English · العربية · Español · Français · 日本語 · 한국어 · Tiếng Việt · 中文 (简体) · 中文（繁體） · Deutsch · Русский

Imagized Language Model (ILM)

Concept target, not a measured model output: writing pixels enter an independent ILM-V runtime and the intended answer is a rendered page image. The glyph panels use local hanziyuan-derived ziyuan data for 言 (YAN, U+8A00). The measured V42--V46 canonical Chinese results and their narrower claim boundary are reported directly below.

Current Evidence: V46 Preserves Raster Language but Does Not Qualify

Scaled Retinal Glyph Language V46 is the preregistered from-scratch test authorized by V45. It keeps V42's exact 24,346,497 learned parameters and causal raster architecture, but replaces its normalized-DCT substrate with the full scaled V45 field v = A(d - mu) / 19.622622.... The deployed model still receives only ordered 64 x 1 x 32 x 32 image streams, emits a continuous 1,024-dimensional image field, exactly inverts it to pixels, and rereads its generated pixels. It has no token or Unicode IDs, strings, OCR, visual codebook, glyph lookup, external runtime LM, or deployed candidate bank.

On the fixed 2,048-window development audit, full top-1 reaches 20.752%, above image unigram (1.416%), symbolic bigram (12.256%), and shuffled history (19.238%). Full target log probability is -4.9355, a 0.3198 nat improvement over V42 that passes the frozen V42-gain gate. But top-1 improves over V42 by only 0.781 percentage point, below the required one point, and exact-suffix counterfactual arm accuracy is only 54.297% against the >60% gate.

Bank-free generated identity top-1 is 8.594%, only 0.391 point above V42 and below the required one-point gain. Generated pixel F1 falls from V42's 0.37308 to 0.35943, against a fixed >0.55 gate. The real held-out sheet shows the same failure: some outputs retain recognizable structure, while many fragment strokes or move toward another identity. V46 passes 10/14 gates and is non-qualifying. Train plus audit takes 955.78 seconds with 0.64668 GiB peak allocated CUDA memory on one RTX 4090 D. The V43 writer and frozen partition remain closed.

This narrows the next question. V45's conditioning is useful and V46 can learn language probability in it, but one isotropic full-field energy objective does not couple identity direction, ink radius, and clean raster rendering. The next bounded test must factor those quantities explicitly before any larger model, writer composition, or frozen-data opening.

Evidence Path: V42 Language, V43--V44 Binding, V45 Geometry

Canonical Glyph Language V42 is the first bounded positive natural-language result in this repository. Its 24,346,497-parameter causal model receives only ordered 64 x 1 x 32 x 32 glyph rasters and predicts a continuous next-image field. On 2,048 fixed development windows, full-history top-1 is 19.9707%, above image unigram (1.4160%), symbolic bigram (12.2559%), and shuffled earlier history (18.3594%). Ordered target log probability improves over shuffled history by 0.17391 nat. The run completes 10,000 updates in 22.91 minutes with 0.632 GiB peak allocated CUDA memory on one RTX 4090 D.

This establishes a narrow but real result: ordered writing pixels carry usable next-glyph language information without a tokenizer, token or Unicode IDs, OCR, a visual codebook, glyph lookup, an external runtime LM, or a deployed candidate bank. V42 is not a complete ILM. Exact-suffix counterfactual arm accuracy is only 53.0273%, and generated pixel F1 is only 0.37308.

V43 tests those two failures. It fine-tunes the reader with train-only same-suffix raster pairs, freezes it, and trains a 5,693,697-parameter spatial rectified-flow writer. The full 30,040,194-parameter model preserves all four language-control wins and raises autonomous generated pixel F1 to 0.44507. Its outputs are visibly more coherent and remain bank-free during generation, but exact-suffix arm accuracy reaches only 54.4922%. It therefore fails the unchanged > 0.60 binding and > 0.55 pixel-F1 gates. V43 is partial, and the frozen partition remains unopened.

A post-result diagnostic explains the V43 failure. V43 scores 99.22% on sampled training pairs but only 58.30% on unseen train pairs and 54.69% on development pairs: the 5,000-pair pool was memorized. By contrast, the spatial writer reaches 0.8824 pixel F1 when the evaluator supplies an exact target ink plan, versus 0.4611 from the autonomous predicted plan. The writer can render; the continuous visual language state is the dominant bottleneck.

V44 performs the resulting frozen-base test. A 1,735,936-parameter residual reads earlier raster memory while preserving V42 exactly, consumes one pass over 24,000 unique train pairs, and leaves 1,024 train-partition pairs untouched. Its 3,000 updates take 210.87 seconds and 0.1985 GiB peak allocated CUDA memory. Consumed and unseen-train arm accuracies are nearly identical (57.47% versus 57.42%), so the V43 repeated-pool memorization gap is removed.

Binding still does not generalize. V44 reaches only 53.42% on fixed development pairs, versus 52.44% for matched V42 and 51.86% after shuffling the earlier prefix. It also damages the accepted natural calibration: matched full-history top-1 falls from 19.43% to 17.43%, and target log probability falls from -5.238 to -5.777. V44 passes 8/14 preregistered gates and is rejected-or-partial. The writer and frozen partition remain closed.

A fixed post-result scale sweep finds no residual strength that reaches the 60% binding gate. More importantly, the learned update-difference has negative cosine with the true target-image difference (-0.0229 on development), while anchor cosine to the corpus-mean image field rises from 0.777 to 0.882. Raw target cosine improves even as rank and probability worsen. The next bounded test must therefore center and variance-balance the continuous raster field, then qualify that representation on held-out image geometry before training another reader or reopening the writer.

V45 performs that preregistered representation test without training a reader or writer. It fits a zero-parameter, invertible matrix-power field from 8,000 training-only canonical rasters and emits continuous direction plus log radius. Across 14,144 fit, held-font, and one-pixel-shift rasters, maximum FP64 DCT reconstruction error is 4.26e-14, binary pixel accuracy and ink F1 are both 1.0, and no output is blank.

The target geometry improves on every fixed measure. Weighted common resultant falls from 0.7190 to 0.01524, while effective rank rises from 145.08 to 185.38. On the pinned 1,024-pair V44 holdout, candidate-pair cosine falls from 0.56319 to 0.06402, fifth-percentile displacement norm more than doubles from 0.36976 to 0.77724, and displacement effective/stable ranks rise from 122.80/60.82 to 144.95/67.79. Held-font and shift retrieval gates also pass. V45 passes 13/13 gates in 66.51 seconds with 0.302 GiB peak allocated CUDA memory on one RTX 4090 D.

This qualifies an image-only representation, not a language model. Applying V45 after the already-trained V42 reader reduces matched natural top-1 from 19.43% to 15.04%; that preregistered report-only diagnostic rejects a retrofit between incompatible coordinate systems. V46 performs the required separately preregistered from-scratch test and, as reported above, preserves ordered language controls but passes only 10/14 gates. The V43 writer and frozen partition remain closed.

See the V42 protocol, V43 protocol, measured V43 result and diagnosis, V44 protocol, measured V44 result and diagnosis, V45 protocol, measured V45 result, V46 protocol, measured V46 result and diagnosis, tracked V42 evidence, tracked V43 evidence, tracked V44 evidence, tracked V45 evidence, tracked V46 evidence, and compiled paper.

Prior Evidence: V41 Qualifies a Visual Glyph Motor, Not Language

V41 tests a practical output component for the image-native loop. A qualified V34 continuous codec produces clean or perturbed canonical glyph rasters; MX-Font, an externally developed 22,761,566-parameter image-conditioned generator, receives only those source rasters and four target-style reference rasters. The model path receives no token IDs, Unicode IDs, OCR output, retrieval result, character lookup, or candidate bank. Both checkpoints, external source revision, fonts, style references, and evidence are hash-pinned.

Across ten held glyphs, the motor raises target-style ink F1 from 0.43182 to 0.54794 for clean V34 projections and from 0.43170 to 0.54610 after sigma=0.05 latent perturbation. The noisy route retains 99.66% of clean projected-motor F1; every output is finite and nonblank. The complete audit takes 1.650 seconds and 771,669,504 peak allocated CUDA bytes on one RTX 4090 D. This passes the visual motor gate.

It does not pass a language gate. The audit supplies the intended source glyph image to the motor, so MX-Font cleans and restyles content but does not choose what comes next. The preceding corrected V39.1 trajectory pilot fixed count prediction (6.303 expected versus 6.313 target), yet held-out answer MRR fell to 0.07406 and segment MRR remained 0.01234; it was not scaled. The next proof must autonomously predict a continuous next-glyph state from canonical-font raster history, decode it to pixels, and beat unigram, bigram, shuffled-history, and blank-history controls without receiving the target glyph.

See the V41 research decision, V39.1 diagnosis, tracked V41 evidence, and compiled paper.

Prior Evidence: V38 Improves Visual Reading but Fails Answer Generalization

Visual Path Alignment V38 is a 90,753,281-parameter image-only reader and prompt-conditioned answer-state model. It completed all 8,000 BF16 updates in 99.71 minutes on GPU 0 of one RTX 4090 D with 2.969 GiB peak allocated CUDA memory. Its deployed tensor path receives only a 3 x 16 x 1024 prompt raster and clean 64-patch mask, then emits continuous 1024-dimensional prompt and answer states plus visual length. It contains no strings, token or Unicode IDs, OCR, vocabulary logits, candidate bank, visual codebook, target tensor, teacher call, or network client. It does not render answer pixels.

The frozen EMA development decision is not-qualified. Prompt top-1/top-5 reaches 60.71%/86.73%, held-font prompt consistency reaches 0.793, and the answer transition is no longer near identity: prompt-answer cosine falls from V37's 0.997 to 0.577, with transition-direction cosine 0.340. But the actual prompt-conditioned answer state reaches only 21.94% top-1, 49.49% top-5, 0.3460 MRR, and 0.244 paired cosine. Held-font answer consistency (0.732), paraphrase answer consistency (0.491), counterfactual assignment (89.80% against 90%), and visual-length MAE (3.370 patches) also miss their frozen bounds. EMA passes 25/39 conditions. Raw weights are materially the same and do not repair the answer path.

V38 intentionally reuses strong external work with exact attribution. Pixel-Linguist-v0 initializes the reader through V37; BGE-M3 builds detached targets; and Qwen-family models prepare and audit training-only paraphrases. None is claimed as project work or called by the deployed student. External models that work are welcome when provenance, licensing, training role, and runtime status are explicit. "Independent" describes the final self-contained image-only runtime, not a requirement to pretrain every supporting component from scratch. Pixel-Linguist's checkpoint states no weight license, so derived weights remain local-research-only.

The bounded advance is real: V38 improves prompt reading, held-font prompt consistency, and answer-transition geometry. The failed answer generalization is equally real. The next proof should expand deduplicated instruction-relation diversity and test ordered multi-state answer dynamics before any raster writer is opened. Zero sealed rows were rendered, and the renderer remains unauthorized.

See the measured V38 result, frozen protocol, research decision, tracked evidence, and compiled paper.

Prior Evidence: V37 Reads Visually but Fails the Complete Semantic Gate

Visual Semantic Distillation V37 is an 89,768,706-parameter image-only reader and answer planner. It completed 8,000 BF16 updates in 61.59 minutes with 2.748 GiB peak allocated CUDA memory on one RTX 4090 D. Prompt-state top-1/top-5 reaches 47.45%/77.55%; candidate-independent answer plans reach 20.41%/45.41% and 0.3377 MRR. Counterfactual assignment and answer rank pass, while absolute answer alignment, font and wording invariance, and length fail. EMA passes 20/33 checks and is not-qualified. Its sealed split and renderer remain closed. V38 directly tests the diagnosed invariance and near-identity transition failures reported by V37.

See the measured V37 result and frozen protocol.

Prior Evidence: V36 Learns a Weak Relation but Fails the Semantic Gate

Visual Semantic Plan V36 is a 93,473,281-parameter image-only planner. It completed all 6,000 BF16 updates in 25.63 minutes on GPU 0 of one RTX 4090 with 1.541 GiB peak allocated CUDA memory. Its deployed method receives only a 3 x 16 x 1024 prompt raster and visual patch mask and emits five continuous 768-dimensional plans plus visual length. The checkpoint contains no answer teacher, candidates, strings, token or Unicode IDs, OCR, glyph lookup, or visual codebook. It does not yet render answer pixels.

The frozen development decision is not-qualified. EMA top-1/top-5/MRR are 1.02%, 12.24%, and 0.0783, against gates of 8%, 25%, and 0.15. Raw top-1 is only 2.04%, so EMA lag is not the explanation. Counterfactual assignment passes at 78.57%, shuffled prompts degrade retrieval, and paraphrase top-5 reaches 33.33%; however, absolute retrieval, cyclic margin, blank control, held-font transfer, paraphrase consistency, and visual length all fail. Only 13/23 conjunctive checks pass. The sealed split remains unopened and the V36-R renderer remains closed.

A post-result nonsealed audit found a real data defect: augmentation was applied before occupancy masks were measured, causing shifted white background to become active. Mean train answer length is therefore 37.31 patches versus 11.53 in development, and the 768-dimensional train answer targets have effective rank only 4.77. This is contributory but not the sole cause: the frozen visual foundation itself reaches only 4.59% direct top-1. By contrast, the exact local BGE-M3 teacher reaches 83.16% direct prompt-to-answer top-1, while a closed-form map from frozen visual features reaches only 2.04%. V37 implements those clean masks and end-to-end offline semantic distillation. It substantially improves visual reading and answer planning, but its complete font, wording, margin, and length gate still fails as reported above.

See the measured V36 result, frozen protocol, pre-run research note, and compiled paper.

Prior Evidence: V35 Closes the Raster Loop but Fails Binding

Causal Glyph Flow V35 is a 129,092,738-parameter Chinese visual-language student. It completed all 22,000 BF16 updates in 2.770 hours on one RTX 4090, with 2.899 GiB peak allocated CUDA memory. Its deployed path receives only writing pixels and a visual mask, predicts continuous states, decodes them to binary writing patches, and rereads those actual patches. It has no runtime tokenizer, Unicode or character IDs, OCR, retrieval table, visual codebook, or teacher-model call.

The measured decision is not-qualified. Correct-prompt copy accuracy is 0.3125%; instruction accuracy is 0.1116% and is lower than the shuffled prompt condition. The visual-causal and semantic-raster routes pass only 5/12 and 5/9 gates. Outputs are nonblank and respond to interventions, but contain corrupted marks that are not bound to the requested answer. The sealed split was therefore not opened.

V35 is transfer-based. The compact V34 continuous writing codec is project work; the causal core starts from the external PIXAR checkpoint and PIXAR is credited as an external foundation, not claimed as an ILM contribution. "Independent" here means a self-contained pixel-only deployment artifact, not training every component from scratch. PIXAR's source revision is MIT, but its downloaded weight archive states no weight license, so the local diagnostic artifact is not authorized for redistribution.

See the measured result, frozen protocol, implementation guide, and compiled paper.

Research North Star: A Visual Word-Origin Book

The concrete product target is an independent image-native model that accepts a rendered English or Chinese question, or a photographed page, and emits a readable answer as a page image. A word-origin answer should combine modern English/Chinese explanation with real provenance-linked oracle, bronze, seal, clerical, traditional, simplified, manuscript, or unencoded forms. OCR may add a searchable sidecar after inference; the UI may display that text beside the native answer image, but it is not the model's language channel.

The interface still behaves like a normal prompt box. Typed text is rendered into a clean prompt band; an optional book page, inscription, or glyph image is placed beside or below it on the same visual canvas. The model can therefore answer ordinary typed questions or questions grounded in an attached page without receiving hidden text metadata.

The canonical interface is a Visual Language Stream with sequence/time, optional geometric depth, height, width, and sensory channels. A page is the T=1,D=1 case; a book is an ordered stream of fields; a 3D Chinese or English character string uses depth; and a character movie also uses time. These are one continuous input/output contract, not separate token vocabularies.

The evidence trajectory is narrower than the product goal. V24 passes a designed variable-length visual packet grammar with generated-image rereading. V25 through V31 replace that grammar with ordinary Chinese and repeatedly find visual or order-sensitive signals without reliable next-form binding. V33.1 shows that a direct linear raster adapter is readable but below its fixed interface gate. V34 then qualifies a compact, codebook-free continuous writing codec. V35 combines that codec with credited PIXAR initialization and closes the direct raster generation and rereading loop, but still fails copying and instruction semantics. V36 then isolates a candidate-free answer-level visual plan. It learns a measurable counterfactual relation but fails its complete semantic gate, exposing both a post-augmentation occupancy-mask defect and a low-rank visual target space. V37 fixes clean masks, adapts the reader end to end, and distills a strong multilingual semantic geometry offline. It raises prompt top-1 to 47.45% and answer-plan top-1 to 20.41%, but still fails the conjunctive semantic gate because absolute alignment, font/wording invariance, margins, and length remain inadequate. V38 improves visual reading and invariance without solving held-out answer semantics; V39.1 calibrates answer length but fails trajectory semantics; and V41 qualifies an isolated external glyph motor. V42 then provides the first bounded positive ordered-raster language result. V43 adds a capable bank-free spatial writer but diagnoses memorized pair supervision and an inadequate autonomous image plan. V44 removes that memorization gap with a one-pass frozen-base residual, yet fails held-out binding and drifts toward the corpus-common image field. V45 then qualifies a centered, variance-balanced, exactly invertible raster representation on all 13 fixed geometry, continuity, boundary, and resource gates. Its failed V42 retrofit shows why the next reader must train from scratch in the new coordinates; the V43 writer remains closed. Page-scale prompting, historical answer generation, 3D geometry, and motion remain deferred.

Concretely, the intended model maps prompt frames X_prompt[Tp,D,H,W,C] to generated answer frames Y_answer[Ta,D,H,W,C]. Typed questions are rendered into X_prompt; scanned pages or handwriting enter directly. Ta=1 is an answer page and Ta>1 is a text-image stream or movie. A valid understanding result must change the generated answer appropriately under held-out prompt changes; reconstruction, OCR, glyph classification, and attractive writing alone do not satisfy it.

The deployed student must not call Qwen, an OCR engine, a tokenizer, a Unicode lookup, or a glyph database to decide its answer. External models and extracted text may help build and audit an offline curriculum, but every student batch and checkpoint must pass a boundary receipt showing that its learned path contains only writing pixels and continuous visual states. The measurable roadmap and source-book policy are in docs/first-imagized-language-model-goal.md and references/word_origin_ilm_dataset_plan.md.

Earlier Natural-Language Test: V31 Conditional Visual Flow Rejected

V31 tests whether coherent conditional flow can repair V30's deterministic next-field averaging. Two parameter-identical 18,736,577-parameter students start from byte-identical initialized states. Both read 64 ordered 32 x 32 Chinese writing images with a causal QKV visual reader. The spatial arm learns a conditional velocity over a 16 x 192 retinal field; the global control learns one semantic vector tiled across the same 16 cells. Neither student receives text, token or Unicode IDs, OCR, a glyph lookup, vocabulary logits, or a candidate bank.

Each arm completes 10,000 finite BF16 updates on one RTX 4090, using 0.997 GiB peak allocated memory. The fixed audit uses 2,048 natural windows, 512 pixel-identical suffix pairs, eight path probes, and eight-step Heun sampling. Spatial full-context path top-1 is only 0.0977%, below the global control (1.8555%), image unigram (1.6113%), symbolic bigram (13.5254%), and symbolic trigram (20.9961%). Spatial autonomous top-1 is 0.1465%.

The spatial path's target log probability improves by 0.1259 nat over a suffix-preserving prefix shuffle and by 0.0527 nat over spatial permutation. Its autonomous samples are diverse and context dependent. These effects do not bind meaning to output: exact-suffix path assignment is 50.4883% and autonomous assignment is 50.1953%, both effectively chance and below the matched control.

V31 generates continuous latent retinal fields, not pixels. The panel above shows the external evaluator's nearest glyph for each field and exposes repeated wrong modes; those glyphs are diagnostic proxies, not model-rendered output. The spatial arm passes 14/19 common gates, global passes 6/6 integrity gates, matched arms pass 4/8, and language plus generation passes 0/10. V31 is rejected, frozen images remain uninstantiated, and direct pixel writer training remains unauthorized under this protocol.

The next proof separates visual semantic planning from rendering: causal QKV attention must predict a multi-glyph answer state, while a compact continuous renderer must emit the answer raster directly. Diffusion, rectified flow, or a continuous autoregressive head may implement rendering, but held-out semantic counterfactuals and readable direct pixels must carry the language claim. See the complete V31 receipt, preregistered protocol, and research decision.

Prior Natural-Language Test: V30 Spatial Next-Field Rejected

V30 tests the spatial predictive target proposed after V29. Two independently trained, parameter-identical 18,641,153-parameter students start from byte-identical initialized states. Both read 64 ordered 32 x 32 Chinese writing images and emit a candidate-independent 4 x 4 x 192 continuous next-image field. The spatial arm compares corresponding frozen retinal cells; the control tiles a matched-visibility global semantic vector across the same 16 rows. Neither arm receives text, token or Unicode IDs, OCR, a glyph lookup, vocabulary logits, or a deployed candidate bank.

Each arm completes 8,000 fixed BF16 updates on one RTX 4090. Peak allocated memory is 1.595 GiB and 1.597 GiB. On 2,048 natural development windows, spatial full-context 1,024-way top-1 is 1.2695%, below the global control (2.4902%), image unigram (1.3184%), and symbolic bigram (11.7188%). Spatial full target log probability improves by 0.31534 nat over a suffix-preserving prefix shuffle, but is 0.19040 nat worse than the global control.

Reversing the candidate's 16 local cells changes spatial scores and drops natural top-1 to 0.0488%, confirming that local geometry is visible. The target log-probability gain is only 0.01541 nat, below the fixed >0.05 gate. On 512 pixel-identical-suffix pairs, spatial assignment is 50.0488%, global assignment is 50.5859%, and spatial patch reversal changes accuracy by only 0.3906 percentage point. Exact suffix rows, candidate-column equivariance, cross-font visibility, finite-state, boundary, and resource controls all pass.

The spatial arm passes 12/18 common gates, the control passes 12/12 integrity gates, matched arms pass 5/9, and spatial language passes 0/8. V30 is rejected, frozen images remain uninstantiated, and no writer is trained. The result rejects a single deterministic bilinear next-field as the selected mechanism; it does not reject continuous image-native language modeling. See the complete V30 receipt, preregistered protocol, and research decision.

Prior Natural-Language Test: V29 Conditional Visual Density Ratio Rejected

V29 tests the candidate-conditioned incremental-evidence hypothesis proposed after V28. Its 20,080,961-parameter image-only student reuses the frozen V16 retina and frozen V28 semantic adapters, retains the eight-layer causal visual field, and lets an arbitrary candidate image query all 64 context states through two cross-attention layers. It scores full context F, exact suffix B, and the centered increment G = F - B without token IDs, Unicode IDs, OCR, vocabulary logits, glyph lookup, or a deployed candidate bank.

The preregistered run performs 8,000 BF16 updates and the complete development audit in 82.60 minutes on one RTX 4090, with 3.356 GiB peak allocated CUDA memory. On 2,048 natural windows, full-context 1,024-way top-1 is 2.3438%, above image unigram (1.3672%) but far below symbolic bigram (13.8672%). Full target log probability improves by 0.17945 nat over suffix-4 and by 3.10248 nat over suffix-preserving prefix shuffle. The model clearly detects natural order, but that effect does not become useful next-image ranking.

The decisive 512-pair audit holds the final four glyph images and suffix score rows exactly equal. Raw two-candidate identity is 99.9512%, all candidate permutation errors are zero, and the student/checkpoint boundary is clean. Incremental assignment nevertheless reaches only 50.7080%, versus 49.5605% after prefix shuffling; both-correct is 8.9844%. Full score alone reaches 49.7314%. V29 passes 8/14 mechanism gates and 2/6 language gates.

For exact shared suffixes, the baseline terms in G = F - B cancel in the aggregate two-by-two assignment margin. The audit confirms identical full and incremental mean margins (0.005103). A suffix baseline can redistribute row margin, but cannot repair a full critic that has not learned the context-candidate interaction. V29 is rejected, the frozen partition remains sealed, and no writer is trained. The next bounded question is whether a causal field can predict a continuous next-image patch map and compare candidate patches before scalar reduction. See the complete V29 receipt, preregistered protocol, and research decision.

Prior Natural-Language Test: V28 Dense Visual Future Energy Rejected

V28 tests the dense ordered-future objective proposed after V27. Its 17,859,142-parameter image-only student freezes the V16 raw retina, learns an identity-initialized semantic residual with an EMA target, integrates 64 glyph images with eight causal blocks, and scores continuous visual futures at horizons 1, 2, and 4. Four image-derived hypotheses per position make the training signal dense without introducing token IDs, Unicode IDs, OCR, vocabulary logits, or a glyph lookup in the student path.

The one preregistered run performs 10,000 BF16 updates and its complete development audit in 118.91 minutes on one RTX 4090, with 1.144 GiB peak allocated CUDA memory. On 2,048 natural windows, full-context top-1 is 1.4160%, below the image unigram (1.8555%) and symbolic bigram (13.1348%). Full context improves target log probability over suffix-4 by 0.03037 nat and over a suffix-preserving prefix shuffle by 0.21511 nat, so the field detects order, but not enough to select the correct future.

The decisive 512-pair audit holds the final four glyph images bitwise equal while changing earlier history and the target. Candidate permutation error is zero, and the frozen V16 retina identifies the two cross-font candidates at 99.9512%. Full-context assignment is nevertheless 49.5605%, versus 49.9512% after shuffling the prefix. Its mean score margin does improve by 0.02207, but that probability movement does not become reliable rank or binding. On the separate 1,024-way identity audit, the learned EMA semantic route improves over the same-scope raw retina from 92.0410% to 96.4355%.

V28 passes 10/14 mechanism gates and 2/6 language gates. It is rejected, the frozen partition remains sealed, and no writer is trained. V29 executes the candidate-conditioned prefix-incremental test and is reported above. See the complete V28 receipt, preregistered protocol, and research decision.

Prior Natural-Language Test: V27 Joint Compatibility Rejected

V27 tests whether a compact model can learn language by scoring an arbitrary next-glyph image directly from 64 preceding glyph images. Its 18,599,553-parameter image-only student initializes an online retina from V16, builds a context query with eight causal rotary blocks, and scores an EMA candidate-image key. It contains no strings, token or Unicode IDs, OCR, vocabulary matrix, codebook, candidate bank, glyph lookup, or external model. Candidate order is randomized independently, and the evaluator removes that permutation before scoring.

The single preregistered run performs 8,000 BF16 updates on one RTX 4090. Training plus audit takes 39.12 minutes with 2.268 GiB peak allocated CUDA memory. The fixed 2,048-window natural audit reaches 1.6113% top-1, below the image unigram (2.0508%) and symbolic bigram (12.5977%). Full context improves target log probability over suffix-4 by 0.05925 nat, but only 0.00273 nat over a suffix-preserving prefix shuffle.

The decisive 512-pair audit keeps the final four glyph images bitwise equal while changing earlier history and the target. The unchanged V16 retina identifies cross-font candidate forms at 99.9512%, and candidate permutation error is exactly zero, so candidate visibility and row-position shortcuts are controlled. Full-context assignment is nevertheless 50.7080%, compared with 50.5615% after shuffling the prefix. Separately, learned cross-font identity over the 1,024-image bank is 94.8730% and misses its 99% gate; that 1,024-way metric is not directly comparable to the two-candidate raw control. V27 passes only 7/13 mechanism gates and 1/5 language gates. It is rejected, the frozen partition remains sealed, and no writer is trained.

V28 executes this proposed dense-future test while keeping the full N x 1 x 32 x 32 glyph-image stream authoritative and the raw retina frozen. Its result is reported above. A reversible 2D lattice can accelerate that stream later; depth and motion remain observable extensions rather than identity encodings. See the complete V27 receipt, preregistered protocol, and research decision.

Prior Natural-Language Test: V26 Factorized Context Rejected

V26 tests the repair proposed after V25 without changing the ordinary-Chinese or image-only boundary. Its 19,142,721-parameter model gives the last visible glyph and the preceding 63 glyph images separate routes, fuses their continuous states, and predicts eight 192-dimensional visual particles for each of future horizons 1, 2, 4, and 8. The deployed student receives no strings, token or Unicode IDs, OCR, labels, glyph table, candidate bank, or external model state.

The fixed run completes 8,000 BF16 updates in 31.05 minutes on one RTX 4090, using 0.888 GiB peak allocated CUDA memory. On 2,048 development windows, full-history top-1 is 0.0488%, versus 1.4160% for the image unigram and 13.5254% for the symbolic bigram. Full history improves target log probability over last-only by 0.17045 nat, but only by 0.01973 over the same four-cell suffix and 0.00369 over a prefix shuffle.

The decisive audit uses 512 cross-record context pairs with pixel-identical four-glyph suffixes and different targets. Their appearance-state difference is exactly zero and their mean history-residual difference is 4.62826, so the history branch is active. Correct pair ranking and swapped-residual target accuracy are nevertheless both exactly 50%, with mean score margin 0.0000677. A perfect retina-bank oracle rules out a blind evaluator. V26 therefore fails its mechanism and language gates; no frozen evaluation or writer is authorized.

This localizes the failure to conditional binding, not visual detection or the mere existence of a history state. Low memory and materially different hidden states are not useful language prediction. See the complete V26 receipt, preregistered protocol, and research decision.

Pre-V27 Localization: Frozen Compatibility Probe

A post-hoc diagnostic freezes all 19.14M V26 parameters and trains three small candidate-conditioned image scorers for one pass over the existing 16,384 train suffix pairs. On 512 disjoint-record development pairs and 2,048 cross-font decisions, appearance-only accuracy is exactly 50.000%, history-residual accuracy is 50.684%, and fused-state accuracy is 50.342%. The same frozen retina identifies the paired target image across fonts at 99.951%, with a 0.74724 mean cosine margin. Candidate ambiguity therefore does not explain the chance language result.

This diagnostic is not preregistered evidence and opens no frozen data. It motivated V27's joint causal-context and deterministic image-candidate test. The preregistered result above shows that this relation did not pass, so no stochastic writer was authorized. See the diagnostic receipt and V27 research decision.

Prior Natural-Language Test: V25 Visual Cell Stream Rejected

V25 is the first experiment here to train directly on ordinary Chinese book language as an ordered 64 x 1 x 32 x 32 visual-time stream. Its 25,549,714-parameter model uses a frozen image retina, an eight-layer causal continuous field, a vocabulary-free next-state proposal, and a flow writer that can append and reread actual generated pixels. The student receives no strings, token or Unicode IDs, OCR transcript, character labels, glyph lookup, discrete codebook, or external model call.

The fixed 2,400-update language run completes on one RTX 4090. On 2,048 development windows, full 64-cell history reaches 1.123% next-cell top-1, versus 0.146% last-only and 0.342% with prior history shuffled. This is a real ordered-history effect, but it is too small: the image unigram reaches 1.611% and the symbolic bigram 12.158%. Counterfactual switch accuracy is 12.891%, target cosine is 0.2751, and six fixed semantic/causal gates fail. Peak allocated CUDA memory is 0.598 GiB; low memory is not evidence of language efficiency when predictive quality remains below a bigram. The frozen partition stays sealed.

An explicitly labeled exploratory writer run after rejection preserves position-16 ink density (0.977x) and avoids blanks, but reaches 0% generated identity top-1, 0.0802 reread cosine, and 0.3221 pixel F1. Its output is glyph-like texture, not readable continuation. This diagnostic does not alter the fixed evidence verdict.

The result points to the next controlled problem: separate exact visible appearance from a context-predictive residual and prove that both states are causally needed before scaling context. The implemented reversible serpentine visual lattice can later fold up to 65,536 clean cells into a long-context retinal field without losing the authoritative 32x32 glyph stream, but it was not used in V25 and is not a claimed fix. See the complete V25 receipt and unchanged frozen protocol.

Earlier Stream Test: V24 Visual Packet Rereading Accepted

V24 is the first accepted variable-input, multi-frame-output visual stream in this repository. The student receives 15, 18, 21, or 24 grayscale 32x32 frames grouped into visibly headed packets. It locates two bindings, an operation, and a query from header images; emits the selected unseen Chinese glyph as frame 1; rereads the actual generated pixels through the frozen visual retina; and emits the glyph's visibly bound label as frame 2. Its deployed path receives no strings, token or Unicode IDs, OCR, role or operation labels, packet indices, active length, padding mask, glyph lookup, discrete codebook, or external language model.

Only 1,347 parameters are trained. On a fresh 1,024-episode paired audit, query, operation, and generated-history switch accuracy are 0.99219, 0.99316, and 0.99609. The corresponding query-blind, operation-blind, and history-blind controls each have exactly 0.0 switch accuracy and 0.0 output-pixel change for the factor they cannot observe. A header-blind control falls to 0.08203 minimum role localization, versus 1.0 for the candidate. Every arm has identical parameter names, shapes, and count.

An opaque agent visual audit, performed before opening its sealed answer key, scores 47/48 for frame 1 and 48/48 for frame 2, including 12/12 for both frames at the held-out T=24 length. This is an agent visual audit, not a human study.

The single authorized frozen run covers 107 unseen identities and 1,024 episodes. It performs no model selection, changes no threshold, and is not repeated.

V24 frozen gate	Measured	Required	Result
Frame-1 binary choice	0.99805	>0.95	pass
Query switch	0.97656	>0.90	pass
Operation switch	0.97266	>0.90	pass
Generated-history switch	0.99609	>0.90	pass
Held-out minimum switch	0.96353	>0.85	pass
Frame-1 identity top-1	0.98672	>0.75	pass
Frame-2 label top-1	0.99727	>0.95	pass
Frame-1 / frame-2 pixel F1	0.83714 / 0.72860	>0.68 / >0.58	pass
Held-out T=24, frame 1 / frame 2	0.98473 / 1.00000	each >0.90	pass
Packet-permutation consistency	1.00000 / 1.00000	each >0.99	pass

V24 proves a fixed packet grammar and causal two-frame image answer, not arbitrary sentence understanding or a finished language model. Packet arity, header semantics, same/other algebra, and output length remain designed into the task. It does not yet answer etymology questions, continue pages, write unrestricted text, or emit a movie. The next milestone must learn from rendered Chinese prompts and passages, generate a learned-length image-line stream, and pass semantic counterfactuals and blind-history controls. See the complete V24 receipt.

Prior Prompt Test: V23 Visual Relation Circuit Accepted

V23 is the first complete positive image-prompt-to-image-answer result in this repository. Six 32x32 writing images enter the student and one 32x32 answer image comes out. The prompt visibly binds two previously unseen Chinese glyphs to two labels, supplies 同 or 异, and ends with a visual query label. The student compares images, reads the operation from its image, routes one visible source glyph, and renders a canonical answer. Its learned path receives no strings, token or Unicode IDs, OCR, character labels, answer indices, codebook, glyph lookup, or external language model.

The relation-aware candidate selected under the fixed development protocol. On a fresh 1,024-episode paired audit it reaches 0.99805 query and operation switch accuracy and 0.99512 identity top-1. Query-blind and operation-blind controls have exactly 0.0 switch accuracy and 0.0 output-pixel change for the factor each cannot see. An opaque agent visual review then scores 48/48 overall and 12/12 on held-out compositions before the sealed key is opened.

The single authorized frozen run covers 98 unseen identities, 1,024 episodes, and 4,096 prompt variants. It performs no model selection and changes no threshold.

V23 frozen gate	Measured	Required	Result
Binary choice	0.99829	>0.95	pass
Query switch	0.99609	>0.90	pass
Operation switch	0.99707	>0.90	pass
Held-out minimum switch	0.99606	>0.85	pass
Unseen-identity top-1	0.99463	>0.75	pass
Pixel F1	0.78478	>0.68	pass
Target cosine	0.93994	>0.82	pass
Query-label visual match	0.99951	>0.98	pass
Operation-gate accuracy	1.00000	>0.98	pass
Pair-swap consistency	1.00000	>0.99	pass

V23 proves bounded visual relation following, not open-ended language. Frame roles and the two-pair same/other algebra remain fixed; the output is a canonicalized form of one visible source glyph. It does not yet parse arbitrary sentences, answer etymology questions, continue pages, or emit an image stream or movie. V24 takes the next bounded step by removing absolute frame roles, reading a variable-length packet stream, and generating two answer frames while rereading the first. See the complete V23 receipt.

Prior Prompt Test: V22 Binding Mechanism Rejected

V22 is the first bounded implementation of the requested visual prompt stream: six 32x32 writing images enter the student and one answer image comes out. Each prompt visibly binds two previously unseen Chinese glyph images to labels, shows 同 or 异, and ends with a visual query label. A paired counterfactual changes only that final image, so a model that understands the prompt must switch its generated answer. The query-aware candidate and query-blind control each have exactly 3,410,128 trainable parameters and use no strings, token/Unicode IDs, OCR, glyph lookup, answer codebook, or external language model.

The model does not pass. At step 1,600, candidate switch accuracy is only 0.0078, versus 0.0 for the query-blind control. Identity top-1 is 0.1592 versus 0.1533, and pixel F1 is 0.5107 versus 0.5125. Changing the visible query changes candidate pixels by only 0.0089 mean L1, far below the fixed 0.08 requirement.

V22 candidate development gate	Measured	Required	Result
Binary choice	0.4824	>0.85	fail
Counterfactual switch	0.0078	>0.80	fail
Held-out-combination switch	0.0113	>0.75	fail
Unseen-identity top-1	0.1592	>0.45	fail
Identity gain over query shuffle	+0.0225	>0.20	fail
Pixel F1	0.5107	>0.58	fail
Oracle-writer F1	0.6020	>0.64	fail
Paired-output L1	0.0089	>0.08	fail
Frozen images instantiated	0	0	pass

The endpoint audit explains why: the candidate gives the operation frame the maximum selector weight in all 1,024/1,024 original and counterfactual prompts, with mean operation attention 1.0 and mean query attention 1.37e-13. A relational answer needs the operation, query-to-label match, and label-to-glyph binding jointly; collapsing six frames into one selected frame cannot perform that composition. V23 therefore replaces single-frame selection with an explicit, differentiable multi-frame visual relation circuit. No V22 candidate selected, so paired, human, and frozen evaluation remain forbidden. See the complete V22 receipt.

Prior Causal Test: V21 Field-Complete Route Works, Writer Rejected

V21 tests whether a continuous local field can carry the complete visual plan. Its candidate and tiled-global control each have exactly 582,336 trainable parameters. Every local cell emits coarse occupancy plus 63 Walsh--Hadamard zero-DC coefficients for its corresponding 8x8 patch. Global state and style provide only spatially uniform modulation; there are no coordinates, position parameters, cell mixing, or global spatial projection.

At the best diagnostic step 1,400, correct-field dense F1 is 0.7053, versus 0.5314 after shuffling the field and 0.3588 after zeroing it. The fixed gains +0.1739 and +0.3465 pass. Identity top-1 is 79.10%, target cosine is 0.8331, all exact-basis invariants pass, and quadrant locality is exactly 1.0. The equal-parameter control collapses to repeated textures, reaching only 0.1473 overall and 0.3074 dense F1 at its selected structural step.

V21 candidate development gate	Measured	Required	Result
Overall pixel F1	0.6038	>0.66	fail
Simple pixel F1	0.5648	>0.58	fail
Medium pixel F1	0.5945	>0.60	fail
Dense pixel F1	0.7053	>0.70	pass
Dense gain over shuffled field	+0.1739	>0.15	pass
Dense gain over zero field	+0.3465	>0.20	pass
Identity top-1	79.10%	>74%	pass
Target cosine	0.8331	>0.82	pass
Occlusion locality	1.0000	>0.95	pass
Detail block-mean magnitude	3.87e-7	<5e-6	pass

The writer is still rejected because no candidate checkpoint passes all quality gates. The comparison to the control is descriptive, not a formal paired audit; the paired evaluator must refuse an unselected candidate. Human review and frozen evaluation were not authorized, and frozen images remained uninstantiated. V21 proves a field-complete causal route, not prompt understanding or autonomous language generation. See the complete V21 receipt.

Prior Causal Test: V20 Routes Local Detail, Writer Rejected

V20 reserved within-block detail for a local 4x4x192 field while global state supplied coarse occupancy. It passed the field-shuffle (+0.1218), zero-field (+0.3613), and locality (1.0) gates, but failed overall F1, target cosine, an exact-decomposition invariant, and the matched-control margin. V21 removed that remaining global spatial route. See the complete V20 receipt.

Prior Routing Test: V19 Rejected

V19 tested the next proposed correction instead of assuming it worked. A clean 2,358,977-parameter global planner was first trained from scratch on a new salted split. Its weights and the V16 retina were then frozen while a 764,545-parameter adapter learned from the retina's continuous 4x4x192 spatial field. The target image supplied loss only. The student still received no token IDs, Unicode IDs, OCR, strings, character labels, lookup, codebook, or external language model.

On a fresh 512-candidate development audit, dense pixel F1 is 0.7278, but the correct field beats a shuffled field by only 0.0088 and a zero field by only 0.0054. The prospectively fixed margins were >0.12 and >0.03. Overall F1 (0.6710) and identity top-1 (72.66%) also miss their fixed gates.

V19 development gate	Measured	Required	Result
Overall pixel F1	0.6710	>0.68	fail
Dense pixel F1	0.7278	>0.58	pass
Dense gain over shuffled field	+0.0088	>0.12	fail
Dense gain over zero field	+0.0054	>0.03	fail
Identity top-1	72.66%	>75%	fail
Target cosine	0.8416	>0.84	pass

The automatic gate is rejected, so human review was not authorized and the frozen split remains sealed. The result identifies a routing failure: an optional local residual can polish a complete global writer without making local topology necessary. V20 subsequently fixed that causal routing defect but still failed writer selection. See the full V19 result and the 2026 continuous-sensory decision scan.

Accepted Development Proof: Visual Motor Plan V18

V18 is a real 2,358,977-parameter deterministic visual motor planner trained for 1,600 updates on one RTX 4090. It receives a continuous 192-dimensional state read from a different-font image plus a separate style image and emits a 32x32 continuous ink plan. Target pixels provide loss only. The learned path has no token IDs, Unicode IDs, OCR, strings, character labels, output vocabulary, glyph lookup, finite visual codebook, candidate classifier, or external language model.

On a fresh 512-example development-only audit, V18 reaches 73.63% global visual-identity top-1 versus 0.98% after shuffling only intended states. Target cosine is 0.8462 versus 0.0716; pixel F1 is 0.6577 versus 0.3129. Ink occupancy is identical in both branches. Most reviewed simple and medium forms are recognizable, while dense forms can still merge strokes.

Fresh development measurement	Correct intent	Shuffled intent	Result
Global visual identity top-1	73.633%	0.977%	strong causal control
Target cosine	0.8462	0.0716	gain +0.7746
Pixel F1	0.6577	0.3129	automatic topology gate passes
Human review	simple/medium readable	unrelated forms	dense forms still fail
Peak allocated CUDA memory	0.778 GiB	-	far below 4090 capacity

This breaks a categorical claim: structured writing can be learned and emitted as continuous image topology by a small consumer-GPU model without a token output table. It does not prove autonomous language generation. V18 is supplied the intended state, the human gate lacked a prespecified numeric rubric, and the new frozen bank remains untouched. The full protocol, limitations, and reproduction receipt is in docs/visual-motor-plan-v18-result.md.

Prior Causal Actuator: V17

V17 is a real 5,729,921-parameter visual actuator trained for 1,600 updates on one RTX 4090. It receives a 192-dimensional state read from a different-font image of the intended form plus a continuous style image, then generates 32x32 ink pixels. The target image supplies loss only; its spatial pixels do not enter the condition. The learned path has no token IDs, Unicode IDs, OCR, character labels, output vocabulary, visual codebook, candidate classifier, glyph lookup, or external language model.

On one untouched frozen split of 512 generated examples, V17 obtains 58.59% global visual-identity top-1 versus 0.98% when intended states are shuffled while style and initial noise remain fixed. Target-state cosine is 0.7130 versus 0.0861, a gain of +0.6269. The intervention establishes that a small continuous visual state causally controls generated writing pixels rather than merely copying style.

Frozen actuator gate	Correct state	Shuffled state	Result
Global visual identity top-1	58.594%	0.977%	passes causal-control gate
Target cosine	0.7130	0.0861	gain +0.6269
Pixel F1	0.4385	0.2800	fails required 0.5000
Human readability	rejected	rejected	pseudo-characters remain
Peak allocated CUDA memory	1.588 GiB	-	fits far below 4090 capacity

This breaks a categorical claim, not the whole language problem: a consumer GPU can train a compact non-token visual state to control image generation. V17 is still rejected as a readable actuator. Its mostly pseudo-character outputs show that retinal identity can be optimized without preserving exact stroke topology. It is also an isolated actuator test supplied with the intended state, not autonomous next-language generation. V16 remains below a symbolic bigram.

V18 implements the correction: it decodes continuous intent into a directly supervised spatial visual motor plan and makes readable writing emerge on development data. V17 remains the frozen causal-control baseline. Its complete selection, frozen receipt, limitations, and reproduction command are in docs/visual-state-actuator-v17-result.md.

Language Core: Predictive Visual Field V16

V16 is a real 16,471,809-parameter image-only causal state model trained on one RTX 4090. It adds a 6,001,536-parameter residual multiscale memory, with dilated local visual fields and global causal attention, to the proven V15 recurrent base. Its learned path receives sequences of 32x32 writing images and produces continuous next-image states. It has no token IDs, Unicode IDs, OCR, character labels, output vocabulary, visual codebook, candidate classifier, or external language model.

On the unchanged frozen 512-form, four-view Chinese benchmark, the selected continuous proposal obtains 6.264% top-1 (112/1,788), versus 3.971% with only the last image, 1.734% unigram, 0.224% random dynamics, 0.195% chance, and 13.143% symbolic bigram. Full image history adds +0.0773 normalized target log-probability. The parallel stochastic field obtains 3.691%, versus 3.244% last-only, with sampled-state context cosine gain +0.0795.

Frozen gate	V15	V16	Result
Proposal full-context top-1	5.872%	6.264%	+7 correct contexts; directional only
Proposal last-image top-1	4.418%	3.971%	V16 context separation is larger
Proposal context log-probability gain	+0.0707	+0.0773	full history helps
State-flow full-context top-1	3.412%	3.691%	beats last-only and unigram
Symbolic bigram	13.143%	13.143%	not beaten
Peak allocated CUDA memory	1.181 GiB	1.479 GiB	fits far below 4090 capacity
Coupled pixel actuator	absent	absent	isolated V18 planner is readable on development, not yet coupled

This breaks a narrow but important claim: a small model can learn causal language signal directly from rendered writing images on a consumer GPU. It does not yet establish general language understanding, readable image generation, historical question answering, or parity with an LLM. Seven extra correct frozen contexts are not a statistically established architecture win. The V16 selection, compute, gate, and limitation receipt is in docs/predictive-visual-field-v16-memory-result.md; the V8-V15 history remains in docs/predictive-visual-field-v15-result.md.

Paradigm: Separate Language From Drawing

RFLM V7 exposed a structural error: one conditional pixel flow was being asked to discover the next linguistic identity and render its strokes in the same operation. V14 through V16 now implement the first half of a factorized solution without relaxing the image-only boundary:

1. A retina learns a continuous manifold directly from writing images.
2. A causal field predicts a low-variance continuous visual proposal.
3. A hyperspherical flow models a distribution over alternative next states.
4. A deterministic visual motor planner renders intended topology; V18 makes simple and medium held-out forms readable on development data.
5. Optional stochastic flow can refine style only after topology is stable.
6. The retina will reread the rendered pixels and feed them back into the field.

There is no nearest-character lookup or output vocabulary. The continuous state proof now passes random, last-only, unigram, context-use, and target-signal gates. It still fails the bigram language gate. V18 passes automatic development topology gates but is not frozen-promoted because its human rubric was underspecified and dense forms still fail. V19 then rejects an additive spatial residual as the repair: correct, shuffled, and zero spatial fields produce nearly the same output. V20 makes local detail structurally necessary and topographic, but still misses writer quality and matched-control gates. V21 makes both occupancy and detail field-causal and passes every structural invariant, but its disjoint patches miss simple, medium, and overall quality. The autonomous prompt-to-answer write-reread loop remains withheld until a continuity-preserving local writer and the language core pass independently.

The strict student boundary remains:

writing pixels -> continuous visual dynamics -> continuous ink pixels

The student receives no strings, token IDs, Unicode IDs, OCR transcript, character labels, external language model, or discrete visual codebook. Typed input is supported only by deterministic rasterization before this boundary. An uploaded page can enter directly as pixels.

Earlier precursor: Retinal Flow V7

The earlier runnable model is an 11.69M-parameter Retinal Flow Language Model, a concrete read-predict-write-reread loop:

1. A small convolutional retina reads ordered 32x32 grayscale fixations.
2. A three-layer recurrent visual field integrates the fixation history.
3. A continuous energy function scores arbitrary candidate images; it has no character output table.
4. A conditional rectified flow writes the next fixation directly in pixel space.
5. The model rereads its generated ink, selects a candidate by visual energy, and feeds those pixels back into the recurrent state.

Measured status, not a capability claim

V7 kept the model at 11,690,244 parameters, added 800 updates on one RTX 4090, and generated 25.3 visual cells per second in its matched run. It added normalized context advantage against independent image anchors and backpropagation through sampled flow endpoints. V6 and V7 were tested on the same 512 common Han characters, four font views, 2,423 eligible held-out contexts, and frozen bank SHA-256.

Gate	V6 closed loop	V7 selected step 5,800	Interpretation
Retina oracle top-1	98.18%	98.27%	Basic cross-font perception is not the main bottleneck.
Full-context top-1	1.20%	2.31%	V7 beats last-only (2.02%) and unigram (1.86%), but not bigram (13.58%).
Normalized context log-probability gain	-0.9066	-0.2155	The calibrated deficit shrank by 76%, but full history still lowers mean target probability.
Generated context cosine gain	+0.0077	+0.0303	V7 passes the held-out generated-signal gate.
Late/early autonomous ink	1.168	1.050	Both loops keep nontrivial ink without late occupancy drift.
Sparse autonomous cells	18.75%	15.63%	V7 is denser, but its continuation is still unreadable.

Verdict: V7 is rejected as a language model. It establishes a useful training correction, not a complete language system. Raw target energy was positive while normalized target probability was negative, proving that raw score margins were an invalid acceptance measure. V7 does not prove readable continuation, historical question answering, efficiency over a text LLM, or Qwen-8B parity. The result motivates the Predictive Visual Field separation shown above.

Reproduce The V21 Field-Complete Test

The two fixed arms must be trained separately. They use the same frozen V16 retina and exactly matched parameter counts:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python scripts/train_field_complete_writer.py \
  --pvf-checkpoint artifacts/predictive_visual_field_v16_memory_pilot/checkpoint_step_0002200.pt \
  --route-mode field_complete \
  --out artifacts/field_complete_writer_v21_field_evidence_20260813

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python scripts/train_field_complete_writer.py \
  --pvf-checkpoint artifacts/predictive_visual_field_v16_memory_pilot/checkpoint_step_0002200.pt \
  --route-mode tiled_global_control \
  --out artifacts/field_complete_writer_v21_control_evidence_20260813

The paired evaluator requires two selected checkpoints. It deliberately rejects the measured candidate because no candidate checkpoint passed selection; it cannot access the frozen partition. Full hashes, metrics, and the expected rejection command are in the V21 result receipt.

Reproduce The V20 Topology Test

The prior V20 commands, exact hashes, and expected paired-evaluator rejection remain in the V20 result receipt.

Audit The V19 Spatial Test

Reproduce the fresh V19 development audit. The evaluator verifies the clean global-baseline hash and cannot access the sealed frozen split:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python scripts/eval_spatial_motor_plan_development.py \
  --checkpoint artifacts/spatial_motor_plan_v19_pilot/checkpoint_latest.pt \
  --out artifacts/spatial_motor_plan_v19_step1600_development_audit \
  --samples 128 --batch-size 32 --num-workers 8 \
  --sample-count 32 --sample-columns 8 --device cuda --precision bf16

The fixed protocol, full training commands, hashes, and failed gate are in docs/spatial-retinal-motor-plan-v19-result.md.

Run The Visual Motor Plan

Audit the selected V18 checkpoint on fresh development renderings. This command cannot access the sealed frozen split:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python scripts/eval_visual_motor_plan_development.py \
  --checkpoint artifacts/visual_motor_plan_v18_pilot/checkpoint_selected_development.pt \
  --out artifacts/visual_motor_plan_v18_step1400_development_audit_v2 \
  --samples 128 --batch-size 32 --num-workers 8 \
  --sample-count 32 --sample-columns 8 --device cuda --precision bf16

The evaluator refuses to overwrite an existing receipt. Full training settings and the sealed-development decision are in docs/visual-motor-plan-v18-result.md.

Run The V17 Baseline

Evaluate the selected V17 checkpoint once on its frozen record split:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python scripts/eval_visual_state_actuator.py \
  --checkpoint artifacts/visual_state_actuator_v17_pilot/checkpoint_step_0001600.pt \
  --out artifacts/visual_state_actuator_v17_frozen_eval \
  --samples 128 --batch-size 32 --num-workers 8 \
  --sample-count 12 --device cuda --precision bf16

The evaluator refuses to overwrite an existing frozen receipt. Full training settings and the rejected readability audit are in docs/visual-state-actuator-v17-result.md.

Run The Predictive Visual Field

Evaluate a trained PVF checkpoint on the fixed image bank:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python scripts/eval_predictive_visual_field.py \
  --checkpoint artifacts/predictive_visual_field_v16_memory_pilot/checkpoint_step_0002200.pt \
  --out artifacts/predictive_visual_field_v16_step2200_eval \
  --device cuda \
  --precision bf16

The implementation, exact V16 continuation settings, checkpoint-selection rule, and metric definitions are recorded in docs/predictive-visual-field-v16-memory-result.md. Training and evaluation artifacts remain git-ignored.

Run The Retinal Precursor

Build the provenance-bearing public-domain Chinese manifest:

PYTHONPATH=. python scripts/build_visual_grammar_manifest.py \
  --wikisource-root ../Books/resources/curated-books/chinese-classics/public-domain-canon \
  --out data/visual_grammar/chinese_wikisource_public_domain.jsonl

Train the current combined RFLM objective from scratch on one 24 GiB GPU:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python scripts/train_retinal_flow_lm.py \
  --manifest data/visual_grammar/chinese_wikisource_public_domain.jsonl \
  --out artifacts/retinal_flow_chinese_anchor_identity \
  --sequence-length 48 \
  --energy-positions-per-sequence 8 \
  --batch-size 32 \
  --maximum-steps 6000 \
  --context-anchor-bank-size 512 \
  --context-anchor-views 4 \
  --context-advantage-weight 0.5 \
  --context-advantage-margin 0.5 \
  --sampled-identity-weight 0.2 \
  --sampled-identity-steps 2 \
  --rollout-start-step 800 \
  --rollout-ramp-steps 400 \
  --rollout-batch-size 8 \
  --rollout-steps 2 \
  --rollout-candidates 2 \
  --rollout-sample-steps 2 \
  --precision bf16

The exact measured V7 continuation command, frozen-bank receipt, and autonomous comparison are recorded in docs/retinal-flow-v7-anchor-identity-result.md.

Run the strict fixed-bank evaluation:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python scripts/eval_retinal_flow_lm.py \
  --checkpoint artifacts/retinal_flow_chinese_anchor_identity/checkpoint_latest.pt \
  --bank-size 512 \
  --prototype-views 4 \
  --evaluation-samples 3000 \
  --generation-contexts 192 \
  --out artifacts/retinal_flow_chinese_anchor_identity/fixed_glyph_bank

Generate an autonomous image continuation from typed or image input:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python scripts/infer_retinal_flow_lm.py \
  --checkpoint artifacts/retinal_flow_chinese_anchor_identity/checkpoint_latest.pt \
  --text '天地玄黃，宇宙洪荒。日月盈昃，辰宿列張。' \
  --new-cells 32 \
  --candidate-samples 8 \
  --out artifacts/retinal_flow_chinese_anchor_identity/autonomous_demo

The primary inference artifact is complete_page.png; receipt.json records the model boundary, parameter count, throughput, VRAM, font hashes, every candidate-selection step, and early/late autonomous trajectory summaries. Generated checkpoints and data remain git-ignored.

The earlier whole-page U-Net, latent diffusion, associative-memory, and causal InkStream implementations remain as baselines. They are not the current model.

ILM is a research codebase for language learned and generated as visible writing. Its current experiment predicts continuous retinal states with a causal proposal and hyperspherical flow, then tests visual actuation separately. V18 writes recognizable development forms through a deterministic spatial motor plan. V19 shows that simply adding local retinal features as a residual does not make those features causally responsible for topology. V20 forces fine topology through the local field and verifies local causality. V21 then forces both coarse occupancy and detail through that field and passes every causal and algebraic invariant, but still rejects the writer on simple, medium, and overall fidelity. The next bounded experiments must improve local raster continuity without reopening a global drawing path and must separately learn prompt-image to answer-image state transitions. The writer remains separate from the still-sub-bigram language core. Older structured embeddings, codebooks, and page diffusion experiments remain available as falsified or comparative baselines; they do not define the current model boundary.

The repository intentionally keeps a practical etymology pipeline and long-horizon ILM experimentation side-by-side.

📌 Overview

This repository has three connected tracks:

1. Retinal-flow image-native language modeling and strict held-out evaluation.
2. Historic Chinese glyph etymology ingestion and provenance-preserving assets.
3. Earlier glyph, codebook, diffusion, folio, and InkStream baselines retained for reproducibility.

This README documents all three tracks and keeps the etymology workflow as a first-class, reproducible path.

🔗 Key Links

Area	Path
Conceptual write-up	docs/imagized-language-model.md
Current engineering goal	docs/first-imagized-language-model-goal.md
V41 image-conditioned glyph motor	references/image_conditioned_glyph_motor_bridge_v41_research.md
V39.1 matched trajectory diagnosis	references/visual_answer_trajectory_v39_pilot_diagnosis.md
V41 hash-pinned evidence	publication/ilm-image-native/evidence/v41/
V35 measured result	references/causal_glyph_flow_v35_result.md
V35 implementation and reproduction	docs/causal-glyph-flow-v35.md
V35 preregistered protocol	references/causal_glyph_flow_v35_protocol.md
V34 qualified continuous codec	references/continuous_glyph_codec_v34_result.md
Current paper	publication/ilm-image-native/ilm-image-native.pdf
V31 conditional visual field-flow result	docs/conditional-visual-field-flow-v31-result.md
V31 preregistered protocol	references/conditional_visual_field_flow_v31_protocol.md
V31 research decision	references/conditional_visual_field_flow_v31_research.md
V30 spatial visual next-field result	docs/spatial-visual-next-field-v30-result.md
V29 conditional visual density-ratio result	docs/conditional-visual-density-ratio-v29-result.md
V28 dense visual future-energy result	docs/dense-visual-future-energy-v28-result.md
V21 field-complete writer result	docs/field-complete-writer-v21-result.md
V20 topology-router result	docs/retinal-topology-router-v20-result.md
V19 spatial causal-test result	docs/spatial-retinal-motor-plan-v19-result.md
2026 continuous-sensory research scan	references/continuous_sensory_language_scan_2026.md
V18 visual motor-plan result	docs/visual-motor-plan-v18-result.md
V17 causal actuator result	docs/visual-state-actuator-v17-result.md
V16 predictive visual-field result	docs/predictive-visual-field-v16-memory-result.md
V7 anchor-identity experiment	docs/retinal-flow-v7-anchor-identity-result.md
Closed-loop V6 experiment	docs/retinal-flow-v6-closed-loop-result.md
Research dossier and evidence	references/image-native-language-model-research.md
Archived diffusion plan	docs/ilm-visual-diffusion-code-plan.md
Archived embedding "color" plan	docs/embedding-color-plan.md
Historical development plan	docs/development-plan.md
Etymology module readme	ilm/etymology/README.md

✨ Features

• 🏺 Etymology ingestion from hanziyuan and chineseetymology-style sources.
• 👁️ Continuous foveal retina with recurrent visual context and cross-font invariance.
• ✒️ Deterministic continuous visual motor plan for directly supervised stroke topology.
• 🖼️ Hash-pinned image-conditioned glyph motor bridge with noisy-state robustness audit.
• 🖋️ Conditional pixel-space rectified-flow writer with a differentiable write-read cycle.
• 🔁 Autonomous image-only inference with candidate rereading, energy reranking, and pixel feedback.
• 🧭 Training on exact model-induced visual rollouts with state alignment, next-image energy, and recovery flow.
• 🧪 Fixed 512-character visual-bank evaluation against random, unigram, and bigram baselines.
• 🌐 Robust AJAX + HTML ingestion path with retries, throttling, and cache.
• 🧩 Stage-labeled glyph extraction including <img> and CSS background-image data URIs.
• 🗃️ SQLite-backed storage for chars/glyph metadata plus filesystem asset layout.
• 🖥️ Tornado web UI for ad-hoc ingest + gallery preview.
• 🔤 Glyph rendering utilities for multilingual token images.
• 🧠 Product-code style embedding/codebook modules.
• 🧱 Sentence frame packing and diffusion/inpainting training/evaluation scripts.
• 📊 Reporting and visualization scripts for embedding and pipeline inspection.
• 📄 Publication artifacts in LaTeX/PDF under publication/.

🧱 Project Structure

.
├── README.md
├── AGENTS.md
├── configs/
│   ├── color.yaml
│   └── diffusion.yaml
├── docs/
├── i18n/
├── ilm/
│   ├── code/
│   ├── data/
│   ├── datasets/
│   ├── db/
│   ├── diffusion/
│   ├── encoders/
│   ├── english_tiles/
│   ├── etymology/
│   ├── frames/
│   ├── models/
│   ├── visual_lm/
│   └── utils/
├── scripts/
├── publication/
├── assets/
├── logs/
└── *.ipynb

🧰 Prerequisites

Requirement	Notes
Python 3.10+	Core runtime
pip	Package installation
Optional GPU	Helpful for PyTorch CUDA training scripts
Optional LaTeX toolchain	Needed for publication builds

Assumption note: there is currently no single root dependency lock/spec file (pyproject.toml, requirements.txt, etc.), so dependencies are inferred from imports and script usage.

⚙️ Installation

Minimal (etymology toolkit)

python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install requests beautifulsoup4 tornado

Extended (modeling/training workflows)

python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install requests beautifulsoup4 tornado pyyaml numpy pillow matplotlib torch fonttools

If a specific script needs additional packages, install them from the import error shown by that script.

🚀 Usage

Quick Start: Historic Glyph Ingestion (CLI)

1. Hanziyuan (recommended): char-only AJAX flow

PYTHONPATH=. python scripts/ingest_etymology.py --site hanziyuan --char 中

2. ChineseEtymology (direct URL)

PYTHONPATH=. python scripts/ingest_etymology.py --site chineseetymology --url "https://www.chineseetymology.org/CharacterEtymology.aspx?characterInput=%E4%B8%AD"

3. Batch file ingestion (lines can be char\turl, url, or char url)

PYTHONPATH=. python scripts/ingest_etymology.py --from-file urls.txt

Outputs

Output Type	Location
Files	data/historic/glyphs/<char>/<stage>/<label>.<ext>
Cache	data/historic/cache/*.html
DB	data/historic/etymology.sqlite3

Web Demo (optional)

PYTHONPATH=. python scripts/serve_etymology.py

Open http://127.0.0.1:8888, choose site, enter a character (for example 中).

Polite Crawling and Site Respect

• The fetcher uses per-host throttling, retries with backoff, and caching.
• Keep delays >= 0.5s, avoid bursts, and honor site terms/robots/licensing.
• Do not bypass paywalls or interactive protections.
• If you see 403/429, slow down and retry later.

Additional ILM Workflows

These scripts exist and are actively part of the repo surface, but they are research workflows and may require prepared local datasets/checkpoints.

1. Data download/prep

python scripts/download_alpaca.py --outdir data/raw
python scripts/download_corpora.py --out data/raw
python scripts/sample_paragraphs.py --out data/processed/test_100.jsonl
python scripts/build_images_common_freq.py --out data/processed/images_common_freq --size 128 --en 5000 --zh 5000

2. Glyph DB lifecycle

python scripts/glyphdb_init.py --db data/glyphdb/glyphs.sqlite3
python scripts/glyphdb_ingest_index.py --db data/glyphdb/glyphs.sqlite3 --index data/processed/images_common_freq/index.tsv

3. Code/color model training

python scripts/train_color_codes.py --config configs/color.yaml
python scripts/train_codes_from_qa.py --en-json data/raw/alpaca_en.json --zh-json data/raw/alpaca_zh.json --epochs 1
python scripts/train_ilmglyph_codes.py --en data/raw/alpaca_en.json --zh data/raw/alpaca_zh.json --out artifacts/ilm_glyph_train

4. Diffusion/inpainting

python scripts/train_diffusion.py --config configs/diffusion.yaml
python scripts/train_inpaint_frames.py --ckpt-code artifacts/ilm_glyph_train/ckpt_epoch1.pt --out artifacts/inpaint

5. Evaluation/reporting

python scripts/eval_color_codes.py --checkpoint artifacts/color_codes_e1.pt
python scripts/eval_diffusion.py --checkpoint artifacts/diffusion_unet.pt
python scripts/eval_qa_retrieval.py --checkpoint artifacts/color_codes_qa.pt
python scripts/report_ilmglyph_pipeline.py --ckpt artifacts/ilm_glyph_train/ckpt_epoch1.pt --lang en --text "hello world"

🧩 Configuration

Primary YAML configs:

• configs/color.yaml

  • data path: data/processed/images_common_freq/index.tsv
  • model/code params: d_glyph, d_code, K, C, temperature/anneal
  • optimizer/log settings

• configs/diffusion.yaml

  • input JSONL: data/processed/test_100.jsonl
  • frame/grid + model size settings
  • train mask ratio range and checkpoint settings

Override settings via CLI flags where supported (--epochs, --batch-size, --lr, etc.).

🧪 Examples

• Build a single English tile glyph:

python scripts/build_english_tile_glyph.py "language" artifacts/language_tile --save-tensor

• Run inpainting demo with trained checkpoints:

python scripts/inpaint_demo.py \
  --ckpt-code artifacts/ilm_glyph_train/ckpt_epoch1.pt \
  --ckpt-inpaint artifacts/inpaint/ckpt_epoch1.pt \
  --lang en \
  --text "the quick brown fox jumps" \
  --mode infill \
  --out artifacts/inpaint_demo

• Bulk ingest common characters from Hanziyuan:

PYTHONPATH=. python scripts/bulk_ingest_hanziyuan.py --limit 200 --resume

📝 Development Notes

• This is a research repository with both robust CLIs and exploratory artifacts (including notebooks and prototype scripts).
• Generated large files are intended for data/ and artifacts/ (both ignored in .gitignore).
• Publication source and PDFs are under publication/; helper build script: scripts/latex_build.sh.
• Collaboration/process conventions are documented in AGENTS.md.

🛠️ Troubleshooting

• ModuleNotFoundError: ilm...

  • Run scripts from repo root.
  • Use PYTHONPATH=. for scripts that expect local package resolution.

• FileNotFoundError for data/index/checkpoints

  • Run prerequisite data/build scripts first.
  • Confirm defaults such as data/processed/images_common_freq/index.tsv and data/processed/test_100.jsonl exist.

• CUDA/device issues

  • Switch to CPU with script flags/config (device: cpu or --device cpu).

• Missing package errors

  • Install required dependency from the specific script import path (torch, pyyaml, Pillow, etc.).

• HTTP 403 / 429 while scraping

  • Increase --delay, retry later, and keep requests polite.

🗺️ Roadmap

• Test candidate-conditioned prefix-incremental visual energy inside suffix-collision buckets before authorizing another writer.
• Add fast, line, and page visual states only through measured ablations, starting with the smallest causal state flow.
• Require full visual context to beat last-fixation, unigram, and bigram baselines.
• Require stable, readable 32-cell autonomous continuations before scaling width or corpus size.
• Add multiscale page memory and provenance-gated historical glyph composition only after the causal gate passes.
• Improve environment reproducibility with one authoritative dependency specification and focused tests.

For deeper conceptual and staged planning details, see:

• docs/imagized-language-model.md
• docs/ilm-visual-diffusion-code-plan.md
• docs/development-plan.md

🤝 Contributing

• Follow AGENTS.md for conventions (atomic commits, push after change, no credentials in code).
• Group related edits in focused commits with conventional messages.
• Prefer reproducible script invocations with explicit flags and input paths.
• For scraping-related changes, preserve throttling/cache behavior and site-respect constraints.

❤️ Support

Donate	PayPal	Stripe

📄 License

No top-level license file is currently present in this repository.

Assumption note: treat the project as research code with unspecified licensing until a LICENSE file is added by maintainers.