Research · Spartan
What is a society of minds a substrate for?
Spartan is the other research line: not evolving networks, but a federation of persistent LLM-driven minds. The question the corpus sits on is whether its central claims can be made falsifiable at all, and then whether they survive the test. The method is to interrogate the architecture, find where it cannot be disproven, and redesign toward falsifiability.
The result so far
One mechanism is signed dead
Verbatim memory recall, replaying stored past under drift, loses to forgetting and relearning. It loses even with a perfect recognizer, because the past returns changed. That was the programme's central claim, it was pre-registered, and it failed.
The kill is deliberately narrow. It is licensed to generalise along its own mechanism and no further, and an earlier claim that the whole programme was finished has been retracted as an over-reach. Recalling the shape rather than the numbers, memory as learned dynamics, and every faculty that reads the present rather than the past all remain open and untested.
The log
Where the line stands
These are Spartan corpus numbers. Faber numbers its own findings separately and the two sequences are unrelated, so a number is only meaningful with its corpus named.
The signed negative: memory is falsified
Inject fired 507 times and added nothing. A 100% precision oracle is reliably worse than engine-plus-reset, and restoration loses to restart under recurrence jitter.
What survives after the kernel dies
Inject died, but retrieval queried from outside was the winner all along, and that is a public commons rather than an improving private kernel.
The scope of the negative, and the frontier
Saying the programme was over turned out to be a mood rather than a measurement, and is retracted. The kill is narrow: it binds verbatim recall under drift and nothing wider.
M1: self-audit economics
Does draft-then-verify cut hallucination enough to justify twice the compute? Mechanical checker, constant-free kill rule, and the programme's first cost ledger. Cleared, build-ready.
M2: peer-review economics
Does a second independent model's review earn its compute beyond self-audit, by removing decorrelated blind spots rather than by context poverty? Pending adversarial clearance.
How it runs
The method
Pre-register first
The metric and the kill threshold are written down before anything is built, so a disappointing result cannot quietly become a reframed one.
Adversarial clearance
An independent model reviews each design and each claim, with the job of attacking it rather than helping. Experiments wait for clearance before they run.
Plain language, always
Every note carries a plain-language section. If a finding cannot be explained simply, the corpus treats that as evidence it is not yet understood.
Credit
Spartan is Dr. Gene Sher's. Its mind, its mechanisms and its philosophy are his, and this research stands entirely on that work. What the commons contributes is a BEAM port, a federated mesh substrate, and the falsification work recorded above.