Evidence-grounded research guidance + verifiable dry-lab design artifacts. Proposes hypotheses and pre-experiment designs — NOT an expression/titer prediction or a guarantee. Validate at the bench.
From transcript to artifact
A chat answer scrolls away. A dossier is an artifact: it survives the session, and it can be checked line by line. The Kairos workbench now lays the conversation (30%) beside a structured display (70%) — the backend’s dossier JSON (evidence · design · reasoning · review · claims_report) renders into four panels instead of being flattened into one Markdown stream.
Each panel answers exactly one question a reviewer would ask:
- Evidence — what is this based on, and can I open the source?
- Design — what should I order, and can I re-run the ranking?
- Structure — does the protein fold as expected, can I see it?
- Audit — did the system check its own claims, and can I verify the check?
The run behind this dossier
Question: How do I express cytochrome c in Pichia — secreted? Target: UniProt Q6Q4H8, a native Komagataella phaffii protein, 110 aa. This is the same run published as our cytochrome-c worked example — every number below can be cross-checked there.
Panel 1 · Evidence — sources you can open
Retrieval runs live against the Kairos knowledge layer; each hit lands as a card with a type tag, title, abstract snippet, host / target / year metadata, and a PubMed link. Four cards from this dossier:
- PMID:30587177 — AOX1 promoter regulation: methanol-regulated PAOX1 (Mxr1 feedback circuit) — grounds the expression strategy.
- PMID:37888283 — nitrogen source & MUT pathway: ammonium sulfate raises MUT / PEX gene expression.
- PMID:38650288 — fed-batch rhythm: glycerol for biomass, then methanol induction — the standard Pichia fermentation pattern.
- PMID:31590267 — proteolysis risk: recombinant protein degradation in P. pastoris — watch degradation, consider inhibitors.
None of this is decorative. Every card links to a real PubMed record that you can open and hold against the claim it grounds — and in the workbench, cards flagged by the reviewer show it: WARN / FAIL cards carry amber / red borders.
Panel 2 · Design — a panel you can re-run
Twelve synonymous CDS candidates, Pareto-ranked, seed 42 — the run is deterministic and reproducible. Each candidate is scored on four independent axes; the top of the table: uniform (rank 1), host_weighted (2), gc_balanced (3).
The winner’s scorecard: sequence validity strong · clean; manufacturability strong · clean; host-likeness moderate · clean; developability strong · 1 flag — extreme_pI, an intrinsic property of the target protein, not a design defect.
Every candidate is synonymous — it translates to the exact input protein, verified by translation. Ranking is multi-objective (Pareto): multiple candidates, never one opaque “best”.
One click copies the CDS to the clipboard, ready to paste into a synthesis order (Twist, IDT…). The panel is designed to end in an order form, not in a screenshot.
Panel 3 · Structure — folding, checked in silico
The structure panel looks the target up in AlphaFold DB and screens folding confidence with Boltz-2; the 3D viewer carries per-residue pLDDT so you can see which regions are confident and which are not. When no structure data exists for a run, the panel says so — “structure check not run” — instead of drawing you something reassuring.
Panel 4 · Audit — the system checks itself
Since v0.5.0 the deterministic review layer runs on the public path, and its output is the fourth panel: a verdict badge (PASS / WARN / FAIL) over a claim table — Claim | Type | Source | Verdict. The reviewer mechanically checks citation authenticity, numeric consistency, and claim grounding; a next-generation reviewer already runs beside it in non-mutating shadow mode. How the review layer works →
A PASS looks like this (live output from the published signal-peptide run):
## Review (automated citation/number audit — verdict: PASS)
- ✅ Citations traced to evidence: PMID:29601964, PMID:41660601But honesty matters more than badges. In a real HSA session, a follow-up turn — “And for secretion instead?” — retrieved 22 domain sources, yet its six reasoning claims rested on earlier conversation context rather than on this turn’s evidence. The audit stamped all six WARN: no evidence source — reasoning claim.
That is the system working as intended: conversation memory is not evidence. A WARN is not a failure — it is a precise pointer telling you which claims to verify yourself before acting on them.
What a dossier changes
You don’t read a Kairos dossier — you audit it. Evidence can be opened, designs can be re-run, structures can be inspected, and the audit trail can be verified. The trust question shifts from “should I believe this answer” to “show me where each claim lives.”
The loop is still honest about its boundary: fermentation remains on the roadmap, and wet-lab data — not any score — decides real expression. When the ferment stage comes online, validation data flows back into the same dossier.
References
- Aw R, McKay PF, Shattock RJ, et al. A systematic analysis of the expression of the anti-HIV VRC01 antibody in Pichia pastoris through signal peptide optimization. Protein Expr Purif, 2018. PMID:29601964
- Dorogova NV, Fedorova SA. Overcoming the problem of heterologous proteins folding to improve the efficiency of yeast bioproducers. Vavilovskii Zhurnal Genet Selektsii, 2025. PMID:41660601
- Ahmad M, et al. Chang CH, et al. Enhancing the efficiency of the Pichia pastoris AOX1 promoter via the synthetic positive feedback circuit of transcription factor Mxr1. BMC Biotechnol, 2018. PMID:30587177
- Zha J, et al. Advances in Metabolic Engineering of Pichia pastoris Strains as Powerful Cell Factories. J Fungi (Basel), 2023. PMID:37888283
- Maity N, et al. Statistically Designed Medium Reveals Interactions between Metabolism and Genetic Information Processing for Production of Stable Human Serum Albumin in Pichia pastoris. Biomolecules, 2019. PMID:31590267
- UniProt
Q6Q4H8— cytochrome c, Komagataella phaffii. UniProt:Q6Q4H8