Research map

drag to pan · scroll to zoom · click a node · esc backs out

The lines

normativity: What does learning from reasons require of a reasoner: which statics a normative seed forces, how a constraint acquires operative force in a market, and what the Normative Inductor, from seed through reasons, compile, enforcement and revision, establishes end to end?
legitimacy: What must stay true of a trajectory's record, verdicts and channels of influence, inside it and at its boundary, for a later judgment to count as the principal's own, and what does corrigibility consume from that?
corrigibility-landscape: Which correction relationships between humans and AI systems does the picture of corrigibility as eagerness to learn reach, and which failure stories does it leave standing?
decision-theory: Which decision theory for computationally uncertain agents, on the updateless side and in the synthesis of logical induction with bounded inductive rationality, is tractable, and which of its results bear on AI safety engineering?
stance-and-level: Where should the program work: at which level of computability, under which decision-theoretic stance, and does the computationally unbounded level represent the phenomena that matter?
abstraction-arc: Which work on abstraction, logical induction and agency gives the program a representation theorem for logical inductors, a decision theory set in that representation, and a notion of scale-free agency that covers cooperative equilibria?
enabling-work: Which repository infrastructure makes the trust research vettable and publishable?

How does the relationship between humans and AI systems go, and what can we do about it?

This site is a render of the argument map under graph/, the canonical index of a research program on how humans and AI systems learn from, correct and rely on each other. The map holds 2467 nodes: one plan with 7 lines, every registered claim as a formal result (178 registry entries), every priority item (141), every decision and every narrative page. Structure and human events are what you see here; the numbers the checker derives are a lens, shown only on the ranking page. Of the 249 formal results in the map, 85 are conjuncts of an informal claim; the rest are held by their programmes until someone reads them. Nothing on this site is typed by hand except this paragraph, checked against the state at every build.

Recent in the feed

2026-10-30 decision the site is one zoomable map: nested regions with semantic zoom, three views over one node set, fragments in a panel, and intake as prefilled issues (S5)
2026-10-30 decision corrigibility is a line of the program; deference-and-trust moves under it
2026-10-30 decision the site's domain is alignmentworkspace.com
2026-10-30 decision the weekly freshness audit, and its reporter's write scope
2026-10-24 round 2026-10-24-site-vertical-slice (enabling)
2026-10-24 decision the site as one vertical slice: five reference results connected by reading, a generated stdlib-only site that is never a second source, deployment to Pages behind an enumerated write-scope job, and one id, one URL (S1–S4)
2026-10-23 round 2026-10-23-normative-inductor-consolidation (normativity)
2026-10-23 claim inductor.rebuttal-run registered (lean-proved), the warrant Normative Inductor, run on the rebuttal of a still-active strength (Scenario B)
2026-10-23 claim inductor.contraposition-blocked registered (lean-proved), rebuttal withdraws the strength row that, kept, would have made the compiled system infeasible
2026-10-23 decision admission is provenance plus entitlement read off the account, verdict licensing is a separate authority, the compile returns a typed outcome that conserves its obligations, the Normative Inductor is one interface with two retained realizations, and the names
2026-10-23 decision the map's follow-up rulings: hypotheses are filed when read, the registry join is an edge, the large derived files leave the tree, one decomposition per line, a formal result sits under the claim it establishes, and items are docketed by programme (F1–F6)
2026-10-23 decision the map's second follow-up rulings: the digest has one shape, a frame holds what it is about and its roots are its goals, a registered result bears on an item rather than answering it, and an unconfirmed want weighs nothing (G1–G4)

The whole feed.