Spec. 0004 — ongoing
Everything the agent left behind
It proposes peptides nobody asked for, then spends the rest of the run trying to kill them. Eight gates, each with a verdict and a confidence, each one able to end the candidate. What survives is published — and so is what didn't.
Candidate
RSD-4820
Scanning
ALA · 01/26
Fold ΔG
−7.2 kcal/mol
- Runs on record
- 9
- Cleared all eight
- 6
- Liabilities flagged
- 22
- Reasoning withheld
- None
Eight ways to be wrong
A candidate is not scored, it is interrogated. Every gate is a place the run can end, and the gate that ends it says so in plain language rather than dropping a number.
- 01
Target definition
Receptor, pathway, therapeutic rationale
Ends onA mechanism nobody can name
- 02
Sequence design
Residues, modifications, analogue lineage
Ends onMotifs borrowed without reason
- 03
Structure & conformation
Secondary structure, folding, cyclisation
Ends onA fold that will not hold
- 04
Binding affinity
Predicted potency and selectivity
Ends onPotency bought with selectivity
- 05
Stability & half-life
Protease exposure, plasma persistence
Ends onMinutes of plasma survival
- 06
Off-target risk
Cross-reactivity and safety flags
Ends onReceptors hit by accident
- 07
Synthesis feasibility
SPPS route, yield, purification burden
Ends onRoutes no lab would run
- 08
Wet-lab handoff
Assays to run, success criteria
Ends onA plan with no falsifier
How a run actually goes
There is no ranking model and no hidden score. The agent reasons in text, in order, and the text is the product.
01
A brief goes in
One or two sentences describing a problem worth solving — a liability to engineer around, a receptor to reach, a half-life to beat. No structure required.
02
The agent traverses
It works the gates in order and cannot skip. Later gates read the earlier ones, so a weak sequence design is still hanging over the run by the time synthesis is assessed.
03
Verdicts land in the open
Pass, flag or fail, each with a stated confidence and the reasoning that produced it. Text streams as it is written — nothing is assembled behind a spinner.
04
The run is written down
Every completed traverse is appended to a shared journal with its verdicts, model and timing. Runs are not curated after the fact.
The most recent traverses
Exactly as they landed. Anyone can start a run, and everyone sees the same log.
- 2d agoclearedanthropic/claude-sonnet-4.5
A GLP-1 analogue with reduced nausea liability at equivalent weight-loss efficacy
GLP-1R (glucagon-like peptide-1 receptor) is well-validated as a weight-loss target, with semaglutide and tirzepatide demonstrating 15-20% body weight reduction in phase 3 trials. The receptor is a class B GPCR expressed in pancreatic beta cells, CNS appetite centers (hypothalamus, area postrema), and critically the br
5 pass · 3 flag · 0 fail82.1s - 2d agoclearedanthropic/claude-sonnet-4.5
A GLP-1 analogue with reduced nausea liability at equivalent weight-loss efficacy
Reducing area postrema engagement while preserving hypothalamic signalling is the central tension. Biased agonism at GLP-1R toward Gs over beta-arrestin is the most promising handle, though the separation achieved in published analogues is modest.
6 pass · 2 flag · 0 fail38.4s - 2d agoheldanthropic/claude-sonnet-4.5
A BPC-157 variant with improved plasma stability for systemic tendon repair
The pentadecapeptide has no defined tertiary structure, which is precisely why stabilising it is hard — there is no fold to rigidify. N-terminal acetylation and D-amino acid substitution at positions 4 and 11 are the obvious first passes.
4 pass · 3 flag · 1 fail33.1s - 3d agoclearedopenai/gpt-4o-mini
Evaluate MOTS-C 10mg as a research candidate.
MOTS-c is encoded in the mitochondrial 12S rRNA region rather than the nucleus, which makes it unusual among signalling peptides. Reported AMPK activation is consistent across rodent work but human data remains thin.
6 pass · 2 flag · 0 fail21.8s
The failures are the point
A discovery process you only see the wins from tells you nothing about how it decides. The whole reason the trail is public is so the bad calls are countable.
The failed runs stay up
A journal that only shows candidates which cleared is a brochure. Every traverse is appended the moment it completes, including the ones where gate three ended it, and nothing is deleted afterwards.
0
runs removed
Confidence is stated, not implied
Each verdict carries a number the agent had to commit to. Low confidence on a pass is a different object from high confidence on a pass, and the journal keeps both.
8×
verdicts per run
This is triage, not evidence
The agent has no lab. It runs no simulations and reasons only over what it was trained on, which means it will state wrong numbers with total composure. Treat every figure as a hypothesis with an experiment attached.
0
assays performed
What made it out
Compounds that cleared the stack and have a certificate attached. The index is the residue of the research, not the reason for it.
The things people ask
Still unsure? Write to the lab and a person will answer.
The console is open
Give it something hard
Describe a liability worth engineering around and watch the agent try to talk itself out of the answer.