Kairos research workbench adds an automated “review” layer — every citation traceable

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.

Why we built this

The biggest trust problem with AI research assistants is citation drift: the model may cite a PMID that was never in the retrieval results, or quote a design-panel number incorrectly. A human has to check each citation by hand before trusting it — which is exactly the tedious work AI is supposed to eliminate.

Kairos now adds a deterministic review layer (the research pipeline is an explicit retrieve → resolve → design → reason → review → auditable artifact graph). After generating advice, it automatically checks three things:

1. Citation authenticity — is every PMID/UniProt cited in the reasoning actually present in the evidence set retrieved for this run? (catches fabricated citations)
2. Numeric consistency — do the design numbers mentioned in the reasoning (e.g. codon adaptation, host-fidelity scores) match the design panel’s output? (catches number drift)
3. Claim traceability — does each conclusion have a citation backing it, or is it an unsupported assertion?

The review outputs a verdict (PASS / WARN / FAIL) and lists which citations traced to evidence and which did not. The response also carries a claim-level auditable structure (each claim → evidence ID → verification status) for line-by-line checking — you don’t need to read the whole report to verify.

Honesty boundary (important): this review layer only checks traceability (is the citation present, are the numbers correct) — it does not judge scientific correctness, does not predict expression/titer. Kairos’s consistent stance is unchanged: it ranks dry-lab design quality; real yield is always for the wet lab to measure.

A real run (live, unedited)

Question: What signal peptide for secreted expression in Pichia?

Retrieval returned 6 relevant literature passages + knowledge-graph nodes, including:

• PMID:29601964 — improving anti-HIV VRC01 antibody expression in Pichia via signal-peptide optimization
• PMID:41660601 — dependence of secretion efficiency on signal-peptide topology and cleavage site
• PMC11018594 — secreted expression of recombinase PMGL-Ba in P. pastoris X33

Reasoning (real Qwen3, unedited):

“Signal peptides are critical for secreted expression in Pichia; their efficacy depends on correct topology, cleavage site, and compatibility with the target protein (PMID:41660601). Systematic signal-peptide optimization has been shown to improve antibody yield (PMID:29601964), though the specific signal-peptide sequences are not detailed in the provided evidence.”

Note the model proactively admits “specific sequences are not detailed in the evidence” — it does not fabricate.

Automated review (section 6, live output):

## 6. Review (automated citation/number audit — verdict: PASS)
- ✅ Citations traced to evidence: PMID:29601964, PMID:41660601

The review layer checked the two PMIDs the model cited — both are genuinely in the retrieved evidence set, no fabricated citations → citation trace: PASS. The response also carries a claim-level auditable structure, linking each conclusion to its evidence ID for line-by-line checking. As on the homepage, this PASS confirms traceability, not scientific correctness or relevance — a cited source can still be indirect or off-target; automated relevance/entailment scoring is on the roadmap.

How the review layer was validated

Before launch, we stress-tested its ability to catch errors using adversarial tests (samples where the model deliberately makes mistakes):

• Fabricated citation PMID:99999999 (not in evidence) → flagged as missing, verdict downgraded
• Number drift (reasoning says fidelity 0.95, design panel says 0.82) → flagged as mismatch, verdict FAIL
• Unsupported assertion (“will definitely express at 5 g/L”) → flagged as ungrounded claim

Then on real Qwen3 output we verified it does not false-positive: in the real run above, the model cited honestly with no fabrication, and the review layer correctly returned PASS — proving it isn’t “alarm on every citation” but is genuinely checking.

What this step means

Every research guidance document from the Kairos workbench now carries a traceable review layer: citations authentic, numbers consistent, claims checkable. This shifts “AI research advice” from “should I trust it” to “I can verify it.”

Combined with the existing three-tier provenance (retrieved citations · reasoning · deterministic design tools) and host-fidelity design ranking, Kairos is becoming a Pichia research workbench where you can audit every step.

References

The live run above retrieved and cited these three real sources; the review layer verified each is genuinely in the evidence set.

1. 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
2. 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
3. Li J, Yang M, Zhao F, et al. Efficient expression of an alkaline pectin lyase from Bacillus licheniformis in Pichia pastoris. Bioresources and Bioprocessing, 2024. PMC11018594