A position paper from the ABI/SOV program — completion, not refutation
Status: drafted 2026-08-12 for presentation to the Theoretical Neurobiology group. Registered as T158's sixth extension (THEORIES.md §XLVIII). Every empirical claim cited here carries a receipt in the program's registry (P48–P93) or findings ledger (F1–F32).
The Free Energy Principle is correct about the axis it describes. Biological systems do minimize something free-energy-shaped; perception and action do trade off against each other in reducing it; precision-weighting is real. Our own organism's fit dynamics — the distance between lived observations and each slot's expected geometry, driving both model update and behavior — are the free-energy story working exactly as advertised. Nothing below disputes the G-sort. The claim is that the G-sort is half of a two-sorted structure, and the formalism never knew it had a second half.
Our operator algebra factors every epistemic operation into a geometry component (boundary tests, feasibility sets, distances — computable from shape alone) and a ledger component (receipt histories, provenance chains, lifecycle states — append-only, earned). The factorization is load-bearing: three derivation constraints (shape-only computability, conservativity, provenance functoriality) each protect one kind of epistemic integrity, and each is independently forced by a failure mode (respectively: needing content to process unknowns; laundering imagined structure into belief; losing the ability to trace where a belief came from).
Now observe: every quantity in the FEP formalism lives in the geometry sort. Prediction error is a distance. Free energy bounds a distance-derived surprisal. Precision weights distances. Model evidence scores expected distances. There is no FEP variable whose value depends on a belief's lifecycle state (open, funded, closed, reopened, dormant) or its provenance (which lived episodes funded it, through which operations). FEP is a theory of the G-sort. The L-sort — the ledger — has no counterpart in the formalism.
Cross prediction error with explanatory state:
| Funded explanation | No explanation (open) | |
|---|---|---|
| Low prediction error | predicted and understood | accurately unexplained |
| High prediction error | model drift (world moved) | full epistemic gap |
FEP sees the main diagonal clearly — it is the free-energy axis. But the top row's two cells are FEP-identical: same predictions, same free energy, same precision. No variable in the formalism differs between an agent that predicts well and understands why, and an agent that predicts well by rote. The difference is entirely ledger-state.
Two consequences:
- Within a single receipted system, the cell cannot persist: in our architecture, sustained accurate prediction matures into funded explanation automatically (the closure lifecycle). Accuracy becomes understanding because the ledger converts it.
- Without a ledger, the cell is a permanent residence. A pure predictor — however sophisticated — is structurally confined to accurately-unexplained, with no interior signal that could reveal the confinement. This is our diagnosis of current large-scale AI, stated as structure rather than complaint: all geometry, no ledger; superb free-energy minimization; empty why-inventory; and no possible felt deficit, because the deficit is in a sort the system does not have.
We found this cell empirically before we had the frame: a retrospective audit (EX-0) showed procedural capability pervading our own archives with no by-name receipts — performance without account, between organs. The repair concept (an audit receptor: a per-region comparator of behavioral competence against ledger coverage) is registered as P92, with its calibration case fixed in advance: it must first detect the known instance.
"Expected free energy already covers this. EFE agents explore at low prediction error when posterior uncertainty is high — the epistemic value term is exactly a curiosity drive."
We agree EFE drives exploration at low surprise — over uncertainty the model represents. Epistemic value is an expectation over the agent's own random variables; the agent can be information-hungry only about variables it has. The accurately-unexplained cell is precisely the state where represented uncertainty is low — predictions are confident and correct — while the missing explanation is not a represented variable at all. Assigning epistemic value to an absent explanation requires an object FEP does not possess: typed absence — an open variable with connector geometry (relational position, boundary shape, neighbors) and no content. That object is our framework's founding primitive.
In one line: EFE is curiosity about your variables. The missing dimension is curiosity about your vocabulary.
This is not only conceptual. Two of our billed results are its empirical form: a separation theorem (information-gain and value-weighted selection provably choose different dimensions at fixed capacity — three independent instances measured), and a blind-spot result (3,600 evidence-driven inquiry events, zero aimed at regions emitting no signal — the correction had to be a constitutional reserved quota, not a term in any expected-value objective, because evidence-driven allocation structurally starves what produces no evidence).
FEP treats precision as estimated — hierarchically learned from prediction success. Our staged-fit experiments (P76, supported after four instrument iterations) measured what happens when precision is instead earned from receipt structure: consume evidence only where the already-funded prefix made it predictable (the "fringe" discipline). The decisive pattern, on one byte-identical stream with paired accountants: fringe-licensed consumption improved world-agreement, while complete-information consumption of the identical evidence at the identical dose actively damaged it — noisier predictions confirm coincidences, and coincidence-licensed updates import noise as signal. Processing order is epistemically load-bearing.
The contrast matters here because precision-estimated-from-success is exactly what is high in the rote cell — the agent predicts well, so it weights that evidence strongly — while precision-earned-from- funded-structure is low there, correctly: the agent has no account of why the evidence behaves, so its confirmations may be coincidences. Where FEP assumes (or globally estimates) precision, a ledger earns it locally, with a per-query audit trail.
FEP does not reward false explanation; it is indifferent to fundedness. Model evidence is the only currency, and funded accuracy and lucky accuracy pay identically. Bayesian model reduction prices complexity, which sometimes correlates with mechanism — but two models of equal evidence and equal complexity, one causal and one correlational, remain indistinguishable. Provenance is not parsimony.
The vulnerability this opens is motivational. The pull toward explanation is pre-epistemic: a counterfeit explanation feels identical to a funded one at the moment it discharges surprise. A system whose only check is eventual prediction failure will tolerate — not incentivize, tolerate — self-generated explanations that are accurate enough, often enough, in-distribution. Our framework blocks this with an architectural gate (conservativity: the ledger cannot be funded by imagined structure; the only return path from imagination to belief runs through the world) precisely because no learned self-regulation can be trusted to police a drive that rewards the feeling of understanding. And the economy independently prices degenerate explanations: an explanation that fires on everything carries no information per confirmation and starves under information-priced existence — falsifiability as rent arithmetic, with a computable optimal applicability rate.
Phenomena with receipts in our program that require both sorts:
- Earned bedrock. The first knowledge to stabilize is the most world-invariant content — and which content that is varies by world (measured: different world classes close different families first). Hinge-like certainty as selection, not architecture; the certainty hierarchy co-authored by organism and world.
- Calibrated assertion by construction. A generator that reads the ledger cannot assert what it does not hold; dormant knowledge speaks only in evidential past; assertion strength reads the survival record. Verified in acceptance: assertions equal funded entries exactly, with zero evidentiality violations.
- Three maintenance clocks. Beliefs go bad three ways — wrong (world moved), orphaned (world left), vacuous (belief says nothing) — each with its own detection mechanism, all receipts-driven. A free-energy account sees only the first.
- Motivational pathology with intact knowledge. Stock (held explanations, still serving) and flow (new invoices sensed and settled) are separable; flow can idle invisibly for years while stock pays the bills — motivational collapse that fitness and error rate cannot see. Registered with its rehabilitation protocol (restart from cheaply-closable, high-fertility explanations).
- Trust as a signed exchange rate on imported explanations, with surprise contagion along trust edges, decoupled authority as a provenance failure, and concentration as correlated fragility.
What a free-energy formalism would need to add — the L-sort axioms, minimally: (i) a lifecycle state per belief (typed absence through funded closure to reopening); (ii) append-only provenance from belief to lived evidence, surviving composition; (iii) a conservativity gate between imagined and funded structure; (iv) existence priced by information carried, not contact made. Each is independently motivated by a failure mode above; together they constitute the ledger.
We are not proposing FEP is wrong. We are proposing it is the geometry half of a two-sorted theory — and noting, with respect, that a theory of epistemics with no ledger sits in the accurately-unexplained cell about epistemics: it predicts epistemic behavior well while holding no account of what explanatory structure is. The completion is constructive, the empirical program is underway (registry P48–P93; findings F1–F32), and every mechanism named here is either implemented and battery-verified or registered with its falsifier.
Companion documents: the_ledgerless_economy (the engineering twin
of this critique — six pathologies of ledgerless learning systems as
symptoms of one missing organ), sov_formal_spec (the two-sorted
algebra and its conservation laws), structured_open_variables (the
founding proposal), stakeholder_theorems (the capacity-bounded
separation results). All in docs/sov/.