# Day 003 — Immunoglobulin Locus Explorer

Part of the **[BioML Build-Off](https://github.com/devdootlab)** — a 300-day public benchmark
where three AI coding agents each build the same small biology tool from one identical
one-sentence brief, and we compare the results with a shared, pre-registered test harness.

> **This is a pre-test gallery.** The three builds are in; the head-to-head scoring hasn't run
> yet. This repo is shared so the group can try all three side by side first. The frozen judging
> plan is in **[TEST-STRATEGY.md](TEST-STRATEGY.md)**.

**▶ Live site (this is where the apps run):
[devdootlab.github.io/d003_SloppyIgLocus](https://devdootlab.github.io/d003_SloppyIgLocus/)**

> The apps only *run* on the GitHub **Pages** domain above. Links viewed here on **github.com**
> open the *source code* — that's normal for a repo. Use the Pages link to actually use the tools.

---

## The brief (identical for all three)

> A standalone page giving a pleasant overview of the immunoglobulin locus in **mouse and human**
> (click to toggle), that zooms **all the way down to codons and nucleotides**; **homologous V
> regions coloured alike**; **easy zoom in/out so the user doesn't get lost**; and a
> **guided-but-escapable** mode to pick a **V/D/J (heavy)** or **V/J (light)** and see **how it
> comes together**, allowing **N- and P-nucleotide addition**. The user is a biology grad student.

Full build prompt: [`prompts/prompt.md`](prompts/prompt.md).

---

## The three builds

Each agent worked in isolation and chose its own design. Reader-facing name is the **model**;
`claude` / `agy` / `codex` are internal folder ids (directories, JSON keys) only.

| Build | Model | Runs in | ▶ Open the live app |
|---|---|---|---|
| **Claude** | Anthropic **Claude Opus 4.8** | Claude Code | [Open →](https://devdootlab.github.io/d003_SloppyIgLocus/sites/claude/index.html) |
| **Agy** | Google **Gemini 3.5 Flash (High)** | Antigravity | [Open →](https://devdootlab.github.io/d003_SloppyIgLocus/sites/agy/index.html) |
| **Codex** | OpenAI **GPT-5.5 (Medium)** | Codex CLI | [Open →](https://devdootlab.github.io/d003_SloppyIgLocus/sites/codex/index.html) |

The **Open →** links go to the live Pages apps. (The source for each is under `sites/<id>/` in
this repo.) All three are static and run entirely in the browser. Each build's notes are in its
`stats.md`.

---

## Fairness note (read before you rank them yourself)

Claude built one of the three tools **and** drafted the test strategy — a self-interested
position. Per the build-off rules, the rubric is **frozen and signed off by a non-competitor
before any tool is scored**, correctness is separated from presentation, and a missing feature
counts as a coverage gap rather than being excused. A top score for Claude's own tool is treated
as a flag to scrutinise, not a result to celebrate. Details: [TEST-STRATEGY.md](TEST-STRATEGY.md).

## What to look at

These tools solve a real gap: IGV/UCSC don't carry a consistently annotated immunoglobulin
locus, so it's hard to see where the isotypes and V families sit. Things worth comparing across
the three:

- **Do the loci read correctly?** V→(D)→J→C order, D only in the heavy chain, the right
  constant-region isotypes (IgM/IgD/IgG…/IgA/IgE), and homologous V families coloured alike
  across mouse ↔ human.
- **How far down does it zoom, and does it stay legible?** All the way to nucleotides and codons
  without losing you.
- **Does the V(D)J assembly make biological sense?** The conserved Cys→Trp/Phe CDR3 anchors,
  the 12/23 rule, P-nucleotides that are actually palindromic, and an honest productive /
  non-productive verdict as you add N-nucleotides.

## Disclaimer

Teaching tools, not validated for clinical or research use. Sequences and spacing are curated /
schematic; verify against IMGT before relying on any specific number.

---

*Naming, layout, and push conventions follow the earlier days (`dNNN_…` repo, `sites/<id>/`
builds, model-named cards).* MIT licensed — see [LICENSE](LICENSE).