Auditable on purpose. Every score, every draft, every answer.

A score you can check, line by line

Every match score is arithmetic, not a vibe. It starts at a baseline and every contribution is a labelled, signed line you can read on the tender page: topic fit, geography, your stated preferences, value band, past delivery echoes. Semantic similarity from AI embeddings joins as one more labelled line, calibrated within your own profile's distribution so "similar for you" means something. And a tender cannot score above 85 unless there is delivery proof behind the claim; strong signals without evidence are capped, and the cap itself appears as a line in the breakdown.

A match score states fit. It is never a win probability, and we never call it one.

Where the model is allowed to speak

The deterministic maths scores everything, cheaply and identically for everyone. The language model gets a voice in two metered places: a deep read of a tender you ask for, whose adjustments appear as labelled lines in the same breakdown, and drafting. When your deep reads keep correcting our heuristic in the same direction, the heuristic learns a small downward prior for your account; that correction is itemised too. No hidden reweighting, anywhere.

Drafting: cite or ask, never invent

Tender Mate, Helm's drafting assistant, works from exactly four permitted sources: your company profile, the tender's own documents, your evidence library, and answers you have approved before. A claim that cannot be traced to one of them is not written as fact; it becomes a visible [CONFIRM: ...] marker asking you for the specific thing the draft needs. Compliance matrices quote your documents verbatim, with file and page.

Ask the pack: cite or refuse

Questions about an uploaded tender pack are answered from retrieved excerpts of that pack and nothing else, with the file and page cited inline. When the excerpts do not answer the question, Helm says so in so many words rather than improvising. When two documents in the pack disagree, both are shown; the conflict is never silently resolved for you.

Labels on content

Anywhere AI tidies or classifies, the result is shown as Helm's reading, with the raw source kept alongside. The AI restatement of your own profile applies to scoring only after you have read and confirmed it. Sample data is labelled as sample data.

Private tenders: the same maths, or nothing

A tender you bring in yourself is scored live with the identical itemised heuristic and semantic signal, run on the text you pasted, or it is not scored at all. If the maths cannot run you see "unscored", never a made-up number. Private tenders are never stored in the shared index, never matched for anyone else, and never appear on public pages.

The gates that hold this to account

• Hallucination audit: before every release, drafting runs against rich, thin and deliberately vague test profiles, plus a customer-entered private tender, and every number, certification and name in the output must trace to a source or the release fails.
• Score integrity audit: every stored score is recomputed from scratch and compared. A drifted score fails the gate.
• Versioned prompts: the instructions the model runs under are operator infrastructure, reviewed and versioned like code.

What the AI never does

• No win probabilities. Fit is knowable from the outside; your odds are not.
• No bid/no-bid verdicts. Helm states evidence and gaps; the pursue decision is yours.
• No invented credentials. A certification you do not hold cannot appear in a draft as if you did.
• No automated submissions. Helm prepares; you submit.
• No training on your data. AI analysis runs through OpenAI under their API terms; your data does not train their models, and Helm trains no models on it either.

Where your data lives, and what leaves Australia in transit, is on the security page. What the product does end to end is on how it works. Ask us anything specific through the contact form; we answer in writing.