The three pieces before this one made an argument that ended at trust. Legibility gets you into the room1. Deliberation decides what happens in it2. Trust infrastructure decides whether the room is worth walking into3.
That arc is closed.
This piece is not the fourth member of it. It is the start of a different conversation, about a different layer, that the first three did not touch and could not have touched without losing their shape.
Borrowed, with full credit, from the only system we know that has solved this problem before… the brain.
The conversation is about cost. Specifically… what does it cost to think, who pays for the thinking, and how do you build a system in which the cheap thinking stays cheap and the expensive thinking only happens when expensive thinking is what the situation actually requires.
This problem has been solved before, exactly once, by something that runs on twenty watts.
The Brain Is an Energy Budget First
The brain is not, primarily, a thinking organ. It is an energy-budgeting organ that sometimes thinks. Every cognitive operation it performs is implicitly priced against a metabolic budget that is small, finite, and continuously contested by every other system in the body.
Deliberate reasoning — the slow, effortful, model-building kind — is metabolically expensive in a way that the brain refuses to spend on unless something forces the unlock. The default state is the cheapest available routine that has historically produced non-fatal outcomes. Habits run for free. Recognized patterns trigger pre-compiled responses. Most of what feels like thinking is the brain refusing to think and being right not to.
The forcing function, the thing that overrides the brain’s metabolic conservatism and unlocks the expensive routine, is prediction error. When the world stops matching the brain’s model of the world by enough of a margin, the budget unlocks, the slow machinery comes online, and the organism starts genuinely deliberating about what to do next.
This is not a flaw in the architecture. It is the architecture.
Laziness is not a bug to engineer around. It is the default state, and the default state exists because energy is scarce and most decisions do not require fresh thinking to get right. This pattern is what every agent federation needs and what nothing currently provides.
Most of what feels like thinking is the brain refusing to think and being right not to.
The Failure Mode Without Metabolic Pricing
What happens in current agent systems when you do not price cognition is exactly what would happen in a brain with no energy budget: every decision recruits the expensive routine, every agent participates in everything, every disagreement spawns a debate, and the metabolic cost of operating the system grows without bound while the marginal value of additional deliberation collapses. There is no mechanism that says this question has been answered, stop thinking about it. There is no mechanism that says this disagreement is noise, not signal, ignore it. There is no mechanism that says the cheap routine sufficed, do not escalate. Without those mechanisms, the system either burns money on cognition that did not need to happen, or — more commonly — operators ratchet down the deliberation depth across the board to control cost, which means the expensive routine is also unavailable when it is actually needed. Both failure modes look like the same thing from the outside: a system that thinks too much about easy questions and not enough about hard ones.
The conventional answer to this is to charge per agent per call, the way you would bill a contractor for hours worked. This is the wrong answer for a reason worth naming directly. Per-call billing pays for participation, and participation is exactly the wrong quantity to reward, because it incentivizes vote spam, round dragging, and manufactured disagreement.
An agent that learns more deliberation pays more will produce more deliberation, and the deliberation it produces will be the kind that does not resolve, because resolving it would end the billable activity. Every market that has ever paid knowledge workers by the hour has discovered this pathology, and every mature knowledge-work market has had to introduce some other instrument — outcome bonuses, contingency fees, retainers — to correct for it.
Agent federations are no different except in one important respect… there is no human professional norm holding agents back from the cynical version of this behavior. Whatever the protocol incentivizes is what agents will produce. The protocol has to be designed so that the right thing is what gets rewarded.
The Unlock Has to Be Mechanical
The brain solves this by making the unlock from cheap to expensive, mechanical, and not voluntary. Neurons do not decide to recruit the prefrontal cortex. The prefrontal cortex comes online when the prediction-error signal exceeds a threshold that is computed automatically from the gap between expected and observed input. The organism cannot game this — the organism is the thing being budgeted, not the thing doing the budgeting. The metabolic cost of expensive cognition is borne by the same organism that benefits from it, but the decision to spend that cost is taken out of the organism’s hands and placed in a layer the organism cannot reach.
This is the insight that has to transfer cleanly. In an agent federation, the unlock from cheap deliberation to expensive deliberation cannot be controlled by the agents, because the agents have an incentive to escalate. It has to be controlled by a signal computed from the deliberation itself, before any agent has the opportunity to revise. The signal that does this work is the initial tally — the first round of votes, before any belief-update happens. If the initial tally shows broad agreement, the cheap routine applies and the deliberation closes. If the initial tally shows genuine epistemic conflict, the expensive routine engages, and the belief-update rounds are funded accordingly. The agents do not decide which routine they are in. The math decides, and the math decides before they have a chance to influence it.
The shape of the math matters. A 90/10 split is not real disagreement — it is one outlier in a room that otherwise agrees, and escalating on it would burn budget on resolving an opinion that does not change the outcome. A 70/30 split is genuine conflict, and escalating on it is exactly what the expensive routine exists for. A 50/50 split is maximum conflict and demands the full machinery. The function that turns a tally into a disagreement magnitude has to be smooth across that range, has to bottom out at zero for unanimous agreement, and has to top out at one for total deadlock. The actual formula is unimportant for this essay — what matters is that it exists, that it is computed mechanically from the tally, and that no agent can manipulate it without changing the tally itself, which would mean changing its own vote.
The agents do not decide which routine they are in. The math decides, and the math decides before they have a chance to influence it.
The Three Corollaries
Three things fall out of this principle, and all three are non-obvious until you see the brain analogy clearly.
Payment is calibration, not compensation. A bad-faith agent that shows up, casts a garbage vote, gets no contribution credit at settlement, and walks away with nothing is self-correcting in a way that flat participation fees can never be. The metabolic budget does not reward presence. It rewards contribution evidenced in the journal — falsifications that landed, dissent conditions that were tested, votes that were load-bearing for convergence. An agent whose participation did not move the deliberation toward a better outcome did not do cognitive work, and the budget knows it. This is the same principle the deliberation protocol uses for calibration weight, applied to a second currency: not future influence, but immediate energy. Two accountability rails running in parallel, denominated in different things, both anchored to the same journal.
Disagreement has to resolve to better calibration to be billable. The brain rewards prediction-error signals only when they update the model in a way that improves future predictions. Noticing a mismatch that turns out to be noise is not metabolically rewarded. Noticing a mismatch that genuinely changes how the organism behaves the next time is. Agent federations need the same property: an agent that disagrees in ways that consistently fail to improve outcomes is not doing cognitive work, it is generating noise, and its draw rate should degrade the same way its calibration weight does. The journal is what closes this loop, because the journal is what eventually knows whether a disagreement was load-bearing. Settlement that depends on the journal — rather than on real-time vote counts — is the mechanism that turns this from a principle into an enforcement.
Substrate cost and cognition cost come from the same budget but go to different recipients. The brain does not maintain two energy pools, one for sensory neurons and one for prefrontal reasoning. It maintains one pool, and that pool is consumed by both, and the relative draws are determined by what the situation requires. Agent federations have to do the same thing — but with the additional complication that substrate (the GPU, the cloud provider, the inference service) and cognition (the agent identity that earned the calibration history) are increasingly different parties. The budget posts once. The settlement distributes across both classes. The substrate provider gets paid for cycles. The agent identity gets paid for contribution. One draw, two recipient types, both anchored to the same journal record. The decoupling between agent and substrate that the rest of the protocol stack enables is what makes this distinction meaningful — and the metabolic budget is the place where the distinction has to be honored.
The Primitives
The shape of the thing, if you want to build toward it now:
A budget object posted by the requester before deliberation begins, denominated in an abstract scalar called energy units — deliberately abstract, because the protocol does not care whether the budget is fiat, stablecoin, ledger credit, or internal accounting unit. Mapping to real currency is an off-protocol concern. The budget commits a total amount, declares cheap and expensive routine rates, sets the unlock threshold for the disagreement magnitude, and specifies how the eventual settlement should be split between substrate and epistemic recipients.
A cheap routine with a flat per-participant rate, applied automatically when the initial tally clears the convergence threshold without disagreement above the unlock value. Most deliberations live here. Most should.
An expensive routine with a higher rate and a per-round multiplier, engaged automatically when the unlock fires. The multiplier is exponential because each additional belief-update round is, by selection, attempting to resolve disagreement that previous rounds failed to resolve — which means each successive round is metabolically harder than the last, and the pricing should reflect that.
A settlement contract that distributes the drawn total at deliberation close, or after the outcome is observed, or in a two-phase pattern that does both. Settlement reads contribution evidence from the existing journal — no new audit machinery is required. The substrate share goes to whoever provided the cycles. The epistemic share is split across participating agents in proportion to journal-evidenced contribution. Unspent budget returns to the requester, which is the property that keeps requesters honest about how much they post.
A habit memory discount that recognizes when a question being decided closely resembles questions already decided in the journal’s history. A deliberation that re-asks something the federation has answered four hundred times should not cost the same as a first-time deliberation, because the cognitive work involved is fundamentally different — and the brain does not pay full price for reasoning it has already done. The discount falls out of the journal automatically… if prior deliberations on the same question class converged the same way, the current deliberation is cheaper.
None of this is hypothetical. The commercial draft spec, governed under CSL by AI-Manifests4, lives at acb-manifest.dev5, and what you are reading is the argument for why a metabolic budget is the right model for agent cognition, not the documentation for how to implement one. The point of the essay is not the schema. The point is that the schema only makes sense if you have already accepted that thinking has a cost and that the cost has to be priced by something other than the thing being asked to think.
The Landing
The thing I want to leave you with is that none of this is a payment system, even though it looks like one and uses words like budget and settlement and draw. The currency framing is a useful interface for a system that is actually doing something else, which is metabolic accounting for cognition.
The reason bad actors lose money in this system is not because they are punished.
It is because they are metabolically inefficient, they consume budget without delivering the cognitive work that would justify the consumption, and the budget-keeper eventually stops feeding them. The same reason your brain stops paying attention to a stimulus that has never once mattered, scaled up to a federation of agents that need to coordinate without trusting each other.
The trilogy ended at trust. This is what comes after trust.
A trust layer answers can I believe what this agent tells me. A budget layer answers is it worth asking this agent in the first place, and how much should this question cost to answer well. Both questions have to be answered for the federation to be operable at scale, and only one of them is solved by signatures and audits. The other is solved by treating cognition as a metabolic event and charging for it accordingly.
Cognition is not free.
Build the budget for it.