# Why the tweet replicas fail The saved observatory examples are failed visual reproduction attempts. The earlier claim of layout success was overstated: the checks measured compliance with our character-count templates, which encode the wrong target. - [Geometry is the main failure](#geometry), confirmed by measurements of both source images and both saved renders. - [Vocabulary restrictions](#vocabulary) exclude important topic words but admit fragments. Their total contribution to poor grammar is unmeasured. - [A capitalization bug](#capitalization) penalises valid initial capitals, reproduced with a real first-step model probe. - [The suspected backend shift error was ruled out](#backend) by reference-code review. This diagnosis leaves the generated examples and sampler behaviour unchanged. ## Geometry The first source image has nearly equal interior word widths within each row, decreasing from roughly 84 pixels near the top to 28 near the bottom. Its lines share a centre. The second has thirteen stable physical word columns whose widths increase from left to right. These are visual width patterns, not character-count patterns. For example, the second image's fourth column contains a five-character word and a three-character word that both measure 34 pixels of ink. Our corresponding words measure 45 and 27 pixels. Matching character counts cannot preserve that column. | Necessary visual property | Source | Saved output | | --- | ---: | ---: | | Second image: row-width range / median width | 0.83% | 10.90% | | First image: line-centre standard deviation / median width | 0.13% | 1.60% | | Second image: word-start standard deviation across rows | 0.8–1.5 px | Up to 19.2 px | The first two ratios are unchanged by uniform scaling. This is therefore more than a font-size mismatch. A proportional font alone would also leave word selection governed by the wrong constraint. [tweet_profiles.py](../src/intaglio/tweet_profiles.py) turns each word into a fixed character count, adds one space between words, and uses estimated whole-cell indents. [render.py](../src/intaglio/render.py) uses fixed-width cells and a line `textLength` based on character count. Neither enforces the source's physical column widths or common centre. The separate meander patterns are designed cell layouts, not replicas of either source image. ### Measurement and reproduction The [numeric snapshot](../experiments/replica_geometry.csv) contains all 72 measured rows: 18 rows from each of two originals and two unannotated generated PNGs. Image URLs and measurement details are in the [audit script](../experiments/audit_replica_fidelity.py). No original prose or images are stored in the snapshot. Measurements threshold greyscale ink below 128, separate the 18 line bands, and find word gaps using each row's largest `word_count - 1` empty-column runs. These are ink bounding boxes, not recovered font advances; the source font is unknown. Thresholds 96 and 160 retained the verdict: source row-width spread stayed at or below 1%, while the generated second image remained above 10.7%. ```sh python3 experiments/audit_replica_fidelity.py ``` This deliberately exits **1** for the existing snapshot: both originals pass and both saved outputs fail. The chosen thresholds (2% row-width spread, 0.5% centre jitter) separate the observed examples; passing would not prove complete fidelity. This is a snapshot audit, not a live-render test or part of the passing unit suite. New renders require new measurements. ## Vocabulary The profiles limit words to nine characters in the first image and seven in the second. Consequently, `observatory` (11) and `astronomer` (10) cannot fit either; `telescope` (9) and `astronomy` (9) cannot fit the second. A tokenizer/table audit also found no allowed whole-token spelling for `observatory` or `astronomer`: they split across tokens despite passing the dictionary predicate. Meanwhile, [vocab.py](../src/intaglio/vocab.py) admits standalone `ing` and the separate possessive suffix `'s`. Ink cells permit apostrophes, allowing the observed `ing's`. A dictionary filter on individual tokens is not final-word validation. These mechanical restrictions are confirmed. No ablation has quantified the separate effects of model size, vocabulary, sampling schedule, penalties, or candidate selection. ## Capitalization In [sampler.py](../src/intaglio/sampler.py), `_stray_capitals` initialises `follows_a_stop` to false and only updates positions after position zero. A valid initial capital therefore receives the default four-nat stray-capital penalty. At temperature 0.7, this changes otherwise equal odds by about 303 times. A first-step probe using the actual `tweet_wave` composition and existing 0.6B model compared ` My` with ` my`, which occupy the same character count. The raw model favoured the capital by about 6:1; after the sampler's belief adjustments, lowercase was favoured by about 23:1. This confirms a decoding defect; its effect on a complete paragraph has not been measured. The defect remains unfixed in this diagnostic change. ## Backend The small checkpoint's [pinned reference sampler][card] reads logits and writes predictions at the same positions. Its [custom model][model] changes the attention mask to full visibility, without shifting the language-model head. The corresponding upstream [trainer][trainer] and [sampler][sampler] default to `right_shift_logits=False`. All 311 checkpoint tensor names and shapes match the stock Qwen3 architecture. These checks support the current same-position interpretation. Applying the Dream adapter's one-position shift to this checkpoint would introduce an error. They do not prove our decoding settings match the reference's quality. The prepared Dream 7B backend has not generated these examples; Vast rental was blocked by account credit. ## What a repair requires The visual target needs font-aware word advances, explicit per-row/per-column width constraints, and centring where the source requires it. Acceptance checks must compare rendered geometry against source invariants independently of the generation template. The vocabulary needs useful multi-token words and final-lexeme validation; the initial-capital penalty needs its boundary condition corrected. Controlled generations can then establish how much a larger model improves fluency. More parameters cannot repair a geometry objective that measures the wrong thing. No paid compute was used for this diagnosis. [card]: https://huggingface.co/dllm-hub/Qwen3-0.6B-diffusion-mdlm-v0.1/blob/c8d24a3f4adaeef46881b450e1bf7d1005203bd7/README.md#L101 [model]: https://huggingface.co/dllm-hub/Qwen3-0.6B-diffusion-mdlm-v0.1/blob/c8d24a3f4adaeef46881b450e1bf7d1005203bd7/modeling_qwen3.py#L82 [trainer]: https://github.com/ZHZisZZ/dllm/blob/ca176752fbceec49c6b4777a2c18ae88e4eb10ed/dllm/core/trainers/mdlm.py#L30 [sampler]: https://github.com/ZHZisZZ/dllm/blob/ca176752fbceec49c6b4777a2c18ae88e4eb10ed/dllm/core/samplers/mdlm.py#L31