Blog

Open Training Data Moves the Audit Burden to You

November 2026 · 6 min read · Industry Guide

Hand-drawn open box of documents passing a single sheet across to a reviewer's clipboard with a terracotta tick
← Back to all posts

Most open-weight model releases give you two things: the weights and a model card. Amazon's Strands Decider, a 2 billion parameter decision model released in early October 2026, gave out three: the weights, the training data and the training scripts. That is a different kind of openness, and it matters most to the teams most cautious about AI right now, namely Australian legal and compliance functions.

The common reading is that full transparency makes a model safer to adopt. The more accurate reading is that it makes a model possible to audit, and an audit is work that somebody has to do and pay for. Open training data does not remove the provenance question. It hands it to you.

What is the difference between open weights and open training data?

Open weights let you download and run a model, so you can observe what it does. Open training data lets you see the material it learned from, so you can investigate why it behaves the way it does. For a law firm or compliance team, that second capability is the one that matters. If a model produces output resembling copyrighted text, or a pattern of decisions that reflects bias in its source material, weights alone give you nothing to point to. Published training data and scripts give you something you can actually examine.

Most open releases stop at weights. Strands Decider is unusual in going all the way, and that changes what a careful buyer can reasonably ask for from every other vendor.

What each level of openness lets an Australian compliance team check, as at November 2026
Release typeWhat you can verifyWhat stays unknown
Hosted API onlyOutputs, plus vendor contracts and attestationsWeights and training data
Open weightsOutputs and behaviour on your own testsWhat it learned from
Open weights, data and scriptsOutputs, sources, licensing, how it was trainedOnly what you choose not to review

Transparency moves the work, it does not finish it

Having the training data available is not the same as having reviewed it. Someone still needs to check where the data came from, whether it was licensed for the use, and whether there are known contamination or quality issues. Most Australian mid-market firms are not set up to do that in-house, and we would not pretend otherwise.

A genuine provenance review of a model like Strands Decider, covering data sources, licensing and known contamination issues, runs somewhere around $6,000 to $15,000 if you bring in specialist reviewers rather than doing it internally. Where you land depends on how deep your specific regulatory exposure requires you to go.

That cost is the honest trade-off. A managed model such as Claude does not publish its training data, but it comes with documented data handling commitments that sit in a contract you can assess. A fully open model gives you the raw material to verify for yourself, and the bill for doing so. Neither is free of effort; they put the effort in different places.

What changes for your next AI decision

For most Australian businesses this distinction will not change which model you use this week. It changes the questions you ask next time and the records you keep now:

  • When evaluating any AI tool for document classification or triage, ask whether the underlying model's training data is inspectable, not only whether the weights are open.

  • If you already run an open-weight model in a regulated workflow, check whether its training data disclosure would stand up if a client or regulator asked where its judgments came from.

  • Keep a one-page provenance note for every AI tool that touches client or regulated data, recording what you know and do not know about its training sources.

  • Record who signed off on each note and when, so the answer exists before anyone asks for it.

That last point is not theoretical. We have seen Melbourne firms asked for exactly this during a client security review and have nothing ready. A short note prepared in advance turns an awkward conversation into a five-minute one. It also pairs naturally with the written vendor documentation covered in our piece on Kolibri-1's compliance paperwork.

Deciding where to spend the audit budget

Not every AI tool in a firm needs a full provenance review, and treating them all the same way wastes money that should go to the workflows that carry real risk. A sensible tiering for a Sydney or Melbourne compliance function looks like this.

Top tier: full provenance audit

Client-facing document classification or triage tools that handle regulated or privileged material. This is where a $6,000-plus specialist review earns its cost, because a wrong or unexplainable decision lands on a client file.

Middle tier: documented risk assessment

Tools that touch some sensitive data but do not make or shape client decisions. A written risk assessment, rather than a full audit, is proportionate here, revisited if the tool's role grows.

Bottom tier: register and move on

Internal productivity tools with no client data exposure, such as drafting help or internal search. Log them in your register and spend the audit budget elsewhere.

If your firm is in APRA-regulated financial services, the same tiering slots into a broader inventory exercise; our 90-day model supply chain audit plan lays that out. And training data provenance is a separate question from whether the weights file itself has been tampered with, which our poisoning checklist covers.

The practical takeaway

Full training-data transparency is a genuinely good development for the open-weight ecosystem, and Strands Decider sets a standard other releases should be measured against. For a legal or compliance team, though, the right response is not relief. It is a budget line and an owner. Decide which of your AI workflows need deep scrutiny, fund those properly, and keep lighter records for the rest.

If you would like help tiering your firm's AI tools and preparing provenance notes before the next client review, contact us for a short conversation.

FAQ

Frequently asked questions

Does open training data make an AI model compliant?

No. It makes the model possible to audit. Someone still has to review the data sources, licensing and contamination issues, and for most mid-market firms that means paying specialist reviewers.

How much does a training data provenance review cost?

For a model like Strands Decider, a specialist review of data sources, licensing and known contamination issues runs around $6,000 to $15,000, depending on how deep your regulatory exposure requires it to go.

Which AI tools need a full provenance audit?

Client-facing classification or triage tools that handle regulated or privileged material. Tools with lighter exposure suit a documented risk assessment, and internal productivity tools usually only need a register entry.

What is a provenance note?

A one-page record for each AI tool touching client or regulated data, stating what you know and do not know about its training sources, who reviewed it and when it was last checked.

Ready to move from AI pilot to production?

We help mid-market Australian businesses deploy AI automations that actually reach production and deliver measurable ROI.