SCIENTIFIC COUNCIL: EVTTC, CAUSAL FUSION, CADENCE, AND JEPA
Review cutoff: 2026-11-02; branch research/evttc-transfer-v2 (now merged in main).

COVERAGE AND HONEST SCOPE
Inspected the complete dated commit-title inventory: 127 branch-reachable commits
since 2026-10-28 00:00 +01:00. Inventoried all changed Markdown paths. Read the
principal transfer, submission, JEPA, native-support/context, canonical-pair,
canonical-population, Nov1 progress, physics review chapters 05-09, and all ten
followup-v2 mechanism-result documents, plus campaign/continuation/summary reports.
Some large bundled reads were output-truncated; do not claim that this lane alone
read every line of every changed document or every diff. Runtime and detection
are covered by other council lanes. Inspected actual ExpertRouter and
CanonicalPair code; cross-checked primary numerical claims directly against JSON.

EDITORIAL DECISION
The central new story should be physical coordinates and measurement contracts
in causal Event-RGB TTC, with compact signed inverse-rate fusion as a useful
adaptation, and with cadence/representation diagnostics showing why improvements
of intermediate objectives do not suffice to promote an estimator. Avoid a paper
framed as a newly successful JEPA model. These five days produced a scientifically
valuable negative result and targeted falsification evidence, not a validated
new world model. Describe novelty as this system and these controlled findings;
do not claim no prior work uses physical TTC, quality fusion, or temporal scaling.

1. CAUSAL ADAPTATION -- INCLUDE MAIN TEXT
Original F0: nine TRAIN40 components frozen: E_A5, E_C2F, PAIR_E, R_A5, R_C2F,
PAIR_R, E_CTX_MATCHED, R_CTX, PHASE_ONLY_EVENT_S7_TRAIN40. Three new QA routers,
seeds 7/13/21, 688 parameters each (33->18->4 tanh gate), four experts: F0,
RGB context, event context, causal box geometry. Softmax masks unavailable
experts. The gate receives magnitude/sign agreements, availability and 20
quality quantities; it does not receive the sequence ID or target errors.
Fusion uses signed inverse TTC. Internal expert magnitude limits [0.1,60] s;
zero rate cancellation returns F0. No adequate history returns F0 exactly;
stale box endpoint >150 ms invokes the history fallback. Geometry uses at most
650 ms, at least two boxes separated >=10 ms. These are annotation-box inputs;
do not imply detection was solved by the TTC comparison.

Final Adapt fit: 2608 queries, 2531 from 24 road recordings and 77 Slider500,
300 AdamW updates per seed, lr=.01, wd=.01. Eight competition recordings and
Slider750/1000 excluded from THIS FINAL FIT. Historical tuning/exposure remains;
earlier startup development explicitly used Slider750/1000. Therefore final-fit
exclusion does not establish clean generalization.

Corrected local comparison (query-weighted, same organizer-described processing
of negative outputs in both columns; ZIP itself unchanged):
Population n; original MAE -> Adapt MAE; original RAE -> Adapt RAE
all434: .8249759271589365 -> .33589880629301444; 49.89424239376255 -> 17.249803112310904
road365: .30540136387734423 -> .18520004117438355; 11.917131743756457 -> 7.469418642264264
without750 n383: .3296427520836695 -> .19326319205961828; 18.730071105277833 -> 9.4171567321828
Slider750 n51: 4.544831 -> 1.407064; 283.9310581484616 -> 76.071442

SOURCE: docs/research_progress_20261101/v1/evidence/evaluator_audit.json
KEY: frozen_baseline_metrics.<population>.{original_email_processed,frozen_submission_email_processed}
fields mae_seconds, relative_error_pct, population, scored.
population keys all434_public_diagnostic, road365, pdf_reproduced_sequences383, Slider-750.
IMPORTANT: docs/evttc_submission_20261031/REPORT.md and early research logs show
obsolete 423/434 target results. They MUST NOT be the new paper's current table.
The fixed evaluator uses official download mappings to correct swapped low/medium
target files without fitting clock offsets; 434 targets now supported. Nine
sequence RAE values reproduce original organizer PDF rounding; Slider750 still
differs (local 283.931058% vs official132%). Therefore local all434 is NOT an
official benchmark score. Retain raw-vs-processed distinction and one remaining
negative Adapt output. New adapted official evaluation is not confirmed.

2. CADENCE -- INCLUDE MAIN TEXT, WITH NONPROMOTION
CanonicalPair freezes weights, replaces dt,logdt,1/dt at PAIR input with d0=.1 s,
obtains raw phase phi_c, converts q_c=-expm1(-phi_c)/d0, uses
q_actual=q_c*d0/dt_actual, then phi_actual=-log1p(-d0*q_actual). Apply the original
emission/clipping policy once afterwards. At dt==FP32(.1), return original bytes.
Reject invalid real-domain transformation (1-d0*q<=0 or overflow).
Recompute raw17 disagreements, temporal contexts and final head, not just output.

The two-query timewarp diagnostic changes prepared observations and their clocks;
it does NOT reselect fixed-duration windows from a time-scaled raw stream. Under
speed factor a, ideal ratio a*T_transformed/T_control=1. F0 improved all four
contrasts after correction, but road event context at a=.5 worsened .673790->1.891432.
This counterexample is essential: intermediate consistency does not guarantee
downstream consistency.

Full exposed population: macro-recording metrics, paired original recapture.
Road2531/2531: MAE .810620018160979 -> .8104896119362653;
RAE30.600389999028852 ->30.591045605308967.
Slider50076/77: MAE1.17815589826358 ->1.1208089503802752;
RAE83.885893959795 ->82.44204743312503, relative MAE decrease4.8675%.
Road delta MAE=-.00013040622471352978 s;
95% descriptive recording bootstrap [-.0005047002743243974,.00009314967648063326].
10000 resamples, seed7, 24 recordings. No new missing values or sign flips.
Only one Slider recording, reused from screen: NOT independent replication.
SOURCE docs/research_log_canonical_population2608/evidence/report.json
KEY routes.F0.domains.{road,slider}.{original,canonical}; road_bootstrap.

Frozen adapted router failed the prespecified 149-query road screen despite
aggregate population improvement. Matched canonical refit: 48 fits, 8 folds,
3 seeds, 2 arms. Full population road .539608->.537801 s, Slider .364987->.356515;
fixed screen road regressed .484752->.511821 s and RAE16.720860->17.431738.
No seed passes both views; no export/promotion. Source
docs/research_progress_20261101/v1/evidence/refit_result.json:
population2608.domains, screen149.domains, admit_further_validation=false.

Local eAP sensitivity covers5805 physical queries (6762 aliases), not accuracy.
F0 p50/p95 absolute change .0000143051/.0000767708 s; max .0064964291 s;
no new missing/sign flips. Original local replay already differs historically:
5805 fail some intermediate check, 504 fail TTC tolerance. Cannot claim official
score preservation. Source native_result.json and router_result.json in same v1/evidence.

3. JEPA -- INCLUDE NEGATIVE RESULT, NOT OPTIONAL FOOTNOTE
Fixed V-JEPA2.1 ViT-B distilled teacher, event student, no TTC labels in student
training; 1000 updates*3arms*3seeds, batch16. Frozen encoder ridge alpha1 on P512
from24 sequences, H256 from8 other sequences; localization uses privileged boxes.
Probe target is DIMENSIONLESS ANNOTATED LOG-HEIGHT CHANGE, not TTC or q.
Means B_static .03763238961408746, C_one_step .03564004206250102,
C_shuffle .035061885577055175, random .037196408997659317.
C-B delta -.0019923475515864395, CI[-.0049620951828768045,.0008189780757168292].
C-shuffle delta .0005781564854458428, CI[-.002170051001389082,.002925896583929867].
Neither upper confidence bound below zero; continuation criterion failed.
Future teacher-state MSE C-B improves -.01719163358 and C-shuffle -.05317387730;
teacher increment MSE worsens +.03848015858 and +.00447422173. C also beats own
student persistence for future state, -.03530466308. Thus improving future state
alone does not establish useful dynamic representation. Mechanism not identified.
SOURCE docs/jepa_closure_20261031/closure_analysis.json
KEYS primary_mae, primary_comparisons, latent_comparisons.
CI resamples 8 sequences keeping 3 seeds together; conditional on those seeds.
H historically exposed; no independent confirmation. Pilot closed, rollout not run.

4. MECHANISM CONTROLS -- SHORT MAIN DISCUSSION / APPENDIX
Observability: q=c^T beta identifiable iff c lies in row space of H. Algebraic
estimability and low residual do not protect against model bias. Here q estimable
63/149; 26 wrong signs, 10 outside canonical domain,8 zero residual degrees freedom.
Physical-feature probe for correction gain MSE .919267453 vs support .492043720;
delta+.427223732 CI[.08055069039,1.01041723207]. Not TTC error.
Robust reading: only8 outputs,4 wrong signs; fallback-system MAE .779543461 vs
R0 .766024827. No new observability gate admitted.
SOURCE continued_results_v1/OBSERVABILITY.json and
followup_results_v2/{observability_failure,robust_measurement}/METRICS.json under
docs/release_research_implementation_20261031.

Photometric ambiguity: 24 synthetic paired streams coincide at pixel/polarity/
timestamp tolerance1ps;12 have distinct q and inverse texture-gradient changes.
Known reference can disambiguate within restricted bank. This is not a universal
impossibility theorem. Photometric proxy72/149 road queries,24 acquisitions;
physical sensor residual worse than activity. Adding observed RGB change helps
sensor residual vs physical, but q readouts yield fallback MAE9.1685 s (affine)
vs R0 .7660 s. Sensor reconstruction is not TTC validation.
SOURCE followup_results_v2/{sensor_identifiability,sensor_reference_control,photometric_q}/METRICS.json.

Surface three-frame estimate:133/149 measurements, fallback16, MAE3.103982972 vs
R0 .766024827; delta CI [.93423130897,4.19824494379]. No varying crop zoom found
in146 auditable histories;53 constant anisotropic resizes. Does not rule out
occlusion/background/parallax. SOURCE continued_results_v1/SURFACES.json.

Event-expansion solver: original endpoint warp p_ref=p+(q*p-b)(t_ref-t), free
focus, NCC early/late images; synthetic confirmation12/12. At50% static background,
both held-out fixtures accepted wrong q. Real Slider50062/77 available, RAE15.693702,
MAE .268963; geometry on same62 MAE .207334; fallback-to-Adapt all77 MAE .256119
vs Adapt .205044. Thus not promoted. Mixed-precision float32 failed objective parity;
hybrid float64 GPU/CPU retains failures but is separate experimental solver.
SOURCE docs/research_log_20261031_event_expansion/README.md plus its evidence JSON.

5. PROTOCOL AND NOTATION
Do not conflate historical official GarlTTC OverallMiD35.80145 (TRAIN40 frozen)
with new adapted local RAE17.2498%. Different metrics, data, and training regimes.
Distinguish original phase coordinate phi=-log(1-d0/T), physical q=1/T,
and finite emitted TTC. Constant-depth-velocity dynamics q'=q/(1-delta*q)
are classical algebra under axial assumptions, not new theorem. Finite magnitude
cap60s cannot encode q=0 physics. Camera rotation/point changes violate simplified
depth assumptions. No uncertainty calibration inherited by all-data refit.

FILES AUTHORED BY THIS LANE
paper/revision_20261102/build_evidence.py
paper/revision_20261102/evidence_tables.tex
paper/revision_20261102/evidence_values.json
paper/revision_20261102/source_manifest.json
paper/revision_20261102/evidence/{official,adapt,canonical,jepa}.json
Script reads only JSON, copies exact source bytes and records SHA256; verifies
official weighted score and corrected local434 audit. Generated four table
commands and numeric macros; successfully executed with python -S. No training.
