantic, verified procedures/scaffolds across sessions with provenance and later correctness.
Next step in the dated plan
keyed multi-session fixture; every summary statement keeps source ID; fixed memory budget.
Completion and acceptance criteria
recall, provenance through hierarchy, procedure execution, stale conflict handling, equal tokens/memory.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-41Incremental indexing and deletionRetrieval & memory · Research design
The question
Additions/changes/deletes/staleness/parser invalidation versus full rebuild on seeded drift.
Next step in the dated plan
content hash/version/tombstones and deterministic change sequence; clean rebuild oracle.
Completion and acceptance criteria
incremental result equals rebuild, deleted records unavailable, caches invalidated; quality and cost.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-42Executable database queriesRetrieval & memory · Research design
The question
SQL/schema-aware queries, EXPLAIN/validation and deterministic records versus prose retrieval.
Next step in the dated plan
freeze schema/units/exact oracle; readonly query fixture with malformed/null cases.
Completion and acceptance criteria
query correctness, units, record provenance, safe limits, query vs answer errors separate; cannot mutate source/tests.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-43Calibrated abstention and confidenceRetrieval & memory · Research design
The question
Calibrate confidence/selective prediction and abstention for semantic uncertainty/source conflicts.
Next step in the dated plan
freeze splits, independent correctness/conflict labels, abstention utility.
Completion and acceptance criteria
calibration and risk/coverage, false confidence and subgroup/conflict intervals; no threshold tuning on test.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-44Verifier-backed local task solvingTool-using AI · Engineering preparation
The question
Model-only/deterministic/model-plus-tools on arithmetic, grounded QA, planning; equalize assistance and attribute gains.
Next step in the dated plan
preserve preparation and 45/45 engineering receipt; resolve independent replay, capacity/CPU access, dispatch and evaluator gaps.
Completion and acceptance criteria
independent answers, equal help, sealed set, model-only vs full-system separation.
Evidence and limits
Preparation includes a reported 45/45 engineering receipt. The scientific study remains NOT_TESTED with zero model calls or uploads.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-45Constrained program synthesisTool-using AI · Research design
The question
Generate against fixed specs/tests/schemas/resource limits; compare one-shot/verified search without test editing.
Next step in the dated plan
immutable tests and sandbox contract; adversarial attempts to alter/read tests.
Completion and acceptance criteria
expected outputs external, hidden property tests, process/network/filesystem containment and resource caps.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-46Reproducible data-analysis agentsTool-using AI · Research design
The question
Local SQL/code analysis vs independent numeric answers on locked data/schema/units; rerunnable scripts and receipts.
Next step in the dated plan
tiny frozen synthetic data/oracle, schema/units and rerun manifest.
Completion and acceptance criteria
answer within prespecified tolerance, rerun from locked inputs, units/filters/source hashes recorded; executable vs model-only separated.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-47Cost-aware experiment selectionTool-using AI · Research design
The question
Compare random/round-robin/Bayesian/bandit/information/resource planners and fixed/model-directed selectors using simulation first.
Next step in the dated plan
independent objective and synthetic/history queue replay, fixed cost/stops.
Completion and acceptance criteria
utility/regret per resource, budget violations, stability; planner cannot certify its own score.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-48Bounded failure triage and code repairTool-using AI · Research design
The question
Injected isolated failures; restart-only vs log diagnosis/limited patch; preserve failures and untouched regressions.
Next step in the dated plan
immutable known-fault fixtures, exact allowlist, at most two repair attempts.
Completion and acceptance criteria
hidden untouched regression groups pass, expected tests cannot be changed, all attempts kept.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-49Checkpoint recovery and result commitmentResearch reliability · Engineering preparation
The question
Isolated interruption/journal/lease/identity/idempotent commit/replay; test actual interruption types.
Next step in the dated plan
preserve interface handback and six NOT_RUN groups; specify minimal shared dispatch/reservation/commit contract; fixture crash/timeout/stale worker only.
Completion and acceptance criteria
atomic accounting, lease fencing, idempotent receipts, raw retries; do not claim reboot support untested. Existing evidence found no shared dispatch path.
Evidence and limits
An interface handback exists, with six test groups NOT_RUN. No shared execution-and-commit path was established by this record.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-50Protected evaluation and research integrityResearch reliability · Engineering preparation
The question
Immutable holdout, separate evaluator, anti-tamper/contamination and independent receipts; candidates cannot rewrite tests or self-certify.
Next step in the dated plan
preserve hold recommendation; resolve evaluator custody, atomic accounting, path/hardlink/symlink safety, stale fencing, independent replay.
Completion and acceptance criteria
independent identity/custody, cumulative usage, CPU/model/helper evidence and outside replay.
Evidence and limits
Evaluation-integrity preparation exists. Independent replay and evaluator custody remain unresolved; the record recommends continuing the hold.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-51Learning from negative experimentsResearch reliability · Research design
The question
Complete positive/negative ledgers vs positive-only/stateless planners for repeated errors and justified next steps.
Next step in the dated plan
append-only ledger schema and synthetic repeated-failure/exception histories.
Completion and acceptance criteria
full denominator, repeat mistake rate, next-step calibration, no dropped nulls/inconclusives.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-52Unattended endurance and driftResearch reliability · Research design
The question
Bounded soak for leaks, pressure degradation, task/input drift, injection and restart, with quality and telemetry.
Next step in the dated plan
finite synthetic endurance/fault simulator; specify supervisor stops before live test.
Completion and acceptance criteria
resource trends, graceful stop, cursor integrity, injection containment and quality; simulation labeled.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-53PC hybrid-core schedulingCPU systems · Research design
The question
Compare P/E placement/thread/prefill/decode on qualified PC.
Next step in the dated plan
literature/protocol only until a PC is explicitly available/authorized.
Completion and acceptance criteria
verified topology/affinity, same binary/model/task, thermal/resource record; no Xeon extrapolation.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-54CPU runtime and build comparisonsCPU systems · Research design
The question
Compatible CPU runtime/build/compiler/allocator/portability comparisons after same-model/settings equivalence.
Next step in the dated plan
inventory commits/backends/features/formats; output oracle; no production install/change.
Completion and acceptance criteria
equivalence first, fallback path and provenance, performance separate; CUDA presence does not qualify CPU.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-55Distributed prefill and KV migrationCPU systems · Research design
The question
Simulate then conditionally test split prefill/cache transfer/checkpoint migration vs one-node/recompute.
Next step in the dated plan
synthetic KV identity/transfer model and approved-node inventory; no network test before grant.
Completion and acceptance criteria
cache compatibility, bytes/time, recompute baseline, output equivalence, per-node receipts/recovery.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-56CPU model and adapter mergingLearning & adaptation · Research design
The question
Merge compatible task-arithmetic/TIES/evolutionary checkpoints/adapters; compare every constituent and protect task holdouts.
Next step in the dated plan
identity/license compatibility; validation-only selection; estimate merge/repack costs.
Completion and acceptance criteria
per-task/worst-task vs all constituents, heldout untouched, CPU I/O recorded.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-57Continual learning and forgettingLearning & adaptation · Research design
The question
Sequential adapters/replay/regularization/routing on fixed sequences; old-task forgetting, new quality, cost/errors.
Next step in the dated plan
freeze arrival sequence/splits; tiny exact fixture; compute feasibility before training.
Completion and acceptance criteria
forgetting curve/new gain/routing/cost and no omitted degraded tasks or leakage.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-58CPU preference optimizationLearning & adaptation · Research design
The question
Tiny CPU DPO/preference and quantized adapter-scale vs SFT/unchanged on approved deterministic preferences.
Next step in the dated plan
verify fixture authority/provenance, derive cost, freeze arms; stop if infeasible or labels unavailable.
Completion and acceptance criteria
labels unchanged, compute matched, heldout behavior/regression and resource record.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-59Formal proof verificationFormal verification · Research design
The question
Local generate-check-repair with real proof checker; informal text is not proof.
Next step in the dated plan
read-only tool/version/library discovery and tiny public theorem fixture.
Completion and acceptance criteria
checker-accepted artifact and theorem deps, reproducible env, bounded repairs, compute per checked result.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-60CPU document extraction and OCRDocument intelligence · Research design
The question
Deterministic parsing/table extraction/CPU OCR or vision/local model on frozen documents with field keys and regions.
Next step in the dated plan
define frozen set/independent keys, inspect CPU support, build region fixture.
Completion and acceptance criteria
field precision/recall, region, coverage/unsupported fields and parser/OCR/model effects separate.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-61Embedding and index compressionModel efficiency · Research design
The question
Embedding quantization/vector-index compression/embedder adaptation at fixed corpus/query conditions.
Next step in the dated plan
freeze corpus, queries, labels, embedding model and build; compression round-trip fixture.
Completion and acceptance criteria
recall/nDCG/downstream correctness, RAM/storage/latency/build; adapt only development queries.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-62Low-rank weight compressionModel efficiency · Research design
The question
CPU low-rank approximation plus quantization vs uniform at equal bytes, including decomposition/repacking and task quality.
Next step in the dated plan
tiny matrix oracle; compatible tiny model and full cost estimate; rank selected on development.
Completion and acceptance criteria
total factor bytes, conversion CPU/RAM, inference support and heldout quality; matrix error is not task quality.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-63Dense-to-sparse MoE conversionModel efficiency · Research design
The question
Probe FFN activation overlap, clustering, gating and upcycling; post-hoc quality and speed are separate hypotheses.
Next step in the dated plan
development activation overlap and routing simulator; keep conversion distinct.
Completion and acceptance criteria
heldout routing stability, bytes/params, quality and measured speed separate; no conversion/training absent grant.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-64Sleep-time and anticipatory computationContext & memory · Research design
The question
Verified background summaries/likely answers/caches vs on-demand at equal CPU, measuring hit value, staleness, correctness and waste.
Next step in the dated plan
simulate demand traces and TTL/invalidation; no persistent worker until admitted.
Completion and acceptance criteria
useful hit/task rate, stale harm, total background+foreground CPU, invalidation correctness.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-65Statistical power and adaptive evaluationResearch methods · Research design
The question
Paired uncertainty, variance, seeds, sequential/multiplicity and subsampling vs frozen full suite without relaxing criteria post-result.
Next step in the dated plan
estimands/units from evaluator contract; simulate declared effect/power and freeze stop method.
Completion and acceptance criteria
interval/coverage and error calibration, all seeds/items counted, fixed threshold/no peeking.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-66Programmatic tasks and contamination controlsResearch methods · Research design
The question
Schema tasks with independent oracles; dedup, consistency, train/test separation and contamination; no weak-oracle tasks in acceptance.
Next step in the dated plan
seeded generator and oracle provenance; near-duplicate/split overlap fixture.
Completion and acceptance criteria
independent correctness, deterministic regeneration, contamination audit, uncheckable synthetic tasks explicitly exploratory.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-67CPU-scale paper reproductionResearch methods · Research design
The question
Select CPU-feasible published work, verify primary source/code, preserve assumptions and reproduce principal result; GPU-only excluded unless validated CPU port.
Next step in the dated plan
literature screen for data/license/protocol/compute; deviation table before resource request.
Completion and acceptance criteria
primary-source trail, protocol/data/seed/metric match, deviations, independent reproduction/uncertainty; demo not reproduction.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-68Input-adaptive precisionResearch methods · Research design
The question
Tiny CPU token/layer precision and activation thresholds vs fixed precision at matched average cost; isolate numerical and switching overhead.
Next step in the dated plan
deterministic development policy, switch/numerical oracle, fixed matching rule.
Completion and acceptance criteria
numerical bounds, measured switching overhead, matched average resources, sealed noninferiority; no holdout tuning.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
MM-69Funding and research evidence packagesResearch methods · Research design
The question
Research-only funding map and evidence packages; validate deadline, eligibility, org requirements, cash vs credits, unknowns; draft only, no contact/submission.
Next step in the dated plan
source/date schema; official program research; map only accepted evidence and gaps.
Completion and acceptance criteria
dated primary citations, unknowns visible, award/application not implied, no outreach; accepted experiment receipts only.
Evidence and limits
Research design only. This source establishes no completed scientific experiment or accepted performance result.
Status date: 2026-10-14. Source: R640 Autonomous Operating Plan v1; retained privately.
No matching question. Clear the search to see the whole agenda.