Q1 learns, but its spatial readout has a coverage blind spot

Do not treat the latest run as a data-scale-only problem. The final model fits TRAIN better than a TRAIN-fitted constant, and its TRAIN MSE improves after the selected checkpoint while development MSE worsens. Separately, a frozen-population geometry census finds states where acquired evidence cannot reach the spatial readout: at the last state, all valid candidates have zero in-map target or segment queries in 36/145 TRAIN chains and 3/20 development chains. This is a concrete representation limitation worth addressing before another broad fit. Its contribution to aggregate error and the benefit of a replacement remain untested.

This exploratory, fixed-weight analysis concerns one completed CUDA C1 seed (41), 2,030 successful updates, selected update 1,305, 77 TRAIN scenes and 20 exposed development scenes. It uses the admitted privileged CF-GT selected-depth replay; it is not an end-to-end actor-visible sensor or policy benchmark. It performs no new training or acquisitions. Interactive figures and all tables · portable evidence bundle · source and analysis provenance · experiment index.

What changes the next-run decision

Open full-size figure · Interactive view

The census covers root and last state of every frozen chain: 290 TRAIN observations and 40 development observations. These are query-domain fractions, not measured surface coverage or estimated learning effects. Scene means weight physical scenes equally; the 36/145 and 3/20 counts above use chains. Intermediate states were not censused.

The 27 target-lattice points are all outside the fixed chart in 44/145 TRAIN chains and 6/20 development chains. At root, segment queries still give every valid candidate some in-map access; later camera positions can lose that access. At the last state, the scene-equal fraction of valid candidates with no query is 27.8% on TRAIN and 20.5% on development. Raw candidate-weighted fractions differ (1,437/3,589 and 172/477); none is an independent-scene failure probability. Exact census and denominators.

The source-qualified three-scene probe distinguishes failed fusion from inaccessible readout. In scene 81022, last-state fusion contains 3,925 surface voxels and 24,620 free voxels after 15 selected observations, but target and segment query support are both zero. Withholding depth returns leaves its outputs bitwise unchanged. Two other fixed TRAIN scenes retain query access; withholding changes predictions slightly (maximum absolute changes 0.001101 and 0.000475), without changing either chosen action. This three-scene, out-of-distribution missing-return intervention demonstrates frozen-model input sensitivity only. It does not estimate feature importance, population benefit or retrained performance. Acquisitions, poses, prefixes, labels and support remain factual; no acquisition was omitted. Probe outputs · volume evidence · candidate support.

Open full-size figure · Interactive view

Each candidate error is averaged within a state, states within a physical scene, then scenes equally. All six evaluations use the same guarded GPU evaluator and original labels. The initial mixed-digest batch is superseded. The constant is fitted only on TRAIN; its deterministic tie action is different from uniform choice.

Regret is the best supported original H1 label minus the label of the greedily chosen candidate; uniform choice averages labels over supported candidates. These are offline recorded-state diagnostics, not closed-loop reconstruction outcomes.

The model is learning something beyond a scalar baseline. Final-minus-selected TRAIN MSE is −0.000616, paired-scene 95% bootstrap interval [−0.001123, −0.000178]; development MSE changes by +0.000196 [0.000082, 0.000331]. This supports a late fitting/generalization or objective tradeoff, conditional on these fixed weights. It does not identify a learning rate, sufficient data scale or causal overfitting mechanism. TRAIN and development have different difficulty distributions, so their absolute MSE levels are not a matched generalization-gap estimate.

MAE and decisions tell different stories. TRAIN MAE worsens late by 0.000872 [0.000075, 0.001653], while development final-minus-selected regret improves by a point estimate −0.001747 [−0.004375, 0.000127]. The latter interval includes zero. Improvement against the deterministic constant must not be reported as improvement against uniform choice; the final-versus-uniform development regret contrast is −0.002835 [−0.006380, 0.000230], which includes zero. Inspect the paired table. These intervals resample entire physical scenes (10,000 draws, seed 20261015). They omit training-seed and checkpoint-selection uncertainty and do not make exposed development scenes untouched confirmation.

A descriptive coverage-stratum check limits the interpretation further. At final weights, last-state MSE is 0.000337 on the 36 inaccessible TRAIN chains versus 0.000185 on the other 109; development has only three inaccessible scenes (0.000294 versus 0.000150 on the other 17). These last-state errors are much smaller than root errors. At roots, zero target-query access is associated with lower TRAIN MSE but higher development MSE, alongside different label spreads. Thus access loss is real, but these observations do not establish that it dominates learning error. Scenes can occur in both TRAIN strata; these are descriptive, difficulty-confounded groups, not independent causal comparisons. Stratum table and exact support.

Which explanations survive the evidence?

Open full-size figure · Interactive view

MSE is verified as state-offset squared plus centered residual variance before aggregation. The right panel retains each phase's contribution to total scene-equal MSE instead of renormalizing root and later phases separately.

Training presents one complete chain per update, so scenes with more chains receive more presentations; the scientific scorecard instead weights scenes equally. Final TRAIN state-weighted MSE is 0.002507 versus scene-equal MSE 0.002579. This exposes an estimand difference, not a tested advantage for a different sampler. Roots are only one of sixteen recorded states, but their larger residuals can carry substantial quadratic loss.

Root errors are large but do not all move together. TRAIN root MSE improves by 0.010373 after selection; later TRAIN MAE worsens by 0.002158 [0.001479, 0.002860]. Final development root MSE is 0.008476 versus later-state MSE 0.000722. This phase split motivates separate root/later monitoring and a controlled exposure comparison. Root-error stratification remains evaluation-only, separate from actor inputs. Phase table.

Open full-size figure · Interactive view

All 320 development states are shown at final weights; roots and later states have different label ranges. Candidate rows and states are dependent. Lower within-state prediction spread does not itself establish a defective architecture or an optimal ranking loss.

Optimization and missing metrics

Open full-size figure · Interactive view

Fourteen native validation checks retain their original successful-update coordinates. The gradient view contains 203 distinct captured task-diagnostic updates, not 1,939 independent transport observations. A plotted line connects sparse captures; it does not reconstruct intervening updates.

Executed configuration is AdamW, LR 0.0003, weight decay 0.001, batch one complete chain, direct H1 Huber(delta1), zero auxiliary weights and no active scheduler. Training used 16-mixed precision, clipping norm 1 and accumulation 1; the new fixed-weight comparisons use FP32 with highest matmul precision. The frozen model uses attention width 32 with four heads, decoder hidden width 128, dropout 0.05 and mean-pooled pose history. The fit completed 14 epochs. The generic OneCycle fields do not mean OneCycle ran. Selection used validation loss, with patience five; final weights remain primary and selected weights secondary. There is no retrospective reselection.

Task diagnostics occur every ten successful updates. Captured global unscaled gradient norms range from 0.000688 to 0.571581; captured readout norms are nonzero. Pre/postclip values agree at these task captures. Other optimizer-callback observations exceed one, so this does not mean clipping never occurred. Historical callback fields lack their own observation coordinate and cannot safely borrow a carried task counter. The patch adds a zero-based attempted-step coordinate prospectively; it changes no gradients, clipping or updates. Missing AMP skip/scaler records are not measured zero skips. Unique diagnostic observations.

Instrumentation overhead has not been benchmarked as a production throughput result. Freeze cadence and an acceptable overhead threshold before any new fit, rather than enabling every diagnostic by default.

Three prioritized next interventions

1. Restore and audit spatial access, before broad scaling. Use the census as a preflight for the current representation and a separately identified corrected-chart/readout candidate. First verify frames, bounds and availability on the same frozen actors; then compare a target-covering chart or readout under a written protocol. Keep calibrated volume resolution, physical inputs, encoder, loss, scene split, exposure and selector controlled or declare unavoidable differences. A larger chart at fixed 48³ changes metric resolution; it is not a free correction. Confidence is high that inaccessible paths exist, low that a particular replacement improves learning. Stop the geometry candidate on invalid frames/support or failed declared coverage; numerical acceptance thresholds must be frozen before evaluating its outcomes. No new fit is launched here.
2. Separate fitting from optimization and objective emphasis. After coverage qualification, a separately authorized small-TRAIN memorization canary should establish whether a fixed population can be fitted, with repeated fixed-label TRAIN/dev probes. If that passes, choose one matched LR/schedule or loss-emphasis contrast, rather than changing both. Use the current AdamW/LR/no-scheduler configuration as control; retain maximum-chain capacity, attempt/success/finite diagnostics, root/later calibration and regret, fixed update/presentation budgets and selector opportunities. A failure to fit is informative and stops escalation to a larger campaign. Gradient magnitudes alone do not select the LR or batch.
3. Measure independent-scene scale and replication. Only after accessibility and fitting controls, vary unique TRAIN scenes separately from chains per scene. Freeze nested TRAIN subsets, untouched confirmation scenes, independent seeds and either fixed updates or fixed presentations (report both). Keep a matched control and retention criteria for calibration plus regret. More trajectories without new scenes is a different intervention. Confidence in a current data-scale-only explanation is low; this experiment is what would discriminate it.

All three require a concrete prospective protocol. New training, optimizer transactions (including a temporary capacity canary), rollout collection, merge and deployment retain separate authorization gates. Before a fit, execute the full capacity protocol with representative maximum admitted chains, frozen precision, controlled warmup/cold states, complete optimizer and validation transactions, peaks/headroom and failure receipts. Current inference does not clear that gate.

How earlier experiments constrain this diagnosis

The matched Mac Q1 comparison increased R1 chains from 77 to 145 on the same 77 TRAIN scenes. It did not clear its frozen retention gates: final MAE worsened, and the regret contrast remained uncertain. Its zero auxiliary losses also exclude active joint-BCE competition for that fit. CUDA changes backend/precision, exposure and selection opportunities; its different point estimates are not a causal backend or hyperparameter comparison.

The larger Q1 and interrupted QH synthesis and failed-QH diagnostics preserve separate dispositions. Continuing QH failed; long-horizon fitted residuals and weighted BCE share cannot identify a failure mechanism. The present pure-H1 controls already identify a concrete next representation test, so another QH inference campaign is not needed to rank that test. No failed run becomes a completed comparison.

Read the thesis alongside the executed implementation

Start with the VIN package and finite-horizon package, then follow the Method-to-code pairs below.

Finding → RQ → thesis section → coverage / gap

The retained feature and geometry probes predate diagnostic package attestation. Their recorded script, checkpoint and population checks remain available, but do not establish the package bytes used at execution. Future captures require guarded implementation identity; legacy readback explicitly records this limitation.