What it is

The protocol runs in phases. Phase I sweeps coupling in a sparse leaky-integrate-and-fire RNN (Numba-JIT, SciPy CSR networks) to find its critical point — avalanche statistics, power-law fits — and maps its computational domain (effective information, LZ complexity). Phase II evolves populations (DEAP) under three fitness regimes — direct task performance, EI, criticality — plus controls, on chaotic time-series prediction. Phase III moves the champions into an embodied grid-world and a zero-shot “Champion’s Gauntlet.”

The codebase is real: ~13,000 lines of Python, Hydra-configured, with a Flask server, test suites, and a genuine 10,001-step Phase IIIb run in its outputs. Then the protocol did the thing protocols exist to do: its 2026 calibration campaign tested whether the measurement pipeline could certify criticality at all — and the answer was no.

Code

src/cti/phase_i/simulation.py (exact):

def generate_er_network(config: DictConfig) -> csr_matrix:
    """
    Generate an Erdős–Rényi random graph as a SciPy CSR sparse matrix.

    This function explicitly avoids self-loops (i.e., diagonal elements).

    Args:
        config (DictConfig): A configuration object with 'phase_i.network.n' for number of nodes
                             and 'phase_i.network.p' for connection probability.

    Returns:
        csr_matrix: A sparse matrix representing the network's adjacency matrix.
    """
    n = config.phase_i.network.n
    p = config.phase_i.network.p

Achievements — and the result

  • Phases I–II machinery complete; fast test suite runs green locally (9 passed, 3 xpassed, assessed 2026-11-11).
  • The finding (Stage 1, closed 2026-11-11): 0 of 60 points Clauset-certifiable. At the near-critical points a lognormal fits better than a power law (28/28), and a second calibration round on a redesigned substrate (E/I 80/20, Dale’s law) failed the same way. The calibration question is closed for N=120 substrates of this LIF family: the engine shows quasi-critical phenomenology, not scale-free criticality.
  • The honest consequence is recorded too: the Champion’s Gauntlet — the experiment the protocol exists to run — was never run, because its premise couldn’t be certified. A σ-distance pilot (which needs no certification) remains available and uncommissioned.

Lessons

  • A negative calibration result is a result. Two rounds, two substrates, one answer — that is the instrument working, and it is worth more than a tuned success.
  • Kill criteria only count if you obey them. The protocol’s own criteria said what failure would look like; when it appeared, the campaign closed instead of shopping for a friendlier test.
  • Document the warts in the code. The Phase I sweep’s parallel deadlock is written down where it happened — future readers inherit the knowledge, not just the workaround.

Status: Dormant — calibration concluded, Stage 2 pilot uncommissioned · Python, PyTorch, Numba, SciPy, DEAP, Hydra