A claim may draw on dozens of publications, measurements, and databases while their provenance, experimental context, and relevance remain unclear. ORAKLIS reconstructs that chain to distinguish what is established, what remains uncertain, and what should be checked next.
We applied a rule declared before data collection to twelve targets used in cosmetic claims. Depending on the target, the share of measurements for which the assay species cannot be verified ranges from less than 1% to more than 70%. This exploratory analysis is separate from the JCIM-submitted preprint and its confirmatory CARA and NTAB analyses.
ChEMBL_37 · IC50 with pChEMBL · rule declared before data collection
The final two targets have no publicly usable potency measurements. Elastase, with near-complete traceability, shows that the method also identifies strong evidence records.
75 of 255 measurements attached to the target annotated Homo sapiens have an explicitly non-human or unresolved assay species.
Fifty-nine measurements that the database itself marks as having unknown origin are nevertheless attached to the human target.
Measured in ChEMBL_37, after the official organism re-curation in version 34. This is not a historical remnant.
| Target | Associated claim | Numerator | Denominator | Rate |
|---|---|---|---|---|
| LOX | firmness, elasticity | 107 | 151 | 70.9% |
| Tyrosinase | radiance, depigmenting | 75 | 255 | 29.4% |
| Sirtuin 1 | anti-ageing | 128 | 850 | 15.1% |
| COX-2 | soothing | 798 | 6,116 | 13.0% |
| Aromatase | hormonal | 419 | 3,848 | 10.9% |
| MMP-9 | anti-ageing | 239 | 2,414 | 9.9% |
| 5α-reductase 1 | anti-sebum | 29 | 398 | 7.3% |
| TRPV1 | soothing, anti-redness | 172 | 3,795 | 4.5% |
| 11β-HSD1 | skin anti-ageing | 142 | 4,522 | 3.1% |
| MMP-1 | anti-wrinkle | 48 | 2,268 | 2.1% |
| Neutrophil elastase | firmness | 1 | 1,858 | 0.05% |
| 5α-reductase 2 | hair-loss claim | 0 | 419 | 0.00% |
Median: 8.6%. Negative control—a non-human target (Agaricus bisporus tyrosinase, 983 of 999 measurements) correctly yields almost zero “human confirmed”: the rule discriminates. These rates are not a database error rate. They describe what our declared operational rule observes for a given target. A dedicated public package for this analysis is being prepared; the DOIs below do not reproduce this table.
The review does not chase a perfect score. It turns evidence received from a supplier, partner, or internal team into a usable map before listing, launch, investment, or a new testing program.
Each result is connected to its source, assay, target, organism, and context. You can see precisely which evidence genuinely supports the claim.
Contradictions, dependencies, and non-comparable data are prioritized. Uncertainty becomes measurable rather than a general impression.
You receive the control, test, or claim refinement that adds the most value before committing more time and budget.
On 30 July 2026, we publicly deposited the protocol and code before running the two confirmatory CARA and NTAB analyses. The timestamp, checksums, results, and complete reproducibility package are accessible: rigor rests on a process that can be checked, not on a promise.
3D view unavailable on this device.
The measured phenomenon is described in the preprint, including figures and code.
What this record does not say. It establishes no causality: the retraining experiment that alone could address it was declined and recorded as such in the deposit. It does not claim that public benchmarks are invalid. Nor does the finding automatically transfer to other databases—we tested a reference clinical database, did not reproduce the finding there, and reported that result as well.
Each control addresses a risk observed in real data. Together they show where the record is strong, where its scope needs clarification, and which check will produce the most information.
An evidence record does not receive a single score; it receives a profile. An identified limitation does not invalidate the entire body of evidence. It clarifies the defensible scope of the claim and identifies the next useful action.
ORAKLIS works where a team must decide from heterogeneous data. Each offer states its maturity, inputs, outputs, and what it cannot establish.
For a brand, ingredient supplier, or adviser who needs to determine whether the evidence received supports the exact claim, formula, and concentration at issue.
For modeling teams that need to know whether a split, annotation, or document provenance genuinely supports their evaluation.
A versioned register connecting every word in a claim to the sources, formula, market, and assumptions that make it supportable.
For a protein pocket, kinase panel, or mechanistic record: compare hypotheses, select measurements that could change the decision, and abstain when the data are insufficient.
Declared boundary. The Pocket, Molecular Interrogator, and Adaptive Kinome programs remain research collaborations or prospective pilots. Their internal results are not presented as universal commercial validation. ORAKLIS offers a better-informed decision and a verifiable plan—not biological certainty.
We turned every detected anomaly into a control rule or regression test. This transparency does not replace evidence for our method; it shows how the method became stronger.
Our tool had linked a canine enzyme to its human counterpart. The anomaly was detected, documented, and retained as a regression test for every subsequent analysis.
An early rule confused the studied species with the expression host or culture medium. Correcting it reduced some rates sevenfold; we report the corrected estimates.
Performance above every reference came from selection leakage. It was discarded, and the corresponding control is now part of the protocol.
The authors report limitations, deviations, and negative results alongside favorable findings. AI systems used under human supervision are neither authors nor reviewers; scientific responsibility remains with Mickael Pluym and Nicolas Calandra.
You provide a claim and the materials supporting it. We reconstruct its evidence chain, define its defensible scope, and identify the smallest control that will genuinely strengthen your next decision.
The result specifies what you can support today, what needs narrowing, and the check required to go further.
The first reviews are founding pilots provided at no charge. In return, we ask for written feedback on usefulness and confirmation or correction of the issues identified. We use an anonymized case only under a separate agreement. An anonymized case and a named quotation require two distinct permissions.
Tell us the claim, target, or decision at issue. One sentence is enough to begin; we will reply with the scope of an initial review.