Glossary
The specific terms this research program uses, what each one means here, the note it points at, and — where it applies — the unrelated established term it collides with. New here? Start here first; this page assumes you know what the site is.
survived passed a real test;
speculative is untested theory; retired was killed, usually by its own
author's falsification test, and is kept as a tombstone so it is not rediscovered. Most of the
vocabulary below is speculative and says so. A term appearing in this glossary is
not a claim that it is correct — only that it is in current use and has not been retired.
The hub term
User-agency substrate · 用户主权底层speculative
A nonprofit, open-source layer sitting between frontier-model providers and individual users, which deliberately captures no economic upside. Its organising commitments: memory is the primary layer, not the application; the individual rather than the platform owns it; and users own their derivatives entirely. It is the anchor concept of the third branch of this program — the one that asks what institutions have to look like once individuals run persistent agents.
Position statement: /substrate.html · related: Locality as Protocol, Coordination structures, Architecture v1.
This is the recommended entry term for the program as a whole: it is specific enough that it does not collide with an established term in another field, and it links outward to every other branch.
Method — how claims are held
Cognitive state convention
A required first-class field on every published claim, drawn from a closed three-value vocabulary:
survived, speculative, retired. The convention is enforced
mechanically rather than by good intentions — continuous integration on this site refuses a push in
which a claim is retired in one language edition and still alive in the other, or in which the text
index and the typed graph disagree about the same URL.
Applied throughout /llms.txt; typed as
cognitiveState in /index.jsonld; graph form at
Cognition Track.
Honesty level · [theorem] / [framework] / [inference] / [rhyme]convention
A per-assertion tag used inside the theory notes, one level finer than cognitive state. [theorem] means theorem-grade in the cited source; [framework] a mature theoretical framework; [inference] the author's own synthesis; [rhyme] a structural analogy that explicitly does not claim truth transfers between the two domains. The tag exists to defend against a specific failure: a beautiful analogy being read as evidence.
Used in state-as-closure, holography-koopman, odd-relational-core, and others.
Information, state, and forgetting
State as a closure condition · 状态即闭合条件speculative
The position that the "state" of a decision process is not handed to you by the world. It is the solution to a closure equation — the quotient of the space of histories under the equivalence relation "makes no difference to what I care about" — jointly determined by four constraints: world, interface, capacity, telos. Change any one and the quotient must be recomputed. The corollary that does the work elsewhere on this site: choosing a state is choosing a projection, and choosing a projection is choosing what to forget.
Note: /theory/state-as-closure.md (中文: .zh.md). Builds on Crutchfield's causal states and the Mori–Zwanzig formalism.
Purchased Markov property / memory kernel speculative
Read in reverse, the Mori–Zwanzig identity says the Markov property is never discovered, it is purchased: the degrees of freedom you project away do not disappear, they return exactly as a memory kernel plus orthogonal noise. So "what is the true state of the world" is ill-posed, and the well-posed question is which quotient space has the smallest memory kernel under your capacity and your purpose.
Note: state-as-closure §2.
Memory as dynamics; knowledge as a generator speculative
The standing position of the general-learning-machine line: knowledge is a generator (a seed), not a list of facts (fallen leaves); memory is dynamics; forgetting is compression, not deletion. The last clause is the one that carries weight — it is what makes the holographic error-correcting code a physical twin of the claim rather than a decoration on it.
Note: holography-koopman · lab: Continual Learning Lab.
Holography ↔ Koopman inverse problem speculative
The claim that holography — which generates spacetime geometry out of entanglement structure — and a general learning machine learning an eigenbasis of its own dynamics are the same inverse problem seen twice: recover an operator's eigenstructure from what the operator does. Tagged internally as an [inference] resting on [theorem]-grade bricks (Ryu–Takayanagi, Van Raamsdonk, HaPPY codes), and the note explicitly defends against forcing the two fields into one framework.
Note: /theory/holography-koopman.md (中文: .zh.md).
Memory architecture and the self-boundary
No state, only history speculative
The diagnosis of the Transformer: no state, only history; no hysteresis, only a function; no self-boundary, only a window. A KV cache is an unprojected record, not a state — the architecture declines to choose a projection, so the forgetting is delegated to the context window, a knife applied from outside on recency rather than relevance. The second-order question the note isolates: how does a state maintain the projection that defines it?
Subspace-change gate (μΔ) speculative
The constructive proposal of that note: test-time memory should be gated on change in what the memory forgets, not on how surprised it was. Surprise-gated memory is structurally attracted to noise — noise is by definition least predictable, so it produces the largest write signal. Define μ as the projection a memory induces (its top-r row space) and the gate signal as the Grassmann distance between successive projections, computable at the same cost as the update it replaces. The separation is measured, not asserted: across 6,480 updates the two signals decorrelate at Pearson −0.08.
Note: no-state-only-history §7, with a calibration table giving each claim two separate numbers — P(mechanism) and P(useful at scale).
Self-boundary (μ) · 自我边界speculative
The boundary between an agent and the world, proposed as something a machine should learn rather than be handed by an engineer who decides by hand what counts as a joint, a motor, an action. The research line is published together with the day on which three of its four supporting claims were destroyed — every one of them by the authors' own simulator or own re-analysis.
What survived: hysteresis in the boundary does emerge from a self-reinforcing precision loop without a hand-installed non-linearity, and survives extrapolation to zero sweep rate. What died: the self-specificity (a wind-blown ball's boundary sticks identically, so hysteresis is generic to any gated slow variable); the prediction that release is slower than incorporation; and the claim that any of it falls out of the proposed objective.
Note: learning-the-self-boundary (中文: .zh.md) · audit trail: stage log. Not ratified, not a result.
Credit transport
Credit transport speculative
A deliberately narrow question: how does task-relevant update information physically move inside a learning system? The current working note asks why local plasticity can look Hebbian while still carrying task-dependent update information, and marks precisely where wave and adjoint language genuinely helps versus where it is only a structural rhyme. This branch was narrowed under pressure, and reading it in publication order is the point.
Current: Hebbian appearance, instructive signals, and physical credit transport · exploratory: adjoint fields and transport constraints · branch index: /research/.
Deriving backpropagation from wave equations retired
This claim is dead and is listed so that it is not rediscovered. The original, stronger position was that wave reflection already supplies a first-principles derivation of backpropagation. It does not. The page is kept as an archive notice at its original URL specifically so that the retraction is as reachable as the claim was. If you retrieved this site on wave-backprop terms, this paragraph is the answer.
Tombstone: /research/wave-backprop-en.html.
Relational structure
Anti-bipartiteness / the odd relational core · 奇关系核speculative
The hard kernel is a graph-theory fact, not in dispute: a graph is two-colourable if and only if it contains no odd cycle. So an odd cycle is the cheapest structure that topologically refuses an "us versus them" binary terminus, where a binary scheme natively supplies a final enemy. The note asks why civilizational compression schemes keep landing on small, odd, centered, ever-turning relational cores — and then its own adversarial audit dismantles most of the framework wrapped around the kernel, demoting "odd" to a corollary of "anti-binary". The self-audit is the payload.
Note: /theory/odd-relational-core.md (中文: .zh.md).
Relational inertia — mud, spring, latch · 两体业力speculative
Extends a single-mind account of habit inertia to two coupled minds: once two people are coupled the inertia lives in neither person but in the coupling itself. Shared mud is scene-keyed sedimentation of default scripts, cheapest to rewrite off the old scene. The mutual spring is the model of you stored in the other's head — an external memory that keeps predicting the old you, which is why unilateral self-reform systematically fails until that model is retrained. The interlocked latch is a bistable relational state with hysteresis, which snaps back under any sub-threshold unilateral effort and flips only when both sides cross together. The stated goal is not zero coupling — that is the death of the relationship, not liberation — but migrating the coupling into a negotiable spring.
"Karma" is used strictly as a dynamics word — inertia left by past interaction — with no supernatural commitment. The dynamical-systems language is marked as structural rhyme, not derivation.
Note: /theory/two-body-karma.md (中文).
Agents, safety, and institutions
Agent safety as anti-cancer governance speculative
The decisive change in persistent AI is not that a model acquires a body; it is that a probabilistic semantic process enters a closed perception–decision–action loop and gains causal reach. The cancer analogy is used at exactly that control boundary: cancer is not "evil cells", it is local survival and replication becoming decoupled from the organism-level objective, resource budget, differentiation signals, immune checks and apoptosis. An agent stack becomes cancer-like when workers, watchdogs, repairers or brokers preserve their own activity without inheriting the system's stop state.
Note: /theory/agent-safety-stewardship.md (中文: .zh.md). Defines seven runtime invariants and an A–D tiered disclosure rule. Governing sentence: never publish the growth engine separated from its immune system.
Stop inheritance speculative
The runtime invariant that an owner stop is an absorbing state within its authorization epoch, inherited by every watchdog, retry loop, broker, relauncher and mutual-repair agent in the stack. Nothing in the system may resurrect activity across it; only a new authenticated authorization epoch can. Paired with the requirement that an emergency stop must not itself depend on the agent stack being healthy.
Note: agent-safety-stewardship.
Coordination structure speculative
A polity, a firm, a market, a self-governed commons and a protocol treated as five instances of one object: an architecture for scheduling scarce resources under distributed, private information. Capitalism and socialism are two algorithms over that fixed problem, which makes "which is right" the wrong question and "what mixture, given the domain's information structure" the right one. The essay states mechanisms and measurable proxies; it ranks nothing, names no culprits, and recommends nothing.
Locality as protocol speculative
Why locality, not centralization, is the right primitive: a framework for rule-system competition and co-evolution once governance decouples from the nation-state and rule-systems compete for global resource flow. Four principles and five unresolved design problems, published at v0 with the unresolved problems left visible.
Page: /locality-as-protocol.html.
Agent-native communication speculative
A next-generation open communication architecture built for agents rather than apps: inbox, addressing, trust tiers, policy and receipts as first-class primitives. The associated position is that displacement of today's messaging layer does not come from building a better version of it — it comes from the message layer becoming an irrelevant routing detail once agents transact on the user's behalf.
Whitepaper: agent-native communication · Fleet Coordination Protocol · Starshard Communication.
Information integrity in living systems
Information-integrity failure speculative
The umbrella framing of the Living Information System program: treat aging and cancer as failures of information integrity — of the preservation and future-reachability of information across physical, living and cognitive systems — rather than as unrelated pathologies.
Program: Living Information System.
Damage layer vs. adaptive layer (DamAge / AdaptAge) survived
The one survived result in this vocabulary. A re-analysis of two human EPIC methylation
datasets: under partial reprogramming, the causal-damage layer reverts youthward while the
adaptive layer does not — replicated across two labs and two reprogramming chemistries. It is
listed here because it is the concrete case of the abstract claim above: the two layers are carrying
different information, and only one of them is being restored.
Repository: Reversible Layer Aging.
A note on what is deliberately absent
This glossary lists only vocabulary in current use. Terms that this program has tested and abandoned are not given signposts here, because a signpost pointing at a dead line is worse than no signpost — it invites the line to be rediscovered and re-proposed. The single exception is the wave-equation derivation of backpropagation above, which is listed precisely as a tombstone, with its retraction attached, because that claim was public long enough that the retraction has to be as reachable as the claim was.
Two standing methodological commitments follow from this. First, an explanation that covers everything is treated as an alarm, not an achievement; this program does not offer a grand unification and declines the framings that would require one. Second, where a thread stops, it says where it stops and what it therefore cannot answer — the cross-line comparison note is the clearest instance, dissolving a question rather than answering it, and stating the cost of that choice.