You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$50,000 left to reach the funding goal.
The project were working on is Sanctuary V3. It has been designed to separate what an AI (LLM/Agent) proposes from the authority to carry it out. It checks whether an operation is permitted, constrains execution, handles interruption and recovery. It then goes on to produce evidence which can be examined, to establish what happened & where the evidence is incomplete.
When the AI tells us that it has authority/permission OR if it reports that a task succeeded, it doesn't actually provide sufficient evidence of either. In fact, as AI systems carry out long tasks/sequences of actions across tools & services, it tends to produce mistakes, misleading instructions & uncertain outcomes. therefore, this can have consequences beyond the models response.
In mind of the above, we have developed runnable prototypes, which have been tested with real file operations, separate job services, synthetic ledgers & two mock business services. The work includes Live multi step agents, interrupted operations, withdrawn permissions, repeated requests & evidence checking.
These are bounded research prototypes. The business services are test environments & the ledger does not move money. The current implementations are yet to undergo independent technical validation.
We are seeking funding for 6 months of continued development and independent evaluation. A major part of the budget would pay for an independent university led technical assessment (subject to agreed scope & cost). This would include independently operated testing of Sanctuary V3's controls, recovery behaviour & supporting evidence. The remaining funding would support my working time, computing resources and investigations of the findings.
The overall aim is to establish which claims independent reviewers could substantiate, also where & if the system falls short & what further development is needed.
The aim over the next 6 months is to continue developing Sanctuary V3 & have its current capabilities independently assessed.
A major part of this would be paying for a university-led technical assessment, subject to agreeing the scope & cost. This has not yet been commissioned. We would agree which version of Sanctuary is being assessed, what claims are being tested & what evidence would be needed to support the findings.
The assessment would look at whether permitted operations complete correctly, whether prohibited or repeated requests cause unintended effects & what happens when an operation is interrupted. It would also examine whether the evidence produced is sufficient to establish what actually happened.
I would want the reviewers to run the tests themselves & challenge the system with their own cases. Where appropriate, this would include comparison with existing safeguards under equivalent conditions.
Alongside this, I would continue development & investigate the findings. Any changes would be kept separately versioned, so a repaired version is not confused with the one originally assessed.
The overall aim is to come away with an independent account of what Sanctuary can demonstrate, where it falls short & what needs further work. I intend to share a useful summary of the findings, while agreeing suitable arrangements for access to proprietary technical information.
The main thing would be whether someone independent of the development can reproduce the results & establish whether the claims are supported.
Before testing, we would agree what counts as a completed task, an unauthorised effect, a duplicate operation, an unnecessary refusal or an uncertain result. Simply blocking everything would not be a useful outcome. Permitted work still needs to complete correctly.
The assessment would examine actual effects alongside the control decisions & recorded evidence. This includes whether an interrupted operation can be recovered without repeating an effect, and whether incomplete evidence is recognised as incomplete.
It should also show where existing safeguards achieve the same result & whether Sanctuary adds anything useful in those conditions.
I am not looking to pay for a positive conclusion. If independent testing finds a weakness, that is something I need to know. Any repair would then need further testing before I could claim that the problem had been addressed.
We are seeking between USD 25,000 & USD 50,000 for 6 months of continued development & independent evaluation, beginning once funding is available.
A major part of the budget would pay for the university-led technical assessment. This would cover agreeing the test scope, independently operating the tests, reviewing the evidence & reporting the findings. The university, final scope & price are still to be agreed.
The remaining funding would support my working time, computing resources, server rental where needed & preparation of the materials required for the assessment. It would also allow time to investigate findings, make changes & test them properly.
At the minimum amount, the assessment would focus on a selected integration & a limited set of clearly defined claims. The full amount would allow broader external testing, more workloads & further assessment of changes made in response to the findings.
The final allocation will depend on the assessment quotation & the agreed work. Any material change to the budget or scope would be made clear to funders.
This request is for work still to be done. It does not include patent prosecution, general marketing or reimbursement of earlier development costs.
I am the founder of Novel Symbolic Ltd & the principal developer of Sanctuary V3. I have been working on the project for roughly a year & have funded the development myself. There is currently no paid development team or commissioned university assessment partner.
My relevant track record is the work completed on Sanctuary itself. This has progressed from earlier control prototypes into working file operations, separate job services, synthetic-ledger tests & two mock business-service integrations.
The work includes live multi-step agents, interrupted operations, withdrawn permissions, repeated requests & evidence checking. I have also carried out comparisons with existing approaches given equivalent safeguards.
The results have not all been positive. Testing has exposed failures, incomplete tasks & limits to what the evidence can establish. These have been retained in the records, with later repairs kept separately versioned. AI tools have been used during development, testing & evidence preparation.
A business acquaintance previously ran frozen test protocols for earlier versions on a separate machine. That was useful, but it was not an independent audit of the current system.
I have since prepared reviewer materials, test records & versioned packages to support a more thorough external assessment. Obtaining that assessment is a central reason for this funding request.
One possibility is that Sanctuary does not provide enough additional benefit over existing safeguards to justify its complexity. Some of the internal comparisons already show that equally safeguarded alternatives can achieve similar results.
Another possibility is that the controls work within the tested conditions but do not hold when those conditions change. The evidence may also be incomplete, or the system may prevent legitimate work too often.
There are known limits already. For example, the current work does not establish protection against a privileged administrator or reliable recovery across arbitrary external systems. These are boundaries to investigate, rather than capabilities I am claiming.
There is also a practical risk that university assessment costs more or takes longer than expected. Agreeing a clear scope before commissioning the work would help manage this.
If the assessment finds problems, I would document them & investigate what can reasonably be addressed. If a particular mechanism adds little value, I would simplify, replace or stop developing it.
I have not received external grant funding for Sanctuary to date. The development has been self-funded.
Previously recorded direct expenditure includes approximately RMB 50,000 on the workstation & components used for development and testing. There have also been smaller software & operating costs which I have not separately tracked. This is money I have spent myself, rather than funding raised.
I applied to the EA Funds Transformative AI Fund on 1 September 2026 for a mainline amount of USD 100,000 over 12 months. That application is still pending.
I also applied to NPL’s Measurement for Business programme on 22 September for technical assistance. Receipt was acknowledged, but no support award is confirmed.
This proposed 6-month project overlaps with part of the wider programme described in the EA Funds application. If either application receives an offer, I will disclose it to the relevant funders & agree how the work and costs are allocated. The same expenses would not be funded twice.
There are no bids on this project.