Experiment 02 — Learning the Landscape: fixed protocol

This experiment isolates place representation and retrieval from the frozen V4 local pose solver. V4 source files and its published 2,300-row report remain unchanged. The new representation experiment is not a replacement localization accuracy leaderboard.

Questions set before measuring results

1. Does spatial graph correspondence discriminate places that share object classes?
2. Do verified observations form stable geometric structure and diverse appearance variants?
3. Does training improve held-out observation retrieval after 1, 10 and 100 route passes?
4. Can entropy fall while the inferred place becomes wrong?

Data and controls

Seed 20260919. One immutable normal latent world, independently rendered historical and live acquisitions. A fixed 32 m reference lattice (532 candidates) defines the state space. Truth and nearest reference cell are used only after inference by the evaluator.

Twenty training timestamps (6, 12, …, 120 s) repeat for 100 passes. Ten explicitly recorded acquisition settings cycle: baseline, vegetation change, changed sun, seasonal plus sun, moderate cloud, stronger cloud, positive yaw, negative yaw, larger footprint, and a mixed condition. Thus the first ten passes expose 200 distinct observations, and later passes deliberately re-encounter those same observations. Exact image/nominal-pose/estimated-pose duplicates must not increase independent support or certainty. This is a saturation test, not 100 independent datasets.

Evaluation uses eight separate timestamps (3, 19, 33, 49, 63, 79, 93, 109 s) under six acquisition conditions. No evaluation observation enters memory. The actual timestamp arrays and pixel SHA-256 hashes are exported. Observations are unseen captures on the same world/route, not an unseen-world generalization test. Heading and GSD vary; the current nadir sensor does not implement a true oblique camera.

Memory admission uses the frozen V4 Hybrid solver without memory, an acceptance-score gate of 0.55, independent geometric verification and a two-sigma navigation innovation gate. Verification of repeated exact input is cached; no new independent evidence is claimed. The truth-only evaluator audits wrong admissions. A gate is not a mathematical guarantee against every confident false match.

New retrieval alternatives use identical observations, priors and a fixed candidate space: appearance, local features, semantic counts, graph relations, their combination and learned place representation. Frozen memory is evaluated at passes 0, 1, 10 and 100. Disabling the prior and disabling memory are explicit controls. The measured outcomes, including deterioration or no gain, will be retained.

Information accounting

The state space is the fixed reference lattice. We define a normalized compatibility distribution, not calibrated physical position probability. Its prior is a Gaussian navigation compatibility plus a 5% uniform floor. Observed score channels enter an explicitly recorded exponential composite model, with fixed temperature and weights. Missing graph evidence contributes a neutral factor. Correlated channels are not claimed to be independent likelihoods.

For each stage record entropy H = −Σp log₂p, effective candidates 2^H, number of candidates required for 95% mass, and entropy reduction H(prior)−H(stage). Reductions may be negative. Also report truth-cell rank, truth-neighbourhood mass, and negative log mass of the true cell. The evaluator flags concentration on a wrong place; a lower entropy alone is not success. This is model-dependent hypothesis entropy reduction, not an empirically estimated mutual information or calibrated Bayesian posterior.

Availability/false acceptance are shown across score thresholds rather than selecting the threshold that wins on evaluation data. Broad place retrieval success uses the nearest reference-cell neighbourhood (32 simulated metres); V4's 10 m final pose tolerance is a different metric.

Freeze and reproducibility

Archive protocol and source SHA-256 before execution, verify unchanged after execution, export all admission/evaluation records and candidate traces. Image extraction and verification are cached only for identical data. Timing separates reference-index construction, feature extraction, graph matching, training update and held-out query work; cache-hit counts are visible. No result-driven parameter tuning is permitted in the recorded run.

Executed inference details

The Gaussian evaluation prior uses σ=90 m and seeded x/y perturbations bounded by 90 m; training uses perturbations bounded by 7 m and σ=10.25 m. These are synthetic navigation priors, not additional image labels. The reference lattice remains fixed when learned places are projected back to it; memory cannot shrink the denominator by dropping unrepresented candidates.

place-belief.js fixes temperature 8 and appearance/features/semantic/relational combination weights 0.35/0.30/0.15/0.20. Learned compatibility is an additive term with weight 0.30, built from an estimated-place spatial kernel and measured appearance/learned-geometry agreement. With empty memory it is exactly zero. The graph trace separately adds available river, observed road/water adjacency, building layout, relations and geometric compatibility. An explicit geometric filter reports how many candidates passed; if none pass, it reports no verified candidate and retains the distribution instead of inventing a posterior.

“Unavailable” is neutral when the observation lacks that structure. A candidate missing structure that was observed is negative evidence. All visible relations remain in the correspondence denominator. Query work is shared across alternatives: exported sharedQueryMs is not per-method latency. Cached evidence time is reported separately. The high-mass false-concentration flag uses top mass ≥0.5; an additional wrong-entropy-reduction flag catches lower-mass wrong modes too.

The independent memory verifier means independent of the learned memory, not independent sensors or an independent truth source. Its accepted pose can still be wrong. The evaluator audit remains necessary.

The fixed checkpoint evaluation includes memory-channel ablations (remove learned geometry, remove learned appearance) and a combined-without-relations control. Removed channel weights are set to zero without re-tuning other weights. These controls separate extra appearance coverage from invariant geometry contribution.