SCIENTIFIC COUNCIL: RUNTIME AND PUBLICATION REVIEW
Review date: 2026-11-02. Scope begins 2026-10-28 00:00 Europe/Madrid.
Reviewer: council_runtime, with an independent subreview of the resume archive.

RECOMMENDATION

Support the proposed manuscript focus: interpretable phase/timing fusion, explicitly
development-exposed quality adaptation, the elapsed-time intervention and negative
JEPA evidence. Runtime merits a main-text subsection and one compact table, with
numerical diagnostic detail in the supplement. Its contribution is a causal,
auditable implementation and a measurement lesson, not achievement of a hard
20 ms deadline, physical real-time operation, or novel cryptographic hashing.
Do not let operational completeness displace the scientific comparisons.

The runtime sources expose a useful failure mode: mean service time below the
query period does not imply deadline compliance after acquisition and queueing.
The best recorded replay delivers all 250 single-object updates with preserved
predictions but misses 20 ms on 201 of them. The failure denominator is central.

PRIMARY TABLE: EXACT MACHINE-READABLE SOURCE

artifacts/train40_runtime_council_20261028/paced/memory_cache_overlap_reverse1_250_v1/summary.json

Use the following prefix for all three latency rows:
  paced.latency_populations.populations.successful_all_object_ttc
Then deadline_response_ms, raw_availability_response_ms, service_span_ms.
Each dictionary supplies count, mean_ms, p50_ms, p95_ms, p99_ms, max_ms and
at_least_20ms_count. Values in milliseconds:

Clock                n    mean        p50       p95        p99        maximum
Scheduled arrival    250  23.9071148  24.538    29.143155  32.414209  35.3326
Payload available    250  20.7330316  20.8045   25.252990  27.627191  29.5930
Service execution    250  17.6103860  16.91275  21.426025  23.212746  25.9175

Deadline >=20 ms: 201/250 = 80.4%. Source fields:
  paced.latency_populations.counts.successful_all_object_ttc = 250
  paced.latency_populations.counts.failed_acknowledgements = 0
  paced.latency_populations.counts.dropped_updates = 0
  paced.latency_populations.counts.source_errors = 0
  paced.offered_updates = 250
  paced.offered_duration_s = 5.0
  paced.rate_hz = 50
  paced.actual_wall_duration_s = 5.0493202
  paced.last_successful_all_object_ttc_duration_s = 5.0036785
  separate_parity.passed = true
  milestone.milestone_with_parity_passed = false
  capacity_with_parity_passed = false
  paced.physical_goal_demonstrated = false

200/250 scheduled arrivals observed one earlier update unfinished; 50 observed
zero. Maximum prior work was one. This is finite-run backlog, not proof of
asymptotic instability or of stability. Never report inverse-mean throughput as
sustained throughput. The source explicitly labels inverse-mean fields otherwise.

Clock semantics: scheduled arrival starts BEFORE raw acquisition/copy/sealing,
includes acquisition delay and queue waiting, and ends with all three required
TTC scalars in host memory. The raw_availability field starts AFTER the payload
has been acquired/copied/sealed, despite its convenient name. Service execution
starts only when the consumer begins. Thus the second row should preferably be
labelled 'Payload-ready to host TTC', not 'sensor arrival to TTC'.

Hardware/configuration: RTX 5070 Ti Laptop GPU, Python 3.11.15, Torch
2.11.0+cu128, CUDA build 12.8, FP32, TF32 off, 2 Torch CPU threads, 8 RGB workers,
queue capacity 2, Windows build 26200. CPU model and GPU driver absent. Startup
outside arrival clock: raw preload 2.7692392 s, model/service init approximately
0.516 s, synthetic priming approximately 12.095 s for 7 layouts, 600 ms history
initialization approximately 0.132 s, source warmup approximately 0.051 s.
Use paper/revision/deployment_contract.json for the existing portable projection
and paper/tables/runtime_setup.tex as a cross-check, not as numerical authority.

Scope: preloaded raw CPU arrivals, single object, CPLA-low, five seconds. Future
raw bytes private to feeder; no future tokens/hashes/crops/features prepared.
Disk I/O is outside measured arrivals. The source bound is 2,611,872,720 bytes
under explicit 3 GiB budget; this is a numerical allocation bound, not measured
RSS, and excludes Python/HDF5/model/GPU overhead. No natural 60-second episode,
physical-camera source or admitted multiobject paced run exists in this record.

EXACT REUSE AND CAUSALITY

Final profile at f046b791: PIL ROI-first; temporal prefix/top-k normalization;
query graphs; native producer batches; event encoder reuse off; RGB feature
reuse on; RGB object chunk size 1; parallel branches; event execution overlaps
RGB preparation; transport reverse-summaries. SHA-256 V1 immutable tokens stay
default. Crop cache, uint8 preparation, BLAKE3 V2 wiring and subpacket backend
are excluded from the selected paced profile. Changing defaults did not itself
produce another GPU measurement.

Immutable cache keys bind pixels, timestamps, ROI, model weights, runtime and
batch layout. Metadata/context estimates are recomputed; predictions are not
replayed from a cache. Raw gaps invalidate history rather than emitting stale
TTC. Synthetic shape priming uses no observations and preserves causal reset.
These are defensible method details, supported by code and receipts.

Prepared transport diagnostic:
artifacts/train40_runtime_council_20261028/producers/exact_transport_sensitive_v1/summary.json
  status COMPLETE; query_coverage_exact true; all_replay_gates_passed true;
  complete_update_admitted false; sustained_50hz_demonstrated false;
  two_object_admitted false; weights_unchanged and normalizers_unchanged true.
Seven query records, indices 0,137,488,505,514,533,553; 462 explicit calls across
original/reverse/reverse-summaries and serial/parallel routes. Per-call
records[*].replay_parity[*].feature_parity.exact validates four feature groups
bit-for-bit. TTC uses unchanged atol=rtol=1e-5; do NOT turn this into a claim
that every TTC is bitexact (the RGB TTC has nonzero rounding differences in
some reference comparisons). Includes initial calls, warmups, repetitions.
Already-prepared GPU producer timing excludes RAW, H2D and downstream context.

5 ms subpacket diagnostic:
artifacts/train40_runtime_council_20261028/event/subpacket_backend_sensitive_v2/summary.json
  all_backend_gates_passed, all_final_ttc_passed, all_canonical_inputs_passed true;
  latency_measured false; target_50hz_demonstrated false;
  three_quarters_latency_saving_claimed false; objects 1.
Three queries (0,137,553), common initial 600 ms filled by thirty original 20 ms
packets, then 120 5 ms ingests over another 600 ms. There are 31 complete
boundaries per query, total 93; all 558 persistent buffer comparisons and 465
scalar-state comparisons are exact; 93 representations exact; 6 final canonical
inputs and 18 final TTC components exact. Intermediate queries use fixed full
sensor ROI for diagnostics, only final anchor uses the real query ROI for TTC.
The required aggregate event budget remains per 20 ms, not per 5 ms. Partial
phases cannot predict; overflow or interruption latches invalid state until reset.
It has NO paced 200 Hz feeder integration or measured latency savings. Four
5 ms calls increase some transfers/dispatches; 75% savings would be invented.

NUMERICAL AND NEGATIVE RESULTS WORTH RETAINING

1. Compiled transport is not an admitted improvement. Native-log/native-match
fail final TTC on sensitive examples, despite close intermediate values. The
isolated coordinate investigation diagnoses 202 differing scalars on four
coarse cases: replacing only the two compiled linspace constructions by opaque
ATen calls gives 12/12 exact transport-summary cases (3,888 scalar comparisons).
Grid 16 differs at coordinates 6,7,8,9 by up to 1.4901161193847656e-08; grid 32
does not. This is conditioned on earlier preserved sampling/reduction boundaries,
not universal attribution of all compiled error or admission of a full model.
Source: artifacts/train40_runtime_council_20261028/compiled_pipeline/linspace_attribution_v1/summary.json
keys all_candidates_exact, all_controls_reproduced_exactly,
standalone_coordinates, records; final_ttc_admitted=false, latency_measured=false.

2. Failure acknowledgements can spuriously improve a latency average. The crop
cache replay contains 115 successful TTC updates, 133 failed acknowledgements
and 2 drops. Successful deadline mean is 27.5726547826 ms, p95 37.42381 ms;
mixed responses mean is 16.1513044355 ms. The latter is NOT prediction latency.
The 115 produced predictions preserve parity, but global coverage fails.
Source: artifacts/train40_runtime_council_20261028/paced/memory_crops_reverse1_250_v1/summary.json
same latency_populations paths as primary table. No matched A/B establishes
that cache or garbage collection caused the failure. Keep causal language narrow.

3. BLAKE3 only improves isolated CPU copy+hash: SHA mean 3.9973 vs BLAKE3 1.6342
ms/frame; paired reduction 2.3632 ms. Seven 6,912,000-byte frames, twenty repeated
pairs, 140 per-frame samples per algorithm/scope; they are not independent
episodes. V2 core implemented, no wiring or end-to-end admission. This is an
appendix engineering observation at most, not a novelty claim.
Source: docs/train40_runtime_council_20261028/FRAME_HASH_CPU_GENERATED.json and
artifacts/train40_runtime_council_20261028/frame_hash_cpu/idx553_v1/summary.json.

4. GPU-exact RGB crop passes seven preparation checks, but no end-to-end TTC
admission or paired win over PIL. First real preparation includes 6631.424 ms
compilation. Prepared miss 35.600-39.863 ms vs causal reconstructed-prefix
8.686-10.566 ms (three repeats). They are different work amounts, not a sustained
speedup comparison. Exclude from abstract and central contribution list.

INDEPENDENT RECOMPUTATION DURING THIS REVIEW

Loaded the stdlib-only paced_statistics.py and recomputed populations from raw
monotonic arrival/object/group receipts for all seven saved paced runs. Five
existing modern latency_populations match exactly. The two oldest runs contain
no stored population projection; modern recomputation gives:
baseline1_50_v1: 13 valid, 0 failed acks, 37 drops, deadline mean 999.5149231 ms,
p95 3817.13708 ms.
memory1_50_v1: 9 valid, 40 failed acks, 1 drop, mean 55.6078667 ms,
p95 74.72102 ms.
memory_cache_overlap1_50_v1: 27 valid,22 failed acks,1 drop, mean45.4657815,
p95 69.72315. The 50-update reverse pilot and lifecycle repeat both preserve
50/50 output, with p95 29.892095 and28.85508 ms respectively.
No GPU execution, training or new performance measurement was performed.

PUBLICATION AND CLAIMS TO REVISE

The old paper is careful about runtime. Preserve its deadline failure and
separate timing scopes. Tighten the wording of the second clock as above.
Its sentence saying timing evidence is aggregate-only reflects the portable
paper projection; the original paced source contains per-arrival and object
records, so the new manuscript can use them without fabricating a distribution.
Do not carry a stale present-tense 'unverified labels on 434 queries' claim into
the revision if later exact recording/target mapping now supplies those labels;
the transfer reviewer must resolve that evolution with the new source revision.
The October28 public benchmark score is a dated observation, not a current SOTA
claim. The complete-system comparison confounds upstream producer training,
history duration and modality. A 1,955-parameter fusion head is not a total
system-size or full latency claim. The 468-query positive-TTC development cohort
and 434-query competition-aligned cohort cannot be merged by scenario names.

The release commit89c74c34 added the existing manuscript, source ZIP, 99-source
registry and CI rebuild. Dependencies/newtx plain TeX and clean-clone portability
were fixed afterward. This is publication infrastructure, not fresh experimental
evidence. ADR0044 preserves25 historical runtime-bound wrappers byte-for-byte;
their presence/hashes are not proof of execution. Third-party FCWD illustration
in paper/ASSETS.md has no separately identified redistribution license; a new
original vector schematic avoids unnecessary image reuse.

COVERAGE

Identified56 recent commits on perf/train40-persistent-50hz. The runtime-path
history through release/phase-ttc-final-20261029 contains41 commits. The release
has81 unique changed .md paths over the interval,34 in the council tree.
Reviewed19 current council documents (narrative, contracts, proposals and
generated-results interpretation),2 evidence README files,13 archived council
documents by content deduplication and differing-version comparison. Nine
archived documents are byte-identical to current versions; differing COUNCIL,
FROZEN_RGB_V2_DESIGN and RGB_GPU_FINDINGS were compared line-by-line. Generated
result tables were interpreted against structured source receipts, including
the old/new failed-response semantics. Reviewed paper README/ASSETS, runtime,
reproducibility, discussion and setup-table source, ADR0044, key source diffs,
paced_statistics implementation and main runtime JSONs named above.
The resume-tree subreview is attached separately in a parent message. This
coverage is a targeted runtime/publication council review, not a claim to have
personally read every eAP/transfer document in the full multi-branch repository.

RESUME ARCHIVE SUBREVIEW (resume_docs)

The subreview read all37 changed Markdown paths in
docs/train40_runtime_resume_20261027 (24 unique byte contents, historical
snapshot differences included), its three doc commits471302d7,451cf090,2521de1a,
publication commits89c74c34,820e6a6f,bb4ef634,3434a3be and0f10312d delivery record.
Together with34 council paths, this covers71 runtime Markdown paths among the81
release changed paths. Transfer/adapter docs are assigned to the transfer lane.

Additional conclusions and exact source locations:

R1. Preserving event producer batching matters. Event endpoint reuse failed5/554
E_CTX outputs despite input parity. E reuse OFF/RGB reuse ON restores all-three
TTC parity on554/554, unchanged atol=rtol=1e-5. E_CTX bitexact; whole-system
parity.exact=false. Source:
docs/train40_runtime_resume_20261027/evidence/closure_20261028_002/corrected_all554_v1/summary.json
parity.components.E_CTX_TRAIN40.max_abs_error=0;
PHASE max_abs_error=0.0000667572021484375;
R_CTX max_abs_error=0.00031280517578125.
RAW-ready to TTC mean46.1917847 ms,p95 66.634525,max103.7058 across554 queries;
history rebuilt outside each clock, therefore not a paced-service comparison.

R2. Exact event prefix representation:121 spatial prefix planes,600ms history,
5ms boundaries, integer counts/moments with epoch-crossing correction and
original float64 interpolation. Exact95th percentile needs top
1+ceil((N-1)/20) values; N163840 gives k8193. Dense query553,5 measured samples:
one ROI bucket7.84078 ms versus prefix/top-k/query graph6.46498 ms;
20 translated artificial ROI73.7359 versus30.62268 ms. Representation exact.
Source under same archive: prefix/gpu_query_graph_v1/summary.json,
prefix/IMPLEMENTATION_NOTES.md, normalization/NOTES.md. No TTC or stream claim.

R3. Metadata cache opportunities:32 sequences,3191 scheduled,3129 accepted,
62 exclusions,650ms retention. Exact-ROI event reuse115/18774=0.6125%; effective
projection155/18774=0.8256%; full RGB-call reuse1719/3129=54.9377% accepted,
53.8703% scheduled. Native RGB spacing50ms (~20Hz), model triplets100ms.
Source compute_council/causal_cache_metadata_v1.json and
rgb_cache/metadata_reuse_v1.json under archive. One-object metadata proxy,
not a measured live cache rate.

R4. Compiled transport final-three-TTC gates pass5/7 default,4/7 no-FMA,5/7
no-FMA+exact-division; each rejected. Sources compiled_pipeline/
gpu_sensitive_v2/summary.json,gpu_no_fma_v3/summary.json,
gpu_no_fma_exact_div_v4/summary.json and PRECISION_VARIANTS.md.
All PHASE outputs pass but event/RGB constituent TTC fails, so retain rejection.
PIL vsuint8 four-process ABBA120 updates all parity pass but no consistent gain:
one-object medians PIL25.620/24.704/79.312 versusuint8 29.563/24.784/80.765ms
for three cases. Source multiobject/paired_1_2_uint8_v1_aggregate/summary.json;
causal_speedup_admitted=false and goal_50hz_under_20ms_demonstrated=false.

R5. Twenty translated ROI pilot mean383.714367ms,p95 386.36472ms on3 updates.
Exact RGB cache MISS408.6445ms vs4 repeated identical-input HIT mean189.4901ms;
48,384,000 hashed raw bytes per update costs22.3652-23.0006ms.
Sources multiobject/pilot_20_chunk1_querygraphs_v1/summary.json and
multiobject/pilot_20_exact_cache_v1/summary.json under archive.
These are capacity diagnostics with translated ROI, not twenty annotated tracks.

R6. Scheduler fake control3000 arrivals/60s: Event.wait p95 15.225025ms versus
sleep0.610805ms to fake RAW, all controls covered. Source
compute_council/paced_scheduler_ab_v1/summary.json. No Torch/GPU/TTC in this
experiment; never mix these times with real service timing or imply causal
attribution of the entire real-service gain.

Previous publication:52pages,24figures,27tables, evidence snapshotOct28. The
revision should be a new concise scientific article, not another chronological
operational diary. Preserve dated benchmark interpretation and refresh obsolete
434-query label status only against newer exact correspondence evidence.
