Reproduction tutorial · September 2, 2026
Protein design validation protocol and computational smoke test
These tutorials test the mechanism behind each paper at a practical scale. They do not reproduce frontier-scale headline results unless explicitly stated.
Source of this tutorial
This tutorial tests the mechanism behind reviewed development 03 in the immutable briefing of September 2, 2026.
Scope and limits
- Reproduction level
- protocol and computational smoke test only
- Claim under test
- whether a documented, blinded design-and-selection funnel with an open baseline can be executed end to end and reported completely, including rejects.
- Excluded claim
- Anthropic's sequences, protocols, design funnel, negative results, and kernel code are missing at review time, so the headline hit rate and affinity results cannot be reproduced here. This tutorial does not attempt them.
- Reader, effort, and cost
- advanced computational biology; 8 to 16 hours for the computational smoke test; wet-lab work is out of scope for an individual reader.
- Prerequisites
- Pinned repositories, weights, and sequence databases with recorded licenses and commits. Institutional biosafety review before any physical work.
Canonical resources
- Anthropic, Claude Fable and Mythos 5.1: https://www.anthropic.com/claude-fable-and-mythos-5-1 (opens in a new tab)
- Artifact availability notice: No complete sequences, protocols, design funnel, negative results, matched compute budget, or kernel code were available at review time.
Project plan
| Step | Stage | Effort |
|---|---|---|
| 01 | Pin the environment | 90 minutes |
| 02 | Equal design budgets | 120 minutes |
| 03 | Blinded selection | 90 minutes |
| 04 | Seeds and complete reporting | 120 minutes |
| 05 | Wet-lab path, partner only | external |
| 06 | Cleanup and disclosure | 30 minutes |
01Pin the environment(90 minutes)
Choose one public target and one open design baseline. Pin repositories, weights, and databases by commit and checksum, and record every license.
02Equal design budgets(120 minutes)
Give each design method the same compute budget, the same number of candidates, and the same filtering stack. Record the full funnel from generated candidates to final selections.
03Blinded selection(90 minutes)
Select final candidates without knowing which method produced them. Freeze the selection rule before scoring.
04Seeds and complete reporting(120 minutes)
Run five seeds. Report every reject alongside every accepted candidate, plus the complete funnel counts per stage.
05Wet-lab path, partner only(external)
Any physical validation proceeds only through a qualified partner laboratory, with a preregistered assay, positive and negative controls, replicates, institutional safety approval, and sequence screening. Individual readers should stop at the computational smoke test.
06Cleanup and disclosure(30 minutes)
Publish pinned versions, seeds, funnel counts, and rejects. State explicitly which parts of the vendor claim remain untested.
Controls, failure modes, and safety
Controlled variables
target, baseline method, compute budget, candidate count, filtering stack, selection rule, and seeds.
Failure modes
unreported rejects, selection after scoring, unmatched budgets, database or target leakage, and treating computational scores as experimental evidence.
Safety
computational work only for individual readers. Wet-lab work requires a qualified partner, institutional biosafety approval, and sequence screening. Do not pursue any hazardous target.