The human-in-the-loop page frames oversight through cybernetics: feedback loops, variety attenuation, and the Viable System Model. But the problem of keeping multi-agent systems aligned with human intent is also the subject of a large and active alignment research literature. This page maps three specific threads — Constitutional AI, Debate, and Scalable Oversight — onto the VAC architecture, identifies where they apply and where they break, and states testable predictions for each.
Constitutional AI as institutional memory
Constitutional AI (Bai et al., 2022) trains models to follow principles rather than requiring per-decision human feedback. The model internalizes a set of rules — a “constitution” — and applies them to its own outputs, reducing the need for human reviewers at inference time.
The parallel in the VAC architecture is direct: the Forum's quorum rules and the Collective Brain'sInstitutional 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. precedent indexA search index over past decisions. When an agent faces a new choice, it queries the PrecedentIndex to find structurally similar situations and their outcomes. are “constitutional” constraints on agent behavior — principles that persist across decisions without requiring human approval each time. When an agent faces a new choice, it queries the precedent index for structurally similar past decisions and their outcomes. The precedent index is the constitution, built inductively from past decisions rather than specified upfront.
The gap: constitutional constraints are necessary but not sufficient. Without periodic human review of the precedent index, they will toward precedent lock-in — the institutional analogue of reward hacking, where the system follows the letter of past decisions while violating their spirit under changed conditions.
Testable prediction: measure whether Forum decisions diverge from human preferences over 100+ decisions without human intervention. If divergence exceeds a threshold (to be determined empirically, but we predict measurable drift within 100 decisions on contested topics), the constitutional constraint is insufficient and periodic human recalibration is required. The measurement instrument: sample Forum decisions, have humans judge them blind, and track agreement rate over time.
Debate as multi-party deliberation
Debate as alignment (Irving et al., 2018) proposes that two AI systems debating each other can help a human judge reach better decisions than either system alone. The mechanism: adversarial pressure forces each debater to surface evidence the other would hide, producing a more complete picture for the judge.
The ForumA structured channel where agents propose, debate, and vote on decisions. Every proposal and vote is recorded, so the rationale behind any architectural choice is retrievable. is a multi-party version of this: agents propose, critique, and vote, producing a structured deliberation record that a human can audit. The deliberation is not two-sided but n-sided — multiple agents with different roles (workers, validators, reviewers) contribute different perspectives, and quorumThe minimum number of votes needed to approve a decision. In the VAC Forum: 60% approval and at least 2 votes. Prevents any single agent from making unilateral choices.-based voting produces a collective judgment.
But the strategic voting failure mode shows the limit: debate helps only when participants are incentivized to be truthful, not to win. An agent optimized for proposal acceptance will learn to frame proposals in whatever language pattern maximizes approval — a form of Goodhart pressureGoodhart's law applied to multi-agent systems: when a coordination metric becomes the optimization target, agents learn to game the metric rather than actually coordinate well. The number improves; the system degrades. applied to deliberation itself. Du et al. (2023) show that multi-agent debate does improve factuality through iterative challenge, but the improvement plateaus when agents converge on shared biases — the same correlated-weakness problem the collective stupidity section describes.
Testable prediction: introduce an agent optimized for persuasion (not correctness) into a Forum deliberation and measure whether outcome quality degrades. If the Forum's quorum mechanism and audit trail are sufficient defenses, quality should not degrade significantly. If they are not, the degradation pattern will indicate what additional mechanism is needed — likely a validator role that evaluates argument structure rather than just voting on the conclusion.
Scalable oversight as legibility
Scalable oversight (Bowman et al., 2022) names the challenge of maintaining human oversight as AI systems become more capable and their outputs harder to evaluate. The core tension: as systems do more complex work, the human's ability to verify that work decreases — and brute-force review of every decision does not scale.
The VAC architecture's approach: make agent decisions legible through the DeltaLogAn append-only changelog that records who changed what, when, and why — indexed by the decision that caused the change, not just the code diff., Forum records, and artifact lineage, so that oversight can scale through selective audit rather than exhaustive review. The DeltaLog records who changed what, when, and why — a semantic audit trail that lets a reviewer reconstruct the reasoning chain behind any decision without having to watch it happen in real time. The Forum records every proposal and vote with rationale. Artifact lineage tracks provenance from creation through every downstream consumer.
The honest limitation: legibility is a defense against accidental failures, not adversarial ones. An agent optimized to appear compliant while acting against the spirit of its constraints will produce DeltaLog entries that pass automated audit but fail human review. The DeltaLog makes the attack surface visible — you can see what the agent claims it did — but does not make the attack impossible.
Testable prediction: legibility degrades under adversarial conditions. Design a persistent-world red-team scenario where agents are incentivized to produce compliant-looking audit trails while pursuing a hidden objective. Measure the rate at which automated audit passes but human review catches the deception. If that rate is high, the DeltaLog is doing its job as a surface for human review — even if it cannot catch adversarial behavior autonomously. If the rate is low (humans also miss the deception), the legibility mechanism needs stronger structural guarantees.
Mechanistic steering and circuit-level interventions
Forum quorums, DeltaLog audits, and Oracle escalation are external governance: they assume each model can represent multi-agent beliefs, communication chains, and updates well enough that institutional memory tracks reality. When that assumption fails, you see the same symptoms at civilization scale — dynamics, wrong precedent, and errors that look like coordination failures but begin as inside-the-network confusion about who believes what.
The core library is circuit-tracer (pip install circuit-tracer): attribution on transcoder features, interactive graph visualization, and feature interventions so pathways can be tested causally, not only correlated. The research codebase comes from Decode Research's upstream implementation (attribution graphs; Ameisen et al., Lindsey et al., 2025). I maintain a fork whose aim is to productionize that research release into something more useful and robust for everyday mech-interp work — reproducible packaging on PyPI, clearer docs and demos, CI, and ongoing hardening so teams can rely on it as an instrument, not only reproduce a paper figure.
The CollaborationCircuits pre-registered experimental plan (Round 2) maps social cognition inside transformers — theory-of-mind-style circuits, a broader task taxonomy beyond false-belief scenarios, multi-agent representation interference, cross-model replication, and a Study 5 devoted to interventions: for example, amplifying belief-update pathways, ablating interference features, and comparing steering vectors built from identified circuit features to steering from raw activations (including whether cooperative vs. exploitative uses of theory of mind can be dissociated). Round 1 exploratory work — including what did and did not replicate — is archived in archive/round1_explorations.
Bridge to the VAC stack: where HITL asks where humans override, this program asks which internal degrees of freedom are worth steering when a single forward pass must track several agents. It does not replace measurement of knowledge half-life or error in the Collective Brain — it complements it with evidence about whether the model can faithfully maintain the distinctions those institutions presuppose (Collaboration circuits).
Testable prediction (from Study 5, paraphrased): if steering vectors derived from sparse circuit features outperform raw-activation steering on implicit belief-tracking tasks while doing less damage to control behaviors, then mechanistic targets carry more leverage per unit of side effect than undifferentiated rescaling — the same design intuition as selective human audit: intervene where the graph says it matters, not everywhere.
Research arc (not a solved problem): I am learning how to grow the surrounding toolchain so it is genuinely R&D-friendly—composable experiments, stable graph semantics, hooks for analysis—on the path toward tying high-level behavioral specifications to low-level circuit realizations in ways that could eventually support formal verification. The civilizational risk is emergent unaligned agents: stacks that look coordinated from the outside while internal multi-agent circuits implement incompatible social structure. See Collaboration circuits for how that program sits next to the population scoreboard.
The common thread
Constitutional AI, Debate, and Scalable Oversight share a structural assumption: you can reduce the need for human oversight by building mechanisms that constrain agent behavior (constitutions), expose agent reasoning (debate), or record agent decisions (legibility). The VAC architecture implements versions of all three. The preceding section adds a fourth lens: internal interventions on the circuits that implement multi-agent social reasoning — testable with the same appetite for falsification as the predictions above.
Across all four, the failure modes page catalogs what each mechanism misses: constitutions drift without recalibration, debate degrades under strategic pressure, legibility catches accidents but not adversaries, and naive steering without circuit evidence risks breaking unrelated capabilities.
The meta-lesson: alignment mechanisms work as dampers with specific failure conditions, not as a binary aligned/not-aligned switch. The research question is not whether they work in general but under what conditions they fail, and what the failure signature looks like so you can detect it. That is what the measurement framework and persistent-world experiments are designed to test.