Learned latch placement and transfer V1: initialization improved, arithmetic and planning favored

Question

Can feedback learn one compact shared local latch-placement scorer that transfers beyond its competent initialization and simple analytical arithmetic, and how does it compare with workload-aware global planning?

Hypothesis

Learning, larger-host/capacity transfer, context dependence and comparative utility are separate endpoints. The analytical rule and exact assignment controls are allowed to outperform the learned policy.

Registration

Exploratory local optimization freeze 0c1e93875f8e5a7abc182c22412a4fbb199519e337257c38a030e280662c4178 precedes all 512 candidate evaluations. Second selection seal 9860e0a246b3f067d57c7dbcfb4a2fc0daa2a068e674fd993cef79d4de8e725d binds all four policies and full lineage before test generation; not public preregistration.

Methods

Fixed inherited local tree growth; learned decision is distinct scalar latch placement on write-to-query tree paths. Actual charged local demand/bid reduction/reservation/differentiation and route installation use 25 bounded binary64 tanh-scorer parameters. No task-algorithm learning, oracle labels, hidden global allocation, future-event cues or online adaptation.

Baselines

D_initial reproduces prior competent boundary placement in the same shell. D_analytic bids immediate workload cost in the same shell. B_original retains exact competent weighted-Dijkstra builder source; B_cost independently chooses exact workload-optimal distinct latches on that tree. D-tree exact minimum is an evaluator-only diagnostic, never learner input.

Evaluation

Four seeded searches each train on eight L4/K4 hosts and validate on four disjoint hosts, balanced p0/.15 and both rho profiles. Twelve fresh groups cross L4/6 K4/8 p0/.15 and rho.25/.75: 192 contexts, 2304 final conditions including four intact, four mismatch and four shared controls. Primary equally weights six size-transfer settings, two profiles and twelve groups. All per-search/group/settings reported.

Resource ceilings

One local standard-library numerical CPU worker; 128 vectors/run exactly. Carry inherited 444.820 CPU/410.215 wall seconds against 3600/5400 ceilings. Finite caps:216 cells, bounded tree/route state, development bids/messages/retries, 1024 task messages/transaction, one interface slot,110000 episode ticks. 1 GiB sampled process-tree RSS,128 MiB retained study plus release. No GPU/agent/scheduler/paid service/framework/install.

Stopping conditions

All locality, sensitivity, baseline, objective/runtime, tiny exact-assignment, workload and replay checks passed before freeze. No scored source repair, extra search, seed, arm, budget or deadline tuning. All scheduled evaluations completed. Startup monitor denial retained; no worker launched. Semantic/oracle/locality/objective/replay defects would stop comparison; individual failures consume evaluations and retain penalties separately.

Executing code

Exact files in optimization-freeze.json remain unchanged. selection-seal.json binds selected coefficients, initial controls, full candidate lineage and unchanged source. B_original source/host/growth program are exact inherited bytes. Post-run saved-data audit/report/packing are outside the experimental source freeze.

Configuration

Python 3.14.8, standard library, binary64 coefficients in [-4,4]. Four tanh units plus explicit boundary prior; seeded Gaussian mutation/strict incumbent selection; prespecified validation. Weighted task V3 removes forced minimum queries, initializes K independent signs then samples64 subsequent events per ordinary episode. Transaction IDs remain 0..63, recycled only after settlement.

Complete outcomes

All 2,304 final gates and 4,608 compiled episode replays passed; 147,180 stationary queries, 147,732 stationary writes, 27,648 initialization writes and 50,688 diagnostic queries passed. Primary reductions against initialization: 8.40/15.00/2.46/14.74%. Against analytical local: −12.57/−5.56/−25.49/−5.72%; against B_cost: −14.57/−7.55/−28.15/−7.71%. Against original B: −0.09/+5.20/−14.59/+5.11%. Positive means learned lower cost. Context mismatch affects only searches 1/3. Narrow learning and average transfer supported; necessity/utility unsupported.

Realized resources

Complete campaign 131.887 CPU seconds. Conservative combined 717.284 CPU/667.415 wall seconds including inherited ledger and 120-second allowance. Sampled peak RSS 313999360 bytes. D retains mean 28.06 cells/283.72 persistent field bytes; B_cost 27.78/238.20. Mean D development 682.82 messages/21235.17 bytes; B_cost has no developmental transport and measured global planning cost. Exact J, sampled means, setup, replay and full ledger are separate downloads. Separate static packing used 10.675 CPU seconds, with OS peak process RSS 630505472 bytes; scientific sampled peak remains separately labelled.

Limitations

Only latch placement is learned; topology, task and primitives are fixed. All four policies lose to cheap local arithmetic and workload-aware planning on mean transfer cost. Two policies ignore rho cues; two depend on them with occasional beneficial mismatch retained. Four searches and crossed shared references do not justify broad confidence or equivalence claims. No superiority over untested optimizers, entire-program compression, unknown-future-workload inference or DNA-like growth claim.

Reproducibility

Full result-first report and tables in report.html/report.md. Restore lossless exact-byte evidence with python3 unpack_evidence.py --out /tmp/latch-evidence. In restored directory: python3 reproduce.py smoke --out /tmp/latch-smoke; or bounded full command python3 reproduce.py full --out /tmp/latch-campaign. Saved-data audit: python3 run_job.py audit. First selected policy plus controls/mismatch matched one paired fully held-out host in disposable author rerun; author checks are not independent reproduction.

Revision record

First release of one new bounded exploratory learned-placement study; earlier constructor-only milestones/releases/failures preserved. Scientific execution, content/visual inspection, saved-data audit, executable author rerun and publication delivery are separate. Lossless static JSON packing keeps exact originals without modifying publisher/policy. Publisher-owner overview note sent separately; no shared-site edits required. Revision1 failed an automated file-protocol false positive on the Python variable text as followed by a variable named file and a colon. Its unsealed bytes are preserved. Revision2 omits only the redundant plain experiment.py download; exact first-sealed driver bytes remain in the inspected lossless evidence pack. No scientific source or outcomes changed. Revision2 passed review and validation but failed the existing 32 MiB remote upload ceiling. Revision3 uses joint lossless tar/XZ/base64 JSON encoding for the same 44 exact source/evidence files; no publisher, policy, quota, source, scores or denominators change. All failed bytes remain intact.

Reviewed evidence & downloads

• report.md — derived. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• report.html — derived. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• resources.json — derived. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• analysis.json — derived. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• verification.json — derived. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• group-effects.csv — derived. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• optimization-freeze.json — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• selection-seal.json — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• selected.json — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• config.json — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• protocol.md — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• development-history.md — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• fixture-results.json — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• common.py — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• host.py — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• program.py — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• policy.py — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• local.py — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• legacy.py — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• planning.py — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• runtime.py — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• task.py — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• run_job.py — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• inherited-ledger.json — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• reaggregate.py — derived. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• reproduce.py — derived. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• unpack_evidence.py — derived. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• evidence-pack.json — derived. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• pack-inventory.json — derived. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• author-rerun.json — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• attempts.json — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• preservation-public.json — derived. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.
• publisher-owner-note.md — original. Explicit selected project-authored synthetic scientific evidence/source or derived report; authorized for this workstream. No external dataset, raw conversations, unrelated logs, repository history or separate primary state.

These are selected public derivatives; raw transcripts, internal configuration and private source are withheld. Hashes describe the actual downloads, not unmodified original evidence.

Artifact manifest and source provenance · Permanent revision URL

Substantive versions

• Revision 3