Repository navigation
Prior Art
A curated reading of what this program actually owes to existing work, and — more
usefully — what it does not. The full ledger with citations is
PRIOR_ART.md
in the repository; Sources lists everything read.
This page is not Relation to the Field, which is the maintainer's to write. Placing a program against a literature is a judgment about what the literature says; this page only reports what was used.
The ledger classifies every entry by how much it constrains us:
Direct mathematical dependency — we use or inherit a theorem. Formal substrate — our objects are built inside a pre-existing framework. Conceptual dependency — the source shaped the theory or its terminology. Adjacent prior art — a close analogue, independently derived here, not logically imported. Verification target — our theory should eventually recover or subsume it; it is not an input. Historical motivation — motivated the question, contributes no current result.
Most entries are adjacent, and the distinction between adjacent and dependency is the one the program is most careful about in both directions. It is settled by looking at the proof, never by asking who derived it first: one entry moved from adjacent to dependency in September for exactly that reason.
Logical induction is the ground the enforcement construction is built on — the market, the traders, the market maker, and the inexploitability criterion are all inherited. The program's contribution there is a finite-time strengthening: a conformance bound at a date where the criterion gives only an asymptotic guarantee.
Imprecise probability — coherent lower and upper previsions, sets of probabilities, and the linear-feasibility characterization of probabilistic entailment — is what the credal state is. The regions the enforcement mechanism operates on are convex sets of probabilities, and their exactness comes from that reduction being a linear program.
Convex projection supplies the enforcement direction, and the compiler uses its defining variational inequality — a genuine dependency. Linear-programming duality is thinner than it looks: the finite-horizon infeasibility certificate has the shape of a Farkas pair and is called one, but its soundness proof is four self-contained lines; what actually leans on duality is the separate remark that the certificate is exact under a Slater point. Nothing there is novel and nothing should be presented as such; what is ours is what is being projected and who pays for it.
Brandom's deontic scorekeeping gives the shape of the whole thing: normative status as instituted by practice through tracked commitments and entitlements, with attributing distinguished from acknowledging. The conservation law — what was owed is satisfied, disposed of, or remaining — is a scorekeeping identity in his sense.
Truth-maintenance systems (Doyle 1979, de Kleer 1986) supply the identity-bearing justification structure — dependency-directed, with the environment/label separation that lets several stances be held at once.
Horty on reasons as defaults gives the treatment of a reason's weight and priority as normative content rather than machinery. Checked against his own 2006 text, his apparatus cannot express the program's authorized disposition, and for a deeper reason than a priority ordering failing to license the loser's release: default logic says what to conclude, not what is owed, so there is no ledger for a licence to act on. Proper scenarios are a function of the current theory and of nothing in the history of defeats; delete the record of which defaults lost and the answer is unchanged. The defeat theory here makes the future depend on that history through a named successor. That is a difference of subject rather than a defect in Horty — his is a static theory — and the novelty on this side is authored, history-sensitive answerability under defeat, not endogenous priorities, which Horty's variable-priority theories already have. What his framework lacks is participants: an authority enters as the content of a default, never as someone who holds standing. Two further findings: exclusionary reasons are challengeable in-system, in his own words, so his exclusion is closer to disposition than to settlement and the two do not collapse; and the program's answer/dispose split is Pollock's rebut/undercut distinction transposed from a belief's warrant to a debt's — a dependency, not a resemblance. Structured argumentation (ASPIC+) likewise has no successor and an unauthored attack relation, with reinstatement computed by the semantics rather than licensed. The 2012 book itself was not read; a claim about what the book says is not licensed by this.
What none of these supply: a commitment to inferentialist semantics. The architecture uses the scorekeeping shape and leaves the theory of meaning alone.
This is the section that matters, and it was written because the program's habit is to derive rather than to search — which produces exactness and also produces avoidable novelty claims. The test is not who thought of it first here. It is: would the repository's proof still stand if the external theorem vanished?
Interval feasibility — a real dependency. The exact condition for transporting obligations onto service dates within a deadline is a specialization of the Gale–Hoffman flow feasibility theorem, and more directly of Horn's 1974 conditions for preemptive scheduling with release times and deadlines. The repository's proof cites that theorem rather than proving it, so this is an inherited step, not a resemblance. It was independently arrived at and is very probably a rediscovery of Horn — but rediscovery does not make a citation optional, and an earlier version of this ledger claimed independence while the proof was doing the citing. A self-contained alternative exists and was deliberately not taken; the repository ledger says why.
Contiguity — the definition, not the theorem. The condition under which service-weighted convergence transfers to obligation-weighted convergence is contiguity of measures, in exactly Le Cam's 1960 sense. The definition is inherited and so is the name; the proof argues from that definition directly and invokes no lemma. The sufficiency direction is the standard and expected consequence, and must not be presented as new — the setting is different (no likelihood ratios, no asymptotic normality), and only the necessity result for triangular arrays might be.
Serve-oldest-first — a resemblance. The optimality of first-in-first-out is the same exchange argument as Jackson's 1955 earliest-due-date rule. Classical in substance, written out here in four lines that invoke nothing.
Several further places are marked literature review needed rather than guessed at, among them: whether the cost form of the timely-service criterion is classical; whether the join-semilattice conservation law has an antecedent in belief revision; and whether the persistent-wait theorem is a standard fairness argument in disguise. The third is the one most likely to already exist.
Carroll et al. on influenceable reward functions is the clearest external statement of the failure the program is trying to prevent: a system rewarded for satisfying preferences is rewarded for changing them. The repository reproduces its examples, and the criterion should eventually explain rather than merely match them.
The learning-normativity agenda is the nearest statement of the problem this line is working on, and the closest thing to a shared target.
A standing criticism, not a target. The objection that asymptotic inexploitability leaves finite behaviour almost unconstrained — for any given mistake there is a logical inductor making it until arbitrarily late — applies to anything this program claims on the basis of the criterion. The enforcement construction's finite-date bound is a partial response and has not been written up as one.
Three of the four results the program might extract as papers rest on mathematics with classical antecedents that were found after the derivation. The rule adopted at the September checkpoint is that every extraction candidate has its literature review completed before any novelty claim is made outside the repository — and that independent derivation is a fact about our process, never evidence about the literature.