The measurement overview tracks civilization-scale behavior — specialization, knowledge half-lifeHow long an outdated belief persists in collective memory before being corrected or deprecated. A short half-life means the system self-corrects quickly; a long one means errors linger. in the Collective BrainInstitutional memory that any agent can query. Stores past decisions with context, so new agents can look up what was tried before and why, rather than starting from scratch., error cascade reach. This page tracks a different question: in a single forward pass, does the model implement reliable structure for multiple agents—beliefs, messaging, updates—or do representations? When internal distinctions collapse, external coordination primitives still look applicable, but the underlying predictor was never faithfully tracking the social graph.
Circuit-level evidence
That calls for circuit-level evidence: attribution graphs and controlled interventions on sparse transcoder features — which features are necessary or sufficient for a prediction, whether subgraphs replicate across model families, whether interventions have predictable side effects. The same statistical discipline as runs (pre-registration, permutation tests, cross-model agreement) applies inward. I use and help productionize circuit-tracer (pip install circuit-tracer): the open pipeline for attribution, visualization, and interventions. The implementation descends from Decode Research's research release; my fork is where packaging, documentation, CI, and robustness work land so the library functions as a dependable research tool, not a one-off artifact of the papers.
CollaborationCircuits program
Round 2 is pre-registered in the CollaborationCircuits experimental plan: social-cognition tasks beyond classic false-belief setups, multi-agent representation interference, cross-model validation, and intervention studies (including circuit-informed steering). Round 1 exploratory runs on Qwen3-4B—lessons on chat mode, probe validity, and what replicated at larger n—live in archive/round1_explorations.
R&D trajectory: tooling and verification
Productionizing circuit-tracer is only the base layer. I am deliberately learning how to extend that system—APIs, graph semantics, evaluation hooks, compositional abstractions—so it stays pleasant for sustained research and engineering, not just one-off attribution runs. The longer arc is to narrow the gap between what we can observe in attribution graphs (low-level circuit behavior) and the high-level predicates we actually care about when many agents share a stack (who believes what, what must not be conflated, what counts as deceptive vs. cooperative modeling).
Eventually that should support something much sharper than vibe benchmarks: formal verification (or proof-carrying artifacts) tying high-level specifications to low-level circuit realizations—so we are not surprised by emergent misalignment when behavior-oriented metrics still pass. Civilization-scale coordination assumes the underlying model's internal social cognition matches the roles we assign in the Forum and the Collective Brain; if those levels diverge, we risk emergent unaligned agents even when the orchestration layer looks healthy. Making circuit behavior a first-class object for verification-oriented work is part of closing that hole.
How this connects to the rest of the series
Use this page when you need the mechanistic side of the same failures named on failure modes — belief propagation, mistaken precedent, errors that look like bad coordination but start as conflated agent states. Use alignment connections when the question is where steering and human override meet circuit-level targets.