Context Window Atlas methodology

Version 2026-09-07 · Primary-source ledger: JSON · The three named diagrams are licensed by Kingy.ai under CC BY 4.0.

# Kingy.ai Context Atlas — What Fits in an AI Context Window?

This local pilot combines an explorable context-budget calculator with a reusable visual asset set. It is a static, dependency-free package: the page, interaction, editable SVGs, PNGs, and data exports all work from this folder.

## Preview locally

From this directory, run:

```bash
python3 -m http.server 4182
```

Then open `http://127.0.0.1:4182/`.

## What the calculator does

The calculator estimates whether a source artifact and the rest of a request fit within a chosen context limit:

```text
artifact tokens = artifact quantity × planning tokens per unit
total tokens = artifact + instructions + tools + history + retrieval + answer reserve
midpoint headroom = context limit − total tokens
```

The artifact estimate carries an editable uncertainty band. A result becomes “Fit is uncertain” when that range crosses the context limit. “Fits, with little room” begins above 80% midpoint utilization. These states are planning aids, not provider guarantees.

Presets cover an 80-page PDF, a repository slice, a two-hour transcript, a 40-turn agent run, and a 100,000-word manuscript. Every preset is editable. The “Known token count” mode removes estimation uncertainty when the user already has an exact count from the target tokenizer.

## Planning defaults

| Artifact type | Default midpoint | Default uncertainty | Basis |
| --- | ---: | ---: | --- |
| PDF/report | 667 tokens/page | ±25% | Editorial planning assumption; page density varies widely |
| Prose/manuscript | 1.333 tokens/word | ±20% | Derived from OpenAI's English rule of thumb |
| Source code | 10 tokens/line | ±40% | Editorial planning assumption; language and formatting vary widely |
| Transcript | 200 tokens/minute | ±20% | Editorial planning assumption for conversational English |
| Agent run | 2,500 tokens/turn | ±50% | Editorial planning assumption including variable tool traffic |
| Known tokens | 1 token/token | 0% | Direct user input |

Use the target model's tokenizer for exact text counts. Multilingual text, code, tables, OCR, images, document parsing, and provider-specific message wrappers can differ materially from these defaults.

## Claim verification

Durable claims are separated from Kingy's editable planning assumptions in [`sources.json`](sources.json). The ledger was checked on 2026-09-07 against:

- [Anthropic context-window documentation](https://platform.claude.com/docs/en/build-with-claude/context-windows) for what occupies context, output budgeting, and long-context quality cautions.
- [OpenAI token-counting guidance](https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them) for the English rule of thumb and model-specific variation.
- [Hugging Face Tokenizers API](https://huggingface.co/docs/tokenizers/main/api/tokenizer) for the tokenizer pipeline and why different normalization and segmentation choices produce different counts.

The pilot intentionally does not list current model limits. Those values change often and would turn a durable explainer into a maintenance-heavy vendor table.

## Reusable assets

Each visual is authored as editable SVG and exported as a 2400×1350 PNG:

| Asset | Editable source | Raster export | Best use |
| --- | --- | --- | --- |
| Shared context anatomy | `assets/context-window-anatomy.svg` | `assets/png/context-window-anatomy.png` | Article lead, capacity explanation |
| What fits at each scale | `assets/what-fits-scale.svg` | `assets/png/what-fits-scale.png` | Social card, comparison section |
| Context compaction | `assets/context-compaction.svg` | `assets/png/context-compaction.png` | Agent-memory explanation |

The SVGs use system fonts, semantic titles/descriptions, and direct text and shape elements. They can be edited in Figma, Illustrator, Inkscape, or a text editor. The assets are cleared for Kingy.ai's own editorial and promotional reuse. No public redistribution license is assigned by this local pilot.

## Data exports

“Download CSV” and “Download JSON” serialize the current calculator state, including the formula, assumptions, midpoint and range, each context occupant, and links back to this methodology and the claim ledger. Filenames are `kingy-context-budget.csv` and `kingy-context-budget.json`.

## Package map

- `index.html` — semantic page structure and verified explanatory copy
- `styles.css`, `assets.css` — responsive Kingy visual system
- `app.js` — presets, calculation, live diagram, and exports
- `sources.json` — primary-source claim ledger and editorial assumptions
- `assets/*.svg` — editable visual source
- `assets/png/*.png` — publication-ready raster exports
- `VALIDATION.md` — correctness, usability, and presentation checks

No build step, package manager, analytics, network request, account, or paid service is required.

## Production effort and scale estimate

This pilot required one focused research/build pass, one browser-verification pass, and one targeted responsive correction. The resulting 1.7 MB package contains the interactive explainer, five presets, three source-audited claims, three SVG/PNG asset pairs, two generated data examples, and desktop/mobile QA captures.

With this design and calculator pattern reused, a nearby topic should take roughly 45–90 minutes of agent production plus 10–20 minutes of human editorial and brand review. That is a planning range, not a measured service-level promise. Topics with disputed definitions, current vendor comparisons, safety implications, or original datasets will cost materially more because research and verification dominate the layout work.