Three resource tiers and two feasibility frontiers
Operator's framing, 2026-11-05, recorded as workshop analysis. The current workshop framing places it under the question of what should guide automatic search. This file retains the earlier candidate model for criticism; its proposed maximand and refuters are not adopted. Moving work between tiers is one possible return on search, not the assumed objective. Nothing here is adopted vocabulary; "tier" and "frontier" are workshop-local words.
The tiers
Commonplace spends three kinds of resource, and they differ in kind, not only in price:
| Tier | Resource | Can it be bought? | Reliability |
|---|---|---|---|
| 1 | Operator attention: judgment, authority, objective choice | No | High, scarce, slow |
| 2 | LLM work: interpretation, judgment without authority, generation | Yes | Fallible; needs verification |
| 3 | Classical compute: enforcement of defined transitions, checks, bookkeeping | Yes, cheaply | High within what it can express |
Tier 1 is the binding constraint. The library already measures improvement in its units: improvements are measured per human judgment, not human time, and progress in a self-improving system is which decision-bearing functions humans still supply.
The two boundaries are feasibility lines first
Economic thinking treats a boundary as a price comparison. The two tier boundaries are not. Each is a condition that must hold before any price applies.
LLM to compute. Code can take a transition only when the rule is defined and a reference settles what it permits: the symbolic criterion in representational form as revised by ADR 124. Below that line the error-correction asymmetry prices the move: avoided model checking and repair against implementation and maintenance, compared at a required end-to-end reliability. Above the line no price makes code feasible. The asymmetry note's own scope says so: the argument applies to explicit state and sufficiently specified transitions. The codify-versus-LLM heuristics hold the economic part, including this workshop's stake-and-odds section.
Operator to LLM. A model can take a judgment only when the result can be verified at the required level by something other than the operator: the boundary of automation is the boundary of verification, and error correction works only with above-chance oracles and decorrelated checks. Objective choice stays on the operator side by construction: objective revision is licensed from outside the objective, which is the epistemic-layer note's stated weak point for role 1.
Epistemic work is what moves both lines. Finding that a choice is stable enough to commit makes a transition codifiable (progressive constraining commits only after patterns stabilize). Finding that a judgment is checkable makes it delegable. This is the README's "convert into routine" reading stated per tier: conversion is the movement of work downward across a frontier.
Where economics operates
Only inside a tier, and in choosing which frontier move to invest in:
- Within tier 2. Which investigations to run, how much self-checking to buy, when to stop. The library grounds this structurally (search allocated before evaluation; a controller is tested by what it brings to stronger evaluation). The context bound gives tier 2 its own internal cost structure (two context boundaries, three regimes), which is the step-shaped cost measure the reconciliation adopted.
- Between frontier moves. Which bookkeeping to codify first; which operator judgment to delegate first. Priced by stake, odds and exposure, and by how much of the scarcer tier the move frees.
- Verification purchase. Whether a check runs in code, in a second model call, or in front of the operator. The asymmetry note's test form is the comparison: total cost at a required end-to-end reliability.
The exchange rate between tiers is asymmetric because tier 1 cannot be bought. A move that frees one operator judgment is worth much tier-2 work. A move that saves compute is worth little unless it also moves a frontier.
The maximand, stated per tier
Candidate, not adopted: Commonplace allocates to maximize work moved downward across the two frontiers at the required reliability, measured per operator judgment. The README's weak points become the objections: the frontier does not shrink by itself, since each conversion leaves the remaining work harder; "solved" is graded; goals are not beliefs. Rival maximands to keep beside it: explanatory reach, defeated claims repaired, user value delivered.
Consequences for an agent choosing what to do next
An agent sits in tier 2. Under this maximand its highest-value choices touch a frontier:
- Produce a decision in a form the operator can accept with one judgment rather than a reading.
- Prefer work that leaves a settled transition behind, as a candidate for the compute frontier.
- Do not spend tier-2 work re-interpreting what code could enforce.
Everything else is ordinary search allocation under the existing structural rules. Whether the agent's individual return guesses must be recorded to make this allocation criticizable, or testing the controller is enough, remains the open question from direction 1.
What would refute the framing
- A frontier move that is feasible but never pays: feasibility would then be insufficient as the first gate, and the price comparison would govern earlier than claimed.
- A case where buying more tier-2 work substitutes for an operator judgment without a verifier: the operator-to-LLM line would be a price, not a feasibility condition.
- Conversions that move work downward and still raise operator judgments per unit of work over time: the maximand would be measuring the wrong thing, and the "frontier does not shrink" objection would be decisive.