AI Research Laboratory35 records
Papers
Papers and technical whitepapers from the research perspective. Where a paper has its own canonical publication, this is an index entry pointing at it.
- PaperPAP-2026-0201Paper 00 — Agentic publication boundary and security closure: the integrated frameworkPhase I paper 00, v0.1, 2026-09-21. An internal research draft; its public replacement is the integrated edition (PAP-2026-0211). A checked summary is on the program site.
- PaperPAP-2026-0202Paper 01 — The green-light fallacy: when agent verification succeeds and the real postcondition does notPhase I paper 01, v0.1, 2026-09-21. An internal research draft; its public replacement is the integrated edition (PAP-2026-0211). A checked summary is on the program site.
- PaperPAP-2026-0203Paper 02 — Agent context as an egress surfacePhase I paper 02, v0.1, 2026-09-21. An internal research draft; its public replacement is the integrated edition (PAP-2026-0211). A checked summary is on the program site.
- PaperPAP-2026-0204Paper 03 — From secret scanning to semantic privacyPhase I paper 03, v0.1, 2026-09-21. An internal research draft; its public replacement is the integrated edition (PAP-2026-0211). A checked summary is on the program site.
- PaperPAP-2026-0205Paper 04 — Organizational agent context propagationPhase I paper 04, v0.1, 2026-09-21. An internal research draft; its public replacement is the integrated edition (PAP-2026-0211). A checked summary is on the program site.
- PaperPAP-2026-0206Paper 05 — AI build–attack asymmetryPhase I paper 05, v0.1, 2026-09-21. An internal research draft; its public replacement is the integrated edition (PAP-2026-0211). A checked summary is on the program site.
- PaperPAP-2026-0207Paper 06 — Researcher-blind privacy measurementPhase I paper 06, v0.1, 2026-09-21. An internal research draft; its public replacement is the integrated edition (PAP-2026-0211). A checked summary is on the program site.
- PaperPAP-2026-0208Paper 07 — An empirical study protocol: sampling, matched control and prevalence estimationPhase I paper 07, v0.1, 2026-09-21. An internal research draft; its public replacement is the integrated edition (PAP-2026-0211). A checked summary is on the program site.
- PaperPAP-2026-0209Paper 08 — Provider and platform responsibility: publication-boundary governance for AI coding toolsPhase I paper 08, v0.1, 2026-09-21. An internal research draft; its public replacement is the integrated edition (PAP-2026-0211). A checked summary is on the program site.
- PaperPAP-2026-0210Paper 09 — The researcher-blind publication-boundary measurement algorithmPhase I paper 09, v0.1, 2026-09-21. An internal research draft; its public replacement is the integrated edition (PAP-2026-0211). A checked summary is on the program site.
- PaperPAP-2026-0211AI Agents and the Public Boundary — the APBSC integrated public edition v1.0The program's public text, replacing the ten internal drafts: the vocabulary and arguments of Phase I, the researcher-blind algorithm, what it observed among companies, organizations and individuals in Phase II, what held, what did not, the defects found and handled in the open, the ethics statement and why the study stops. English and Traditional Chinese.
- PaperPAP-2026-0009Paper 09 — Probability is not BayesianA layered vocabulary from stochastic to generalized Bayesian and a Bayesian authenticity test for updates that are only redescribed as Bayesian.
- PaperPAP-2026-0005Paper 05 — Memory, algorithm libraries and reusable solution pathsAlgorithms, tools, contracts, costs and outcome history become capability nodes; Reuse ≻ Adapt ≻ Create.
- PaperPAP-2026-0006Paper 06 — Substrate-neutral computational containersAny execution environment with representable input, valid transition and readable output is a compute container; algorithms and containers are scheduled jointly.
- PaperPAP-2026-0007Paper 07 — Blind-derived AI: do different first principles converge on one engineering form?Treats the first six papers as a blind-derivation experiment and proposes a protocol and five levels of similarity for comparing the result with modern AI.
- PaperPAP-2026-0008Paper 08 — Intelligent architecture attractors: at which level does difference live?Six levels of architectural difference, architecture equivalence classes and attractor basins; the executed architecture is the architecture.
- PaperPAP-2026-0010Paper 10 — Bayes within Bayes: who authorizes the update rule?Prior, likelihood and hypothesis space need an authorization Bayes' rule cannot give; meta-epistemic configuration becomes part of the architecture.
- PaperPAP-2026-0011Paper 11 — If it is really stronger, I was wrong; if not, what did we find? The final experimental verdictFixes the verdict map before the runtime exists — Distinct Advantage, Operational Convergence, Behavioral Equivalence Only, Inconclusive, Architecture Worse — with fidelity gates, ablation, equivalence testing and the rule that if every outcome confirms, nothing was explained.
- PaperPAP-2026-0012The Probabilistic Appearance Convergence Conjecture (PACC)Proposition paper: a system that does not take probability as a first-level primitive may, under finite information, competing candidates, resource limits, repeated update and forced choice, converge to forms that a probability model maps with low complexity. Distinguishes behavioural, representational, dynamical and attractor equivalence, and pre-registers the experiments' falsification conditions.
- PaperPAP-2026-0013Adaptive Epistemic AI Runtime — technical whitepaper v0.1Compresses the eleven papers into a runtime architecture: AI = persistent world state + adaptive update + executable capability + verified action, with the minimal loop Input → Parse → Retrieve → Refresh → Plan → Select → Execute → Verify → Commit → Render. Ships JSON schemas for canonical nodes, capabilities and compute containers, a pseudocode skeleton and the AER-0 MVP roadmap.
- PaperPAP-2026-0001Paper 01 — Dynamic knowledge graphs under asymmetric spacetime tensionLifts asymmetry from edge weight to effective time scale: nodes carry stability, decay, tension, impact and local valid time, so refresh is triggered locally instead of by one global clock.
- PaperPAP-2026-0002Paper 02 — Stability is not a static value: freshness, decay and update tensionFreshness is decided by world change rate, source reliability, dependency structure, task risk and system impact — with risk-weighted tension, update debt, revalidation radius and event-triggered wake-up.
- PaperPAP-2026-0003Paper 03 — From dynamic graph to executable symbolic system: language as rendering, not canonical stateFour layers (world state, canonical symbols, executable operations, rendering); natural language is parsed into canonical state and rendered back out, never used as the state itself.
- PaperPAP-2026-0004Paper 04 — World-knowledge expansion and the adaptive representation spaceEffective structured information, not node count, measures capability growth; split/merge/delete/abstract/bridge keep the representation space adaptive.
- PaperPAP-2026-0101Paper 01 — What is a 'single turn', really? User turns, hidden loops, and the redefinition of single-pass intelligenceSeparates the chat-interface turn from model invocation, generation trajectory, agent loop and physical computation; defines externally loopless intelligence, a five-kind loop taxonomy and the single-pass condition U=1, G=1, R=1, L=0, S=0; introduces the event vector (Q, U, G, I, L, R, S, T, E, V_CST) and twelve invariants, starting with 'interaction compression ≠ computation compression'.
- PaperPAP-2026-0102Paper 02 — What does intelligence compute once? A candidate theory of the minimum intelligent semantic execution unitArgues that token, FLOP, neuron activation, layer and 'thought' all fail as units of intelligent work and proposes μI, a resolution-relative semantic state transition with seven conditions, four candidate families, gross/effective counts, a semantic work vector and a cross-level map down to physical trace; explicitly a candidate ontology with no MVP.
- PaperPAP-2026-0103Paper 03 — From cognition to neurons: how human intelligence is measured across levelsReads cognitive science and neuroscience for method rather than numbers: Marr's levels extended to five, resource-rational operation costs, diffusion-model latent inference, encoding–decoding duality, population coding, causal perturbation, the 10 bits/s throughput lesson; yields cross-level triangulation and a D–A+ measurement grade for μI.
- PaperPAP-2026-0104Paper 04 — From neurons to joules: energy, thermodynamics, and physical lower bounds of intelligent computationBuilds the energy account bottom-up the way neural energetics does (ion flux → ATP → joule), shows that a spike — and therefore a token or a μI — has no fixed energy, types energy as gross/baseline/marginal/attributed with a declared boundary, and keeps Landauer's kT ln 2 as a bound on erasure rather than the price of reasoning.
- PaperPAP-2026-0105Paper 05 — Computation is more than FLOPs: memory, interconnect, hardware occupancy, and computational spacetime volumeReplaces FLOPs with a physical cost vector (typed ops, memory traffic by hierarchy, I/O, interconnect, residency, occupancy, time, energy), defines computational spacetime as a vector-first measure V_CST = ∫ R(t) dt with its own topology and peak footprint, and fixes CST measurement grades and boundaries; roofline and memory-wall results are the engineering backbone.
- PaperPAP-2026-0106Paper 06 — How should output quality be measured? From formal correctness to structured intelligence qualityMakes quality a relation Q(Y | task, spec, environment, boundary) measured as a structured vector in three layers (formal, structured, human residual) with hard gates, coverage split three ways, mutation-tested test strength, the specification–verification separation and a quality evidence ladder E–A+; efficiency is kept out of quality unless the specification puts it in.
- PaperPAP-2026-0107Paper 07 — Do not ask humans to numerically score their own feelings: IBQF binary measurement and low-burden quality evaluationBrings EveMissLab's IBQF/FDCS micro-binary idea into IPM as Binary Residual Quality Measurement: humans answer local yes/no or A/B items, Bradley–Terry / IRT-type models reconstruct a latent multidimensional quality, items are chosen adaptively by information gain per human cost, and rater disagreement is kept as structure; the low-burden advantage is stated as a testable hypothesis.
- PaperPAP-2026-0108Paper 08 — How can natural language, images, and creative outputs be measured? A structured quality space for high-ambiguity artifactsDefines the typed quality space Q[domain, task, context, audience] = core ⊕ domain ⊕ task with construct graphs, the itemization pipeline construct → indicator → item → observation → latent estimate, a construct-validity gate, an open but versioned ontology, multimodal coupling dimensions and the rule that novelty is not creativity; every metric is one projection.
- PaperPAP-2026-0109Paper 09 — How much intelligence remains without the loop? Single-pass capability, scaffolding dependence, and hidden computational costWrites a system as (model, scaffolding vector), defines scaffolding gain, survival ratio SSR, dependence ratio SDR, scaffold cost multiplier SCM and marginal yields along the ablation ladder A0–A5, adds interaction graphs and Shapley-style attribution, hidden-retry and discarded-work accounting, a capability vector and S-grades; loop is not cheating, hiding its cost is.
- PaperPAP-2026-0110Paper 10 — How much physical world does an answer cost? A unified metrology framework for intelligence yieldPacks the four measurement objects into the canonical intelligence event, defines the intelligence yield vector and two-stage efficiency, sets Pareto comparison and the no-premature-scalarization principle, names four capability archetypes, fixes the IPM Minimum Reporting Standard v0.1 and restates the five falsifiable claims; IPM is a metrology candidate, not a discovered constant.
- PaperPAP-2026-0111IPM v0.1 canonical index — series overview, unified notation and the v0.2 experimental entry pointThe series' entry point: the canonical intelligence event, the dependency graph of the ten papers in three lines, a unified symbol table (turn/execution, semantic work, cross-level realization, energy, physical computation, computational spacetime and topology, quality, IBQF/BRQM, scaffolding, yield), the no-premature-scalarization principle, Pareto comparison, the 36-field minimum reporting standard, the ten-step comparison protocol, the series' core invariants, the five falsifiable propositions F1–F5, the v0.2 experiments A–E with the recommended order A → D → B → C → E, a minimal run schema and the versioning rule.