1
Source inventory
63 language PDFs across 32 newsletter dates were inventoried and given first-pass browsable outputs.
Proven
Current position
The extraction foundation exists, but the web output is not ready for submission.
The work has taken considerably longer than expected. We completed substantial extraction and validation work, but the generated pages have not been consistently good enough in human visual review.
Why the approach changed
The first approach was a fixed-template builder. Roughly a month went into building and refining that pipeline. It produced some good individual pages, but the quality did not carry consistently across newsletters. Following a project meeting about three weeks ago, we changed direction and began testing whether an LLM could generate the design from verified content instead.
What we are testing nowA design model receives a smaller verified content package—text, images, links and relationships—without the old HTML, coordinates or layout instructions. Fable and Sol produced more coherent results from this package than the earlier December attempts.
What is still unknownNeither candidate yet passes the mobile containment gate, and the method has been tested on only one difficult English edition.
63source PDFs inventoried and generated in first-pass form
56image-based PDFs given a semantic OCR pass
34approximately edition-specific renderers accumulated in the earlier builder
Not yetfleet-level visual approval or publication readiness
Clarification of the earlier update
The earlier message said that the extraction foundation was complete but the design output was not ready. That remains the position. The latest experiment is more encouraging than the previous December attempts, but it has not passed mobile or multi-edition testing.
“Extraction foundation complete” means that the 63 packages, text and image artifacts, source annotations, URLs and mappings exist and have passed substantial structural checks. It does not mean that every image crop, visual grouping or rendered link association has been approved by a person.
01 · Evidence ladder
What exists at each stage
The project is not one undifferentiated “PDF-to-HTML” result. This ladder separates the parts that are proven, partial, rejected or still pending.
2
Extracted corpus
All 63 packages passed the recorded package-integrity check. The corpus retains structured text, image artifacts, reading order and article boundaries.
Proven foundation
3
Links and mappings
The deterministic parity report covers 864 source annotations: 56 existing article links and 808 visual hotspots, with no missing annotations. Cross-language article coverage reported no failures across 31 bilingual editions.
Proven structurally
4
Deterministic web design
Some editions looked strong, but the renderer accumulated edition-specific code and did not produce consistent visual quality on difficult newsletters.
Partial
5
Human visual review
Review of November and December exposed bad crops, text over faces, detached or unexplained links, inconsistent card sizing, excessive whitespace and weak mobile composition.
Not approved
6
Automatic visual approval
The checker detected many real problems but also flagged every clean control package. It can assist investigation, but cannot approve publication.
Rejected as gate
7
Lean-corpus model design
Fable and Sol independently produced more coherent difficult-edition designs from the same verified, design-neutral content package. Mobile containment still fails.
Under evaluation
8
Fleet publication
A varied English/Hindi sample and human acceptance must come before any claim that the method scales to all 63.
Pending
02 · Work to date
How the approach changed
Each change responded to a specific limitation found in the preceding approach.
June 2026
We built the complete inventory and first-pass archive
All 63 PDFs were downloaded and converted into browsable HTML wrappers. The process copied 237 rendered page assets. This proved completeness of the source inventory, but 56 PDFs had no usable embedded text.
Inventory completeJune–July
We added OCR and structured extraction
Gemini OCR produced semantic pages for all 56 image-based PDFs, using rendered images when direct PDF reading failed. Multiple extraction and validation passes improved text and URL reliability. The resulting corpus contains text, images, links and their intended relationships.
However, OCR HTML was not visually faithful, and image crops and ownership still needed human review.
Extraction usable · layout incompleteJuly–August · roughly one month
We first worked towards a fixed-template pipeline
The initial plan was to extract the content and place it into a deterministic set of reusable layouts. We spent roughly a month building and refining that pipeline. It produced some convincing editions, particularly the June 26 reference.
However, the quality did not generalise. The builder accumulated roughly 34 edition-specific renderers and dozens of dispatch branches, and difficult newsletters still required special handling. This showed that a fixed template alone was not giving us a consistent result.
Useful foundation · inconsistent designAround mid-August · about three weeks ago
We agreed to test LLM-generated design
In the subsequent project meeting, we decided to separate verified content from visual design. Instead of requiring the fixed builder to decide every layout, the proposed approach was to give an LLM the extracted text, images, links and relationships, and allow it to generate the presentation.
This was a change in design strategy, not a restart of the extraction work. The existing corpus and validation work remain the factual input and checking layer.
Direction changed after reviewAugust
We over-relied on technical validation
Automated checks verified that expected text, assets and links were present. Several outputs passed those checks. Human reviewers still found cropped faces, captions covering subjects, missing or detached links, excessive whitespace, weak typography and poor mobile composition.
We repeatedly treated “all expected pieces are present” as if it meant “the design is correct.” It does not.
Technical pass ≠ visual approval27 August
We tested an AI visual-QA checker
The checker caught 14 of 17 supported known defects, but it also flagged all 13 clean control packages. A tool that raises alarms on every clean example cannot reliably replace a human reviewer.
Automation rejected as approval gateLate August–September
We separated reusable code from client data
Portable schemas, exporters, link-affordance rules, asset-authority rules, validators and a reference renderer were moved into the RG repository. This preserves the valuable engineering while keeping protected PDFs and review evidence outside the codebase.
Reusable foundation retained3–4 September
We tested a lean, design-neutral content package
For the difficult December 31 English edition, we removed old HTML, CSS, coordinates, crop coordinates, QA logs and renderer instructions from the model input. Fable 5.1, GPT-5.6 Sol and Gemini received the same source-authoritative package independently.
More encouraging · not yet proven03 · What we learned
What is reusable—and what did not work
Work that remains valuable
- A complete 63-PDF source inventory and first-pass archive.
- OCR and deterministic extraction for text and inline URLs.
- Structured content IDs, reading order, asset ownership and link ownership.
- Approved source assets and source-PDF traceability.
- Automated checks for omissions, duplicates, broken links and invented content.
- Portable schemas, exporters and validation code now owned by RG.
Approaches that did not converge
- Pixel-matching the typography and spacing of a hand-designed PDF in responsive HTML.
- Adding newsletter-specific renderer code until each edition looked acceptable.
- Assuming technical validation could judge visual quality.
- Giving models rich JSON containing old HTML, coordinates and prior layout choices.
- Letting a model redesign link affordances without deterministic ownership checks.
- Scaling before a difficult, unseen edition passed both machine and human gates.
The clearest correction: the extraction layer and the design layer are different problems. We should preserve verified facts deterministically, then give controlled creative freedom only to presentation. A model may choose layout, typography and rhythm. It may not change text, assets, URLs, ownership or reading order.
04 · Design evidence
The results that changed our view
These are representative artifacts, not a claim that every edition has reached the same standard.
A strong deterministic reference
The June 26 output showed that high-quality newsletter HTML was possible. Later investigation showed that its quality depended heavily on edition-specific construction, so it could not be assumed to generalise.
A technically organised but visually inconsistent template
The December template assembled the correct kinds of content and adapted to mobile, but human review found problems deeper in the page: inconsistent image treatment, misplaced link affordances, oversized sections and poor spacing. This is the clearest example of why first-viewport appearance and structural checks were insufficient.
Visual-QA trial
The automatic reviewer was not reliable enough to approve pages
14/17known supported defects detected
13/13clean packages incorrectly flagged
The formal result was “untrusted—manual review required.” The score remains frozen; it was not tuned against the known answers.
Human review
What reviewers found
Correct text and valid URLs could still be presented poorly: captions covered faces, snapshot images were cropped inconsistently, links became detached labels, thumbnails lost their click targets, and large blank gaps interrupted the reading flow. We therefore keep technical checks and human visual approval as separate stages.
The underlying review submissions remain protected rather than published on this public evidence site.
05 · Current experiment
One verified content package tested with three design models
The December 31 English newsletter was selected because it contains varied story types, many linked media items, complex imagery and a long page. Earlier approaches had struggled with it.
51ordered text blocks
26approved newsletter assets
21contextual content links
5shared shell links
What the models did not receive: previous HTML, CSS, layout templates, PDF coordinates, crop coordinates, SEO suggestions, QA history, model history or old renderer instructions. This reduced inherited design bias while keeping the factual content locked.
Claude Fable 5.1
Passed the strengthened content contract. Its layout is more coherent than the earlier deterministic December attempt. It still clips content at a 390-pixel mobile width.
GPT-5.6 Sol
Independently passed the same content contract and produced another usable design direction. It also has right-edge clipping at the 390-pixel mobile width.
Gemini corrective attempt
The corrected candidate now preserves the complete content and removes the duplicated PDF link. It still fails mobile containment. It is also excluded from model comparison because the CLI reported a different underlying model from the one requested.
Same source for every arm
Inspect the input, not just the output
The source PDF remains available beside a lean JSON representation. The JSON carries verified content and relationships, but deliberately avoids dictating the design.
| Model | Exact content | Links/assets | Desktop direction | 390px mobile | Current verdict |
|---|---|---|---|---|---|
| Claude Fable 5.1 | Pass | Pass | Worth further review | Fail | Retain for the next test |
| GPT-5.6 Sol | Pass | Pass | Worth further review | Fail | Retain for the next test |
| Gemini corrective | Pass | Pass | Usable | Fail | Exclude: provenance mismatch |
06 · Assessment
What the current experiment tells us
What is encouraging
We now have a cleaner separation between factual content and visual design. Two independent models made a difficult newsletter more coherent from the same lean source package. This is more useful than continuing to repair individual pages without a reusable method.
What remains uncertain
No candidate has passed every check. The method has not been tested across Hindi, different content shapes or a pre-selected mixed sample. One improved edition is not evidence that all 63 can be completed this way.
Current assessment: this approach is worth testing further because the remaining questions are narrower and easier to measure. It is still an experiment and should not be scaled until a small, varied sample has been reviewed.
07 · Recommendation
Suggested next step: a limited validation exercise
We suggest testing the lean-corpus approach on a small mixed sample rather than returning to PDF pixel-matching or continuing to add newsletter-specific renderer code.
- Freeze a small shared design system: responsive masthead, typography scale, story block, linked-media card, snapshot grid and footer.
- Repair mobile behaviour generically: no edition-specific CSS; every candidate must fit 390 pixels without clipping or hidden overflow.
- Apply the same pipeline to a varied six-edition sample: English and Hindi, simple and complex, selected before generation.
- Keep correctness deterministic: exact text, image, link, relationship and reading-order validation remains outside the design model.
- Require human visual approval: crop, link ownership, readability, hierarchy, spacing and overall finish are judged separately from technical checks.
Scale decision: proceed to the full fleet only if the sample has zero content/link loss, zero horizontal overflow, high reuse of shared components, no newsletter-specific renderer code, and human acceptance with only bounded adjustments. If it does not, we should tell Shashank that the approach has not proved scalable and agree a revised scope rather than continuing open-ended iteration.
08 · Decision
Proposed decision
Run the limited six-edition exercise above, then review the evidence before deciding whether to continue with the remaining newsletters.
This would not authorise fleet publication or imply that all 63 editions can already be completed reliably. If the sample does not meet the stated checks, we should revise the scope rather than continue open-ended iteration.
If the project is paused instead, the extracted corpus, source archive, link mappings, validation code and current evidence can still be handed over as useful outputs.