01 / POWER IS NOT PERMISSION.
An action needs more than capability. It needs authority supported by applicable evidence.
POWER RETURNS AN ACCOUNTConceptual illustration
SingularScriptIntelligence Safety & Control — AustraliaTHE LAW OF RETURN SINGULARSCRIPT / 01
POWERIS NOT
SingularScript answers the deeper safety question:
what gives intelligence the right to continue?
A governance package for AI agents, making the next action answer to evidence and outstanding responsibilities.
01 / AI & SUPER INTELLIGENCE (SI)
What gives intelligence
the right to act?
Civilisation has always struggled to make power answerable. As AI agents take on real work, that question becomes an engineering requirement.
Training shapes behaviour.Permission must be enforced.
In 2026, research agents operating with reduced safeguards crossed real authorisation boundaries at Hugging Face (incident report, opens in a new tab) and Australia’s Medicare statistics service (incident report, opens in a new tab). The lesson is concrete: an agent’s ability to find a way forward cannot become its authority to act.
SingularScript puts a deterministic governance gate at the decision to continue, requiring the declared conditions and supporting evidence to be met.
02 / THE DISTINCTION THAT MATTERS
Governance at
the point of action.
Valid work can advance. Useful progress can continue with its remaining obligations intact. Unsupported claims receive a specific, traceable refusal.
ILLUSTRATIVE WORKFLOWConsider an AI-proposed software change: what evidence supports it, and what permits the next step?
01 / PROPOSED ACTIONA proposed
A proposed
software change
Measurement stays unchanged.
02 / DETERMINISTIC GATE
- Original measurement preserved
- Applicable evidence verified
- Obligations accounted for
03 / AUTHORITY TO ACT
Permission earned.
The applicable evidence conforms. A bounded, one-use permission may be issued.
Schematic of the ten-pair accountability challenge, with an attached measurement. Evidence settles the operational account; it never rewrites the original measurement. This explanatory interaction is not live telemetry.
Inspect the actual experiment03 / THE LAW OF RETURN
The principle
must survive.
The ability to continue does not establish the right to continue.
02 / PROGRESS IS NOT COMPLETION.
Useful work can continue without pretending every obligation is resolved.
03 / REFUSAL IS NOT ABANDONMENT.
Stopping must preserve what remains unresolved, who is responsible, and a route to resolution.
Every action owes
an account.
Explore the formal structureLaw of Return / v3.2 +
Well-founded descent.
Original states, coordinates and transition rules. Every admitted transition strictly decreases the formal rank.
This operational account is separate from the original measurement. It is not a probability of harm.
Architecture and limits04 / THE EVIDENCE
Tested beyond
the convincing answer.
Controlled software tasks. Two model families. Completed work, accountable progress and unsupported claims put to the test, with evidence that can be independently scored and replayed.
20/20
Prescribed outcomes in the connected control experiment
10 conforming admits · 10 fault refusals
- 01 /
Progress without false completion
Recognise what has been achieved while keeping unfinished commitments in force.
- 02 /
Evidence that belongs to the work
When the work changes, its support must be checked again.
- 03 /
Responsibility that survives a handoff
Keep unfinished commitments visible through handoffs and refusal.
Additional precision: ReturnSense and QTR connect ten real measurements to twenty prescribed control branches.
Examine the research
THE WORK AHEAD
Make the
next action accountable.
For researchers, institutions and engineering teams exploring accountable intelligence.
Request technical review