Seven text encoders against a fixed DINOv2 vision side, priced by the anchor bound rather than the retired field correlation: PPMI 0.289 BGE-large 335M 0.308 MiniLM 22M 0.328 Qwen 0.5B 0.325 mpnet 110M 0.341 Qwen 1.5B 0.311 bert 110M 0.341 part-oracle ceiling 0.359 No correlation with scale and the largest model is not the best; a retrieval-tuned 335M encoder is worse than bert-base. All seven sit in a 0.29-0.34 band against a part-oracle ceiling of 0.359, so swapping in a stronger off-the-shelf encoder is closed as a route. That the band is narrow across unrelated architectures and training sets is itself the finding: what limits them is common to all of them. They are generic text encoders whose similarity geometry is organised around distinctions that do not line up with DINOv2's. Changing which one is used cannot fix that; changing how both towers are trained might, and that variable has never been touched. Also records the VG part-level oracle: annotator-supplied box-to-phrase correspondence buys 0.023 (0.336 -> 0.359), so part correspondence was never the problem either. Co-Authored-By: Claude <noreply@anthropic.com>
Each VG region box is index-aligned with its own description, so the crop can be encoded and paired with the phrase describing it -- part correspondence by annotation, which is cross-modal supervision a deployed system would not have. The cross-modal anchor bound goes 0.336 -> 0.359. So the two encoders do not covary at the part level either, and no segmentation, aggregation, kernel or backbone repairs that. The 0.805 within-text ceiling was never the cross-modal ceiling; for this encoder pair on this corpus the cross-modal ceiling is about 0.36 against the 0.9 recovery needs. Six representational interventions this session were chosen without a ceiling in view -- the same error being made on the matching side at the same time. What replaces them is a three-number screening protocol that runs before any pipeline is built. Co-Authored-By: Claude <noreply@anthropic.com>
Pool scaling does not saturate -- 0.167, 0.508, 0.644 at pools 40, 160, 640 -- but it is buying its way around a missing unary term, which is the most expensive way to not fix something. Co-Authored-By: Claude <noreply@anthropic.com>
Adversarial review of the artifacts found three of my numbers to be artifacts of my own code. All three reproduced here before acceptance: - anchor_bound presented probe rows in the same index order on both sides, so exact twins had their tie broken onto the diagonal. 0.997 -> 0.920 on omit-size. Fixed by scrambling the T-side presentation. - The truth is not a strict local minimum: 51 transpositions have exactly zero energy delta. fast_pair_descent only looked stationary because its break test treats zero as no-improvement. - scipy's FAQ takes no n_init, so it was swallowed into unknown_options and 'FAQ x30 restarts' computed bit-identically to plain FAQ. Replaced with a real restart loop over P0='randomized'. The blind ceiling for omit-size is 0.836, not 1.0: the text field has 51 exact transposition automorphisms, so T[s,s] is bitwise identical to T and no objective f(V, P T P^T) can separate an orbit at any order. Every synthetic accuracy was being divided by the wrong denominator. The fifteen failed solvers share one property -- all purely quadratic or purely spectral, none with a node-level term. Eight moments of each node's own field row, blended with the quadratic term through Frank-Wolfe, reach 0.837 with an energy gap of exactly zero. The term must stay in the loop: as a seed for pure-quadratic descent it scores 0.21, pinned through the iterations it scores 0.84 -- which is also why amplification plateaued, being itself pure-quadratic. Co-Authored-By: Claude <noreply@anthropic.com>
Adds MATCHING_RESULTS.md. Entropic GW annealed and PATH both reach 0.961 on the reference field, above the 0.958 the project's own pipeline reached after months, and neither had been run. GW reaches 0.881 on the rank-8 field where GRAMPA reaches 0.076 -- which retires the morning's rank-ladder conclusion, since that ladder was run entirely with GRAMPA and GRAMPA degrades on clustered eigenvalues. The hard instance yields to amplification rather than a better solver: descent multiplies a partial answer by about four, so the job is to feed it a start that is 10-20% correct rather than to replace it. Co-Authored-By: Claude <noreply@anthropic.com>
The caption-omitted field returned under 5% from every solver tried: Umeyama, GRAMPA, five Gromov-Wasserstein variants, FAQ with and without restarts, PATH convex-concave, a moment ladder, semirelaxed GW. The measurements said the answer was not a sixteenth solver. Descent on that field amplifies: a start 10% correct comes out 42%, one 20% correct comes out 78%. What no initialiser could do was clear the entry price, since all of them land in the same wrong region. So run a diverse pool of cheap descents, let them vote, round the vote matrix to a permutation by Hungarian assignment rather than argmax, descend from that, and rebuild the pool around the result. Each round feeds the amplifier a better start. 0.044 -> 0.72 on the best run, 0.54 mean over two. Competitive elsewhere: 0.977 on the easy field against 0.961 for the best GW variant. Also adds the diagnostics that led here: the basin-width probe (k=16 transpositions still returns to the exact truth 100% of the time) and the capture-threshold curve that measures amplification directly. Co-Authored-By: Claude <noreply@anthropic.com>
The gate settles which failure mode each field is in, and refutes the symmetry hypothesis I proposed. On the caption-omitted field the truth is a STRICT local minimum -- descent started at the truth does not move at all -- the anchor bound says the information is 99.7% intact, and our solver stops 0.44 above it at 4.9% accuracy. That is a pure optimiser failure. Natural data is the opposite: descent from the truth falls a further 0.167, so the truth is not even locally optimal, which is the information-deficit signature the 0.291 bound predicted. Steepest descent was brute-forcing all 32,640 candidate permutations through the full energy every step, including a batched cube trace with the triangle term active -- 203 seconds per descent, which is why the gates were hopeless. The pairwise term needs one matrix product for the whole table: swapping p,q changes the alignment sum by 2(C_pq + C_qp - C_pp - C_qq + 2 A_pq B_pq) with C = A @ B. Verified against brute force to 1e-9 before use, and the fast descent reaches the same optimum. 203s -> 0.79s. Adds a matching benchmark with known-reachable answers and the solver families never tried on these fields: Gromov-Wasserstein, entropic GW with an annealed regulariser, BAPG. Co-Authored-By: Claude <noreply@anthropic.com>
Three statistics failed the same way -- fields agreeing on the number and disagreeing on recovery -- because each was invented by staring at the fields rather than by asking what matching needs. The fourth asks directly: declare half the scenes anchors, hand over their correspondence, describe the rest by their field rows against the anchors, and match one-to-one by Hungarian assignment. Seconds to compute, and it upper-bounds blind recovery because blind recovery must also discover the anchor correspondence. Never violated across six fields spanning the full range of outcomes, and it separates every case the refuted statistics collapsed: synth full bound 0.989 blind 0.958 gap +0.03 synth noise 0.35 bound 0.984 blind 0.947 gap +0.04 synth omit size bound 0.997 blind 0.056 gap +0.94 synth rank 8 bound 0.930 blind 0.129 gap +0.80 Visual Genome bound 0.291 blind 0.000 gap +0.29 This corrects two claims from earlier today. Caption suppression does not destroy information -- its bound is 0.997 -- it destroys blind searchability, by making scenes interchangeable under permutation in a way given anchors break. And Visual Genome's problem was never spectral width: with the correspondence handed over, seven scenes in ten still cannot be identified. Co-Authored-By: Claude <noreply@anthropic.com>
The replacement gate proposed this morning is refuted by a control run this afternoon. Independent noise lowers the correlation to 0.828 without narrowing the underlying signal and reaches the same measured width as a field whose captions omit one factor -- 15.7 against 15.0 -- with recovery at 94.7% and 5.6%. A second instrument built specifically to fix that, counting canonical directions that generalise to held-out scenes, fails the same way and rates Visual Genome highest of the three failing fields. Established: the correlation does not govern recovery, in both directions. Corpus overlap does control it -- the caption-suppression ladder is monotone from 96.4% to 0.0%. Not established: any statistic that predicts recovery cheaply. Documents corrected accordingly rather than quietly rephrased; the refuted claim stood for four hours and is recorded as such. Co-Authored-By: Claude <noreply@anthropic.com>
DINOv2-large, twice the depth of base and wider, leaves the shared count at exactly 15 and moves the correlation by less than the segmentation noise floor. Four vision-side interventions now raise the vision field's own rank and leave the intersection alone. The cheap tier of the register is exhausted. What remains is a corpus chosen for naturally wide overlap, promoted to the central bet, and the unbalanced formulation, promoted to the critical path since such a corpus is unlikely to arrive in bijection. Co-Authored-By: Claude <noreply@anthropic.com>