grantmaking.ai Launch Round
The core hypothesis of MAVS is that governance and prediction are two different computational problems. Its foundations are currently divided into two classifications:
- The Main Architecture.
- Diagnostic Sciences.
1- The Main Architecture:
The Main Architecture of MAVS is basically the core foundation of MAVS, without this section, the entire architecture and Diagnostic Sciences and any additions would be rendered obsolete.
This section effectively introduces how MAVS seperates prediction and governance into different computational problems. In accordance to the referenced docs, MAVS does this by making them never solve the same optimization problem. It introduces a governance simultaneous parallel working which is evidence-based, i.e through diagnostic functions (gi) it observes the situation, reweighs the specialists in accordance to it, calculates severity A(z) and then works on governance threshold, i.e how much evidence is needed to trust a given prediction.
Then it calculates R which is how much evidence we have, if R is more than theta and severity is less than t-hard (it represents the hard-risk boundary), then it decides to trust a given prediction.
However, if severity crosses the t-hard, then no matter even if R is more or equals to governance threshold, the given prediction will not be trusted.
2- Diagnostic Sciences:
Diagnostic Sciences, is based on the foundations set by the main architecture where it mainly deals with how high quality is a given diagnostic function (gi) where a high quality gi is defined as:
"A high quality diagnostic function is objective aligned perception extension that detects intended hidden structure, is not redundant (overlaps with already existing diagnostic functions), and beneficially changes MAVS behavior chosen for a given objective while minimizing unintended damage"
Diagnostic Sciences is to MAVS what optimization theory is to optimization algorithms, it transforms the diagnostic layer from a fixed implemention into something that can be changed and optimized to change the behavior of MAVS overall for a given objective.
3- Evidence up until now:
MAVS has been tested on 2 tabular bounded static benchmarks, and two dynamic unbounded benchmark with randomized corruption families under a period of time.
Tabular bounded static benchmarks:
10A, 10B, 10C
Dynamic Benchmarks:
10D, Diagnostic-Science Specials.
10A, 10B, 10C shows that MAVS in static conditions, is competitive in accuracy under normal conditions but not winning. However, under normal corruption, MAVS had 22 times less false positive acceptances (~1.45% of unsafe acceptances) and under high corruption, it was 144 times to 202 times less false positive acceptances (~0.45%) whilst maintaining the highest accuracy in both (85% to 90%).
10D shows MAVS under dynamic benchmarks, where MAVS had a consistent 0% unsafe acceptances across all suits (~12,000 governance decisions), however, it had a significant false rejection rate (10-30 percent) and under Shared Representation Correlated Failure (all models got the wrong representation, all models were wrong but they confidently agreed with each other.
In this case, MAVS intentionally shut everything down by increasing its false rejection to 100% while keeping its unsafe acceptances to 0. However, this was later on fixed massively.
Diagnostic Science Specials is an evolution of 10D, where we used Diagnostic Sciences to optimize z for MAVS to have minimal false rejection rate, and 0 unsafe acceptances, specifically under shared correlated failure.
In this case, MAVS got 0% unsafe acceptances, and 0% false rejection in shared correlated failure, whilst maintaining a mean reward of ~0.83. And this occured consistently in all various corruption families where MAVS had 0 false rejections and unsafe acceptances except multi agent triage where it was deliberatley made to attack optimized MAVS strength, i.e severe evidence masking, where it had a false rejection of 7% and unsafe acceptances of 0%.
4-What's next?
Next, currently I am working on a direction of MAVS where it dynamically uses Diagnostic Science to optimize its diagnostic functions or z for a given situation for each prediction that occurs.
I refer to this as MAVS-Self-Learning, as another feature of this architecture is that it becomes better as it goes through more predictions and likewise.
However, for this, it requires a scale of experiments that currently I am unable to do so feasibly; requiring us to test various aspects of MAVS-Self-Learning for example it's latency, it's compute usage, it's accuracy in terms of optimizing its z, if it generalizes beyond experiments, ablations, adversial testing, cross-domain verification, production-grade feasibility and likewise.