gem/docs/research/completion-audit.md

Gem objective completion audit

This audit keeps the project objective broader than the latest successful optimization. A status is complete only when the named current-state evidence covers the full requirement; planned commands and historical intent are not proof.

Requirement	Status	Authoritative evidence	Remaining proof
Standalone C inference server; production never links ggml/libllama or launches llama-server	Complete	architecture.md; native/bench/profile dependency-guard CTests; otool receipts in benchmark artifacts	Re-run the complete guards after every new production target is added.
Fixed dense Gemma 4 31B graph and strict GGUF/tokenizer ownership	Complete	60-layer/833-tensor contract in model.c; authentic header/tokenizer tests; malformed topology/extent tests	A new GGUF layout is out of scope until explicitly added and gated.
M4 Max-specific Metal execution and benchmark/profiling scaffold	Complete	../build/benchmark-protocol.md; gem-bench, stage/decode/attention profilers; raw three-repetition artifacts; adjacent backend oracle receipts	Continue applying the protocol to every promoted axis.
Small hax-compatible offline server and one-command lifecycle	Complete	gem agent; gem.hax_project_memory, gem.hax_native_server, and lifecycle tests; real-hax artifacts including tool loops and MTP autodiscovery; final selector-19 8K+ smoke returned exactly OK and reclaimed its port	None for the current kernel decision.
Exact multi-turn long-prompt reuse	Complete through 128K	Exact token-boundary cache implementation; 32K/64K receipts; audited 2026-08-26-server-gate-31b-split16-131072 with three fresh processes, exact 131,072-token reuse, and 26-token extensions	Re-run if cache identity or request tokenization changes.
128K memory capacity	Complete	artifacts/2026-08-24-context-capacity-31b: 31B F32 capacity and request probes on the physical 128 GiB host	Capacity alone does not prove cold TTFT or current-kernel speed.
Current 31B/F32 long-decode value optimization	Complete	Promoted selector 19; independent split screens; boundary/tail/repeat correctness; authentic 64K/128K whole-token gains of 22.62%/26.79%; 128K KV gain 32.20%	Re-open value-side work only if a new profile shows another justified ceiling.
Long-prompt progress and disconnect cancellation	Complete	Batch-boundary callback implementation; CPU fixture; authentic 2026-08-26-prefill-cancel-31b-131072 receipt with four keepalives, curl status 28, health recovery, and a successful same-process request	Re-run if batch boundaries or server streaming semantics change.
Fastest justified production prefill kernel through 64K	Complete	Selector 19 routes 31B/F32 post-4K global prefill to BQ8/BK64; isolated attention improved about 59% at 4K/64K/128K, evaluated graph improved 7.12% at 4K and 33.38% at 64K, and the 64K logit gate retained 0.99999958 correlation with 8.35e-7 NMSE	The current BQ8 binary still needs an uncontaminated 128K server acceptance gate.
Complete post-change test state	Complete for the retained diagnostic	All 69 normal and all 69 freshly rebuilt ASan/UBSan tests pass with BK128 available only through diagnostic selector 23; the adjacent backend-op oracle remains 4,793/4,793	Repeat after any production routing change.
End-to-end 128K production acceptance	Incomplete for current BQ8 production	One current BQ8 cold request and exact-prefix extension completed, but extreme unrelated host load contaminated the timings; the older audited three-run receipt predates the BQ8 prefill promotion	Run three fresh schema-4 repetitions on a host that passes load, one-second system-busy, AC-power, and thermal preflight, preserving raw JSON, gem print-config, digests, SSE/progress, and continuous host-monitor receipts.
Latest 128K prefill kernel cycle	Diagnostic open; production unchanged	BQ8/BK64 is promoted. Exact conditional rescaling was bit-identical but rejected and removed after sub-threshold results. BQ8/BK128 is correct, deterministic, uses 20,544 bytes of threadgroup memory, and is exposed only through diagnostic selector 23; its initial screen was host-contaminated	Run the quiet-host operator gate. If it advances, run the prepared selector-19/selector-23 evaluated-prefix/logit gate; otherwise remove BK128. Resolve it before rerunning the current-production 128K server gate.
Latest long-decode score/K cycle	Diagnostic ready; production unchanged	Selector 20 changes only global-decode K staging from tile1 to tile8 while retaining selector 19's split16 value path; threshold/tail parity and a fail-closed ABBA runner exist. The parked D512 fused split-KV design now records the measured baseline, ownership, bounded scratch, resource formula, and fail-closed gates without adding a selector.	After closing BK128 and current-production 128K acceptance, run the schema-2 64K/128K score-tile8 gate. Advance to authentic logits/model timing only if projected whole-token gain exceeds 3% at both lengths; otherwise remove selector 20 before implementing fused split-KV.

Current critical path

1. Preserve selector 19 with BQ8/BK64 as production and retain its completed 8K, 32K, and 64K server gates.
2. On a host that passes the frozen load/power preflight, run the seeded, interleaved BK64/BK128 operator gate with identical inputs and enough warmup to avoid frequency-ramp bias. Operator-gate schema 2 also requires AC power, no thermal/performance warning, complete digest/linkage receipts, and a positive paired 95% confidence lower bound at both long contexts. The clean-host launcher can capture the first sustained quiet window without killing or reprioritizing unrelated work; its observation receipt supplements rather than replaces the gate's host receipts.
3. Promote BK128 only if it clears the protocol threshold consistently. Otherwise remove the diagnostic pipeline. The evaluated-graph step is frozen in scripts/run-bq8-bk128-graph-gate.sh and independently revalidates the advancing operator artifacts, provenance, continuous host state, logits, paired timing, and configuration. A promotion also requires evaluated-graph/logit evidence, all local tests, and the adjacent backend-op oracle before a performance claim.
4. Run the schema-4 three-repetition 128K cold/prefix server gate against the resulting production binary, including raw JSON, gem print-config, digests, progress/SSE validation, and host receipts.
5. Run the already-implemented selector-20 score-tile8 diagnostic under its schema-2 clean-host gate. Remove it if it misses the model admission floor. If it passes, complete authentic logits and populated-model timing before any routing change. Only then consider the larger fused split-KV GQA score/K design.

The standalone engine is complete enough to serve hax natively, and BQ8/BK64 is the justified production prefill choice through the completed 64K gate. Current-production 128K acceptance is not yet complete: the only finished BQ8 run was host-contaminated, and BK128 is routed only through diagnostic selector 23 pending its clean operator and evaluated-graph gates. Seeded profiles cannot substitute for authentic prompt throughput or TTFT.