A deterministic assessment foundation for consequential decisions

Agents propose. Workflows govern. Supervisors decide.

Your rules, your evidence, your state — and you decide how much of the work the system does. Underneath the six words is a foundation that runs your rules the same way every time, records why, and lets an officer talk to a case.

Deterministic
Same rules, same case, same result
Versioned & hashed
Every verdict names its ruleset
Holds the state
A computed position, not a flag

Sign a verdict yourself — the evaluator is at layer 2 ↓

The foundation01

Same rules. Same case. Same result. Always.

The core is a small, closed system. Two officers running the same case get the same verdict and the same reasoning chain. Every assessment records the ruleset version and hash it ran under, so yesterday's decision stays defensible against yesterday's rules. Every verdict traces to the branch of the rule that produced it. And the state of every case is a live, queryable position — not a folder of documents and status flags.

Not features
The foundation
Everything else
What it makes possible

Why governance still matters — the file ↓

The design rule · 02

Six words. The whole platform.

Agents propose. Workflows govern. Supervisors decide.

You are now crossing the line — and every crossing is audited. Model outputs are proposals, never decisions, until the foundation accepts, rejects or revises them against the rules. Governance is not what is being sold; it is what makes the foundation trustworthy in operation.

The evaluator03

It walks each rule against the case, and says why.

Branches are evaluated top to bottom; the first whose condition is satisfied wins. Out come the per-rule verdict, the reasoning chain, the pre-decision actions and the updated state. Rulesets are data, not code — bump a threshold, re-version, re-hash, redeploy, without engaging our engineering.

Everything the foundation can do fits in a document a lawyer can read. A specification, not a black box
Every verdict
Rule · branch · ruleset hash
Every ruleset
Data, versioned, hashed
Audit
Append-only

The state of every case04

A computed position, not a status flag.

What has arrived, what has been extracted, what is missing, which rules have been evaluated, which verdicts were overridden, what is outstanding and who owns it — for every case, at every point in its life. Queryable. Grounded in your ontology, not an open-web index. Which is why voice can speak to it without inventing.

Traditional
Documents and status flags
Here
A computed position

Automation depth · 05

Set per rule, on evidence, at a pace you control.

Because every decision has an auditable reasoning chain, the organisation can raise the depth of automation one rule at a time — and defend the raise.

Advisory

The officer signs. The system is a second opinion, reasoning shown.

Assisted

The system proposes a verdict. The officer approves or overrides, with a reason.

Delegated

The system signs inside a defined band and escalates the rest.

Autonomous within envelope

The system signs and logs. The officer audits after the fact.

Per rule, not per workflow

A refusal stays advisory. Pure arithmetic can run autonomous. You choose; the system enforces.

One audit record for all of it

Depth changes who signs. Nothing else in the flow changes.

The AI layer never signs. Every output is evidence for a human, or a proposal for a human.

Foundations · 06

Not a pilot. An engine with a service record.

The AI layer is new. The foundation underneath it is not — and the foundation is what carries the decision. The state model, the evaluator and the voice interface run on the same engine.

20 yrsof regulated government production — hardened in service, granted in renewals
~285,000accounts administered on a single statutory register · governed as a designated critical system
17 yrscontinuous land and asset operations for one national public agency · 2008–2025
0recorded security breaches or unplanned downtime across two decades · 3 sovereign jurisdictions

The record — renewals, audits, changes of government — is available to your assurance team.

The AI layer, inside the fence07

Where the rules stop, the reasoning chain doesn't.

Real cases have edge cases the rules do not cover. Language models draft the reasoning for those, in the same structure as the deterministic parts, for an officer to accept, edit or reject. Vision models read documents and drawings — site plans, overlays, registers — into the form the rules run against. A voice model speaks to the state. Agents coordinate the work. None of them signs.

Policy loop
Overrides become amendments · quarters
Case-state loop
Every action updates the position · minutes
Interface loop
Richer state, better voice · continuous

One platform, four parts, one signature

Deploy AI without surrendering the decision.

The question is no longer whether agencies will use AI — it is what they will be able to prove afterwards. Bring a workflow; the demo takes minutes, not a procurement cycle.

Questions agencies ask first

What is WorkSym?

A deterministic assessment foundation for organisations that make consequential decisions from written evidence. It runs your rules against a case, produces a verdict for each rule and records why — the same rules, the same case, the same result, always.

Is it an LLM wrapper?

No. The foundation, not the model, decides. Every verdict traces to the ruleset version and hash that produced it. AI extracts, drafts reasoning for edge cases, proposes amendments, and powers voice. It never signs.

What is automation depth?

A per-rule setting for who signs: advisory, assisted, delegated, or autonomous within envelope. A rule that produces a refusal stays advisory; pure arithmetic can run autonomous. Depth changes who signs; nothing else in the flow changes.

Can an officer drive it by voice?

Yes. A spoken instruction becomes a typed operation on the state. Ambiguous references are refused, not guessed; anything consequential is read back and confirmed; and the utterance, the intent and the result go to the same append-only audit record.

Where do the AI models run?

Chosen per role and per classification — inside the tenancy, behind a government DMZ, or in sovereign cloud. Every model writes the same audit record.

The evidence file

Governance is not what is being sold — it is what makes the foundation trustworthy in operation. Here is why it matters to the agencies that will run it: every entry sourced to the public record.

The mandate is real — agencies have been ordered to adopt.

Australia’s Policy for the responsible use of AI in government has applied to non-corporate Commonwealth entities since 1 September 2024, with accountable officials and public AI transparency statements required. Singapore published the Model AI Governance Framework for Agentic AI in early 2026 and updated it mid-2026 with testing guidance — it is the yardstick agency reviewers now reach for. Sources · DTA · digital.gov.au · IMDA press release · MGF for Agentic AI (PDF)

The budget squeeze is structural — standing still compounds.

Worldwide government IT spending was forecast at US$589.8 billion for 2023, growing 7.6 per cent year on year (Gartner) — while the Commonwealth efficiency dividend has cut agency running costs almost every year since 1987–88, at ongoing rates between 1 and 3.25 per cent. Demand compounds up; the base compounds down. The efficiency gap is why the AI mandate exists. Sources · Gartner · May 2023 · Parliamentary Library · The Commonwealth efficiency dividend

The failure mode is already priced — in public inquiry reports.

Queensland Health’s payroll replacement began as an A$6.19 million contract and was put at roughly A$1.25 billion in total cost by the Commission of Inquiry and KPMG’s review — a state Commission of Inquiry sat for it. New Zealand’s Novopay education payroll triggered a Ministerial Inquiry and years of remediation. Ungoverned automation is government’s most expensive mistake. Sources · Qld Health Payroll Commission of Inquiry · 2013 (PDF) · iTnews · KPMG estimate · NZ Ministerial Inquiry into Novopay

A raw agent fails the bar a regulated decision must clear.

A high-consequence decision needs a named accountable person, a reason that can be reviewed, and rules applied the same way over time. The assurance frameworks say so in terms: “Decision-making remains the responsibility of organisations and individuals, not the AI system itself” (NT AI Assurance Framework). Australia’s policy requires accountable officials; Singapore’s agentic framework centres human accountability and bounded agent behaviour. A raw agent, alone, supplies none of this — so the market’s current choices are ungoverned deployment or none. Sources · NT AI Assurance Framework · DTA policy · IMDA MGF for Agentic AI (PDF)

WorkSym

Assembling the stack