NovaBoyle researches, drafts and checks. It does not decide. Every result is a draft until people sign it off, every finding is cited and checked by a separate reviewer, your data stays on your premises, and every sensitive action leaves a record that cannot be quietly changed.
1 · People decide
The decisions that matter in a chemical lab are kept for people, and the software is built so they cannot be skipped.
A formulation is approved only when one person requests it and a different person approves it. There is no administrator bypass. Once approved, its composition is locked; a change means a new version.
NovaBoyle writes a draft SDS. It can only be issued through your own qualified reviewer, named with their qualification and registration number, and approved by a second reviewer. It is never labelled "compliant" or "certified by NovaBoyle".
Under Malaysia's USECHH Regulations a chemical health risk assessment is made by a DOSH-registered assessor. NovaBoyle hands the assessor the facts it holds and leaves the assessment, and every rating, to them.
Lab notebook entries are signed and witnessed, and amended rather than overwritten. Projects, research history, formulation versions and lab results cannot be deleted.
2 · Every finding checked
In chemistry a confident wrong answer is worse than none, so the research team is held to rules a reviewer can check.
Identifiers, properties, regulatory status and conditions must come from this session's database results. An empty result is reported as "no data found".
A verifier agent checks the report against the evidence. Invented references are caught mechanically, not by trust, and a low-confidence review sends the report back for one revision.
Every report has a Data Gaps & Limitations section: sources that failed, properties with no source, single-source claims, lists that were not screened.
Model-predicted properties are labelled "prediction, not measurement", and calculations are labelled "calculated, not measured".
Molar masses, dilutions and limiting reagents are calculated by code, never by the language model. Ambiguous input is refused, not guessed.
A regulatory list that has not been loaded reads "NOT screened", never "not listed". Every dataset shows its date.
3 · Your data
By default the AI runs on your own hardware and nothing leaves. A researcher can choose a cloud AI model for one session for higher quality; this table shows exactly what that changes.
| Data | On-premises AI (default) | Cloud AI (chosen per session) |
|---|---|---|
| The research question | Stays on premises | Sent to the chosen AI provider |
| Public database results | Stays on premises | Sent to the chosen AI provider |
| Your internal documents, past reports and SDSs | Searched on premises | Never sent |
| Approved formulations, lab results, project memory | Stays on premises | Never sent |
| Stored research records | Your database only | Your database only |
Your team can run NovaBoyle with no cloud AI at all. Lookups to public chemistry databases are read-only and carry no research data.
4 · Records you can audit
5 · How we test it
Any change to a model or prompt is measured against a fixed set of benchmark research questions before it ships. Most recent full run (25 September 2026, 12 questions):
A daily health check also tests every public data source through the same code the agents use, so a source that quietly stops returning data is flagged instead of trusted.
6 · Malaysia's AI governance principles
Malaysia's AIGE guidelines set seven principles for responsible AI. Here is where each one lives in NovaBoyle.
| Principle | In NovaBoyle |
|---|---|
| Fairness | NovaBoyle makes no decisions about people: it researches chemistry and drafts documents for people to approve. |
| Reliability, safety & control | Grounding rules, a separate verifier, mechanical citation checks, two-person sign-off, a calculator that refuses rather than guesses. |
| Privacy & security | On-premises by default; internal documents and lab data never go to a cloud model; project-level access; login throttling; security headers. |
| Inclusiveness | Usable on a phone at the bench; draft SDSs in English, French, Chinese and Korean using officially published hazard phrases only, never machine translation. |
| Transparency | Every finding cited; Data Gaps section in every report; predictions labelled; the research plan's reasoning shown; the AI model used for each session shown and recorded. |
| Accountability | Hash-chained audit trail, permanent records, named reviewers, signed and witnessed notebook entries. |
| Human benefit | People make the decisions: approvals, SDS issue and risk assessment stay with qualified people. |
Malaysia is preparing an AI governance law. When it is passed, we will publish how NovaBoyle meets it.
7 · Limits, plainly