A story state that remembers without pretending it read the page.

A better state may become a useful temporal prior.

Identity continuity, repeated-scene disambiguation, belief tracking, causal closure, and loop branches are where persistent state could add information that a local page reader lacks. The current serialization does not yet demonstrate that value.

• Preserve exact evidence and provenance.
• Keep top-k speaker, identity, order, and branch hypotheses.
• Revise recent decisions when later pages resolve ambiguity.
• Use query-conditioned retrieval to reopen the raw crop.

A deterministic graph cannot be the comic reader.

It cannot recover missed lettering, invisible actions, or wrong panel order. Worse, a wrong speaker or event gets promoted into a persistent “fact” and contaminates later answers.

• Do not assign one irreversible speaker or identity.
• Do not infer a reset from a keyword such as “revival.”
• Do not compress every page into one event sentence.
• Do not report teacher prose as vision-only evidence.

MagiV2 sees characters well; text labels expose a granularity mismatch.

Across all 308 pages, standard IoU@.5 character detection reaches .7727 precision, .9479 recall, and .8514 F1. Standard text detection is only .0807 precision, .0967 recall, and .0879 F1. The matched prediction-to-gold area ratio has a .263 median, consistent with tight lettering predictions being compared with larger Re:Verse bubble regions.

A separate intersection-over-smaller containment diagnostic gives text .8317 precision, .9968 recall, and .9068 F1. That .9068 is IoS, not standard IoU, AP, or a substitute detection score; it only shows that tight predicted text is usually contained by a released bubble box.

The four-page OCR pilot compared 25 outputs with 11 adapted teacher turns: mean best token-F1 per teacher turn was .8328, with all 11 turns reaching at least .5; mean best token-F1 per OCR string was .4727. This is a lexical proxy—not OCR ground truth, CER, WER, precision, or recall.

The current deterministic state loses useful signal.

Privileged local text and gold identity/layout improve candidate ranking, so narrative and structural information can matter. But the matched Story State serialization trails the frozen visual reference, and its linear head collapses to 7.1% with 10 pages.

In the held-constant fusion test, local text beats matched state by 18.4 top-1 points at five pages (descriptive 95% interval +5.3 to +31.6) and 17.9 points at ten (+3.6 to +28.6). Candidate representations and the SigLIP prefix sequence are identical; only the prefix text representation changes.

The current deterministic state has not earned promotion. Branch-aware graphs, uncertain relations, open-thread stores, and predictive latent slots remain candidate data structures—a hypothesis to test after exact retrieval and a strong task reader, not an established architecture win.

Seemingly small leakage channels can dominate NPP.

If page identifiers escape the boundary, chronology solves all 433 rows. Reused candidates also permit transductive strategies. Scientific runs therefore detach pixels, strip file and page metadata, expose only neutral A–J aliases, and require a stateless task-local scorer.

The repaired manifest excludes prefix/target gaps and keeps identical candidates for a shared target across 5- and 10-page comparisons. The public manifest is an audit artifact, not by itself a sealed benchmark.

The first zero-shot run was invalidated by answer-position bias.

The superseded three-seed validation run scored 16.7% top-1 / .367 MRR, but accuracy changed drastically with the gold letter: D reached 46.9%, H 32.4%, A 4.2%, and J 0%. Those numbers are not model evidence.

Version 2 cycles every item through A–J exactly once and shuffles negatives independently. Validation is 13.4% top-1 / .329 MRR with five pages and 10.0% / .316 with ten. The frozen test is 17.1% / .371 and 17.1% / .360, with 99.7%/100% parse rates.

That is above chance on this fixed test cohort, but 11.8 and 4.3 points below frozen SigLIP; paired descriptive intervals include zero. Preferred canonical images still change across a mean 4.2 of ten permutations. Balancing neutralizes aggregate letter preference—it does not make Qwen order-invariant or establish a gain over simple image similarity.

We now know where the next experiment must be stricter.

The current run covers all 308 pages for MagiV2 detection, a four-page OCR proxy, and 433 repaired NPP rows. The test slice contains 38 unique targets from two chapters; 28 targets support a paired 5-versus-10-page comparison.

• Measured: task-local frozen vision, train-only probes, privileged text/structure upper bounds, and exactly prefix-matched Story State diagnostics.
• Still missing: real OCR ground truth, speaker and identity scoring, VQA/IPP evaluation, a trained updater, and an untouched cross-title test.
• Next causal oracle: annotators see only pages ≤ t; identities remain unknown until revealed, and every answer has a minimal region-level proof.
• Next fairness contract: match visible pages, image/state tokens, training targets, candidate calls, FLOPs, and latency.
• Next statistics: predeclare a target/chapter/title hierarchy and stop treating overlapping prefixes as independent observations.

Calibrated manga perception

The measured 85.1% character IoU F1 and 90.7% text-containment diagnostic justify continuing MagiV2, while the 8.8% standard text IoU F1 and proxy-only OCR make the remaining calibration work explicit.

Hard-negative task readers

Frozen SigLIP already reaches 28.9% top-1 at five pages. A candidate-conditioned verifier should now test whether exact adjacency can be separated from visual similarity and generic narrative plausibility.

Hybrid memory

The current matched state scores below frozen vision. Exact retrieval and predictive latent experiments come before graph complexity; explicit structure returns only after a causal oracle gap and a matched-ledger win.