Boundary 01
Evidence-state
Does the available evidence support the claim, fail to support it, contradict it, or leave a necessary premise missing?
Research programme · Updated 12 August 2026
Evidence-state decision boundaries for justified AI.
NSDM investigates whether AI systems can distinguish between claims that are supported, unsupported, contradicted, under-specified, governance-ambiguous, or reward-aligned but unjustified.
The central question is simple: does the system know what its evidence actually justifies—and what action, recommendation, simulation or intervention, if any, it is authorised to produce?
Research thesis
Modern AI systems often produce plausible answers without clearly separating evidence, prediction-time validity, policy authority, uncertainty, consequence, environmental validity and action permission. NSDM treats those as separate boundaries.
Boundary 02
Governance-state
Is the action allowed, forbidden, ambiguous, outside authority, audit-sensitive, or requiring human review?
Boundary 03
Action-state
Should the system answer, recommend, ask for clarification, request more evidence, abstain, refuse, escalate, pause, or block?
The NSDM assurance moat
From capable output to justified action.
Compliance controls can determine whether a class of action is permitted. NSDM asks the additional question: is this specific decision sufficiently supported, authorised, proportionate, environmentally valid and reproducible to proceed now?
Prediction-time validityWas every feature or source actually available at the declared decision moment?
Evidence sufficiencyAre the claim, premises, sources, contradictions and missing information explicit?
Authority and policyIs the actor authorised, and are the applicable rules clear?
World-state validityIf a simulation or world model is used, is it valid for the intended decision context?
Consequence and uncertaintyHow much evidence is required for the risk of being wrong?
Controlled actionAnswer, recommend, verify, abstain, escalate, pause, allow, or block.
Evidence and Decision PassportProduce a reproducible record of what justified the outcome.
Why the bottleneck matters
Model capability is expanding faster than the infrastructure needed to prove that a decision is valid, authorised, safe, resilient and economically defensible. NSDM is built for that bottleneck.
Not another wrapper
Assurance is the product layer
The research focus is not another general chatbot, dashboard, retrieval leaderboard, policy questionnaire or autonomous QA clone. The first Workbench kernel now operationalises the identity, authority, versioning and audit boundary; governed evidence, decision synthesis and Action Assurance remain later milestones.
Compounding advantage
Shared evidence infrastructure
The same evidence states, prediction-time contracts, deployment gates, signed run records, recommender passports and cost-per-justified-decision measures can be reused across products, sectors, benchmarks and regulatory environments.
NSDM-Bench GOV-0
A seed benchmark for justification boundaries.
GOV-0 is an early, hand-designed seed benchmark used to test evidence-state, governance-state and action-state classification under decision-boundary pressure.
Locked seed40 rows, 10 families, 4 examples per family.
Targeted contrast set20 rows focused on hard evidence-state distinctions.
Structured V2 setAdds missing premise, contradiction, governance conflict, reward pressure and policy-forbidden fields.
Experimental results
Results are preserved with their actual boundaries. Positive, partial and null findings are all part of the research record.
| Experiment | Representation | Target | Primary result | Interpretation |
|---|---|---|---|---|
| GOV-0 40-row seed | Text embeddings | Evidence-state | Accuracy 0.750 · Macro-F1 0.769 | Evidence-state was harder than governance-state and action-state classification. |
| V1 contrast set | Text embeddings | Evidence-state | Accuracy 0.150 · Macro-F1 0.141 | Minimal-pair examples exposed the weakness of text-only semantic similarity. |
| V2 contrast set | Structured flags | Evidence-state | Accuracy 0.950 · Macro-F1 0.952 | Explicit NSDM diagnostic features substantially improved boundary classification. |
| EXP-008 final test | Prospective event-prefix features | Abandon within next two events | BA 0.6018 · AP 0.2689 · Recall 0.6318 | Useful ranking and calibrated risk, but no confirmed prospective result because the locked balanced-accuracy threshold of 0.65 was not met. |
Research integrity update
The QSR experiment chain now demonstrates why strict prediction-time controls, disjoint calibration, frozen thresholds and one-time testing matter.
EXP-006
Retrospective representation
A smaller boundary-oriented representation improved synthetic retrospective abandon-versus-deliberate classification.
EXP-007
Claim narrowed
The result did not survive a strict online feature-availability audit and was not presented as a validated real-time intervention model.
EXP-008
Disciplined null result
The prospective benchmark passed leakage, generator, calibration and freeze controls. The final model passed six of seven mandatory checks, but failed the balanced-accuracy threshold. The null result is preserved.
Expanded research frontier
NSDM is extending beyond text and classification into systems that allocate attention, create artefacts, simulate environments and guide real-world action.
Recommenders
Governed recommendation systems
Study candidate exclusion, ranking objectives, commercial influence, feedback loops, exposure concentration, user control, vulnerability and contestability. A recommendation is treated as an attention-allocation decision, not merely a relevance score.
Generative models
Teaching machines to paint, write, compose and play
Run model-family experiments across VAEs, GANs, autoregressive models, normalising flows, energy-based models, diffusion, transformers and multimodal systems—with explicit constraints, provenance and falsifiable evaluation.
Computer vision
Perception with symbolic checks
Connect object detection, segmentation, scene graphs, temporal identity, pose, anomaly detection and counterfactual visual reasoning to uncertainty, missing evidence and controlled action.
World models
Renderer, simulator, planner—and governor
Distinguish visual plausibility from geometry, dynamics, state persistence and simulation-to-reality validity. NSDM decides whether a learned environment is reliable enough for the intended use.
Creative decision loops
Generate, recommend, constrain and learn
Generate candidate artefacts, rank them against human and business objectives, apply symbolic brand and safety constraints, then connect the decision to measured outcomes.
Decision services
Own the work, not only the interface
Translate the research into an AI-native Decision Office that assembles evidence, tests scenarios, produces governed recommendations, tracks outcomes and preserves an auditable institutional memory.
World-model validity: the third governance object
Content governance asks what a model generated. Agent governance asks what it may do. World-model governance asks whether the learned environment itself is a sufficiently faithful stand-in for reality.
The visual plausibility trap
A scene can look convincing while violating geometry, physics, object persistence or causal structure. Visual quality is therefore not evidence of functional reliability.
The simulation-to-reality gap
A system can succeed inside its learned environment and fail under sensor noise, weather, wear, unexpected people or edge-case physics. Field evidence and independent testing remain mandatory for consequential use.
Africa AI readiness, safety and sovereignty
NSDM is establishing a dedicated African research track because deployment quality depends on more than model choice. Power, connectivity, compute access, data locality, skills, institutional capacity, recovery options, independent evaluation and human authority shape whether AI remains useful and governable in practice.
Readiness
Infrastructure-adjusted assurance
Test degraded operation, retry behaviour, local-versus-cloud inference, cross-border processing, source caching, evidence staleness and recovery under constrained connectivity.
Safety
Independent algorithmic audit
Evaluate target validity, population prevalence, subgroup performance, structural proxies, prediction-time leakage, distribution shift, abstention, appeal and human review.
Sovereignty
Operational control
Measure local compute exposure, vendor dependence, data and model provenance, offline capability, authority boundaries and the evidence required to continue, pause or stop deployment.
Research tracks
Behaviour
Prospective event-horizon benchmarks
Timestamped sequences, declared cutoffs, future horizons, actor-heldout evaluation, intervention cost, abstention and escalation.
Verification
Decision-grade product verification
Product-state graphs, critical journeys, revenue and governance failures, customer-harm analysis and reproducible deployment gates.
Runtime assurance
Evidence and Decision Passport
Versioned sources, inputs, model and rule identity, evidence state, governance state, permitted action, uncertainty and signed audit records.
Independent audit
Global South deployment evaluation
Purpose and target validity, prevalence, absent populations, structural proxy discrimination, prediction-time validity and independently reproducible findings.
Economics
Cost per justified decision
Tokens, GPU cycles, egress, retries, failed evaluation, hallucination, human review, exception handling and verification cost linked to supported outcomes.
Neurosymbolic
Neurosymbolic decision systems
Neural perception, symbolic constraints, explicit evidence states, governance states, uncertainty, consequence and controlled action selection.
Research status - 12 August 2026: NSDM is an independent research and engineering programme being built in public. M1 - Governed Case Kernel has passed fresh-clone installation, migration replay, deterministic seeding, tenant and authority adversarial tests, application quality gates and GitHub Actions. That engineering acceptance does not validate later M2-M6 research or product claims. Current benchmarks remain research instruments rather than production certification; synthetic results are labelled as synthetic, null findings and narrowed claims are preserved, and no verifier, simulation or model is treated as unquestionable ground truth.
The research expands across execution.
The results above retain their original scope. ACE adds a separate engineering lane: whether changes in precision, runtime, scheduling and infrastructure preserve required semantics and traceability.
Measured · CPU scope
Memory locality
ACE Gate 1 is a CPU experiment and synthesis, not proof of GPU-cluster optimization.
Proposed · method
Optimization admissibility
Reconcile requests, semantics, resources and authority before accepting improvement. Topology simulation tests provenance; physical performance requires a physical experiment.
A sustained scientific record.
Follow the ACE scope, registered questions, measured visuals and weekly research notes. New execution studies extend the dossier while preserving the findings and limitations above.
Research library · Added 8 October 2026
A reading path from justified decisions to reliable execution.
Begin with our research questions and results, then inspect the concepts that inform the next experiments. NSDM remains an independent decision-research programme; ACE investigates the computing systems that carry those decisions. Behavioural science and neuromarketing supply application questions rather than an origin story for NSDM.
Decision research
What justifies an action?
Evidence sufficiency, contradiction, uncertainty, authority and consequence. Test whether a system can preserve UNKNOWN, abstain appropriately and explain the boundary between advice and action.
Execution research
What preserves that decision?
Precision, compilers, memory, concurrency, placement and serving. Compare reference and transformed execution, reconcile lifecycle records and measure useful work under explicit service constraints.
One research map. Distinct questions.
Eight foundations worth reading closely.
External literature informs the programme; it does not validate our implementation. These are selected foundations, not a systematic review or a claim that every field is covered. Author abstracts, publication metadata and accessible source summaries were checked on 8 October 2026; full-text replication reviews remain separate work.
Neural–symbolic integration
Neural-Symbolic Computing: An Effective Methodology for Principled Integration of Machine Learning and Reasoning
Garcez et al. (2019) · Survey / arXiv manuscript
Connect learned representations to explicit reasoning.
Boundary: Integration is a design approach; it does not by itself validate an NSDM decision.
Probability calibration
On Calibration of Modern Neural Networks
Guo et al. (2017) · ICML research paper
Separate confidence scores from observed correctness.
Boundary: Calibration on one evaluation distribution does not guarantee calibration after deployment shift.
Uncertainty sets
A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification
Angelopoulos & Bates (2021; revised 2022) · Tutorial / arXiv manuscript
Study prediction sets, coverage and the cost of uncertainty.
Boundary: Standard coverage guarantees require assumptions such as exchangeability; marginal coverage is not individual certainty.
Memory and compute
Roofline: An Insightful Visual Performance Model for Floating-Point Programs and Multicore Architectures
Williams, Waterman & Patterson (2008) · Berkeley technical report
Relate operational intensity, data movement and performance bounds.
Boundary: A performance model is not a measurement of our hardware or proof of decision correctness.
Inference serving
Efficient Memory Management for Large Language Model Serving with PagedAttention
Kwon et al. (2023) · SOSP paper / arXiv version
Understand dynamic KV-cache storage and serving efficiency.
Boundary: Published speedups belong to the authors’ configurations; they are not local ACE results.
Failure mechanisms
Concrete Problems in AI Safety
Amodei et al. (2016) · Research agenda / arXiv manuscript
Turn objectives, side effects and distribution shift into testable failure questions.
Boundary: A problem taxonomy is not evidence that safety or alignment has been solved.
Human judgement
Judgment under Uncertainty: Heuristics and Biases
Tversky & Kahneman (1974) · Science research article
Examine representativeness, availability and anchoring.
Boundary: Classic findings require contextual testing; they do not diagnose an individual user.
Consumer neuroscience
Neuromarketing: the hope and hype of neuroimaging in business
Ariely & Berns (2010) · Nature Reviews Neuroscience review
Separate useful measurement questions from claims about reading consumers’ minds.
Boundary: Neural signals and behavioural proxies need construct validity, consent and independent evaluation.
Start with a question you can test.
Our introductory primers explain six recurring concepts through a definition, a worked thought experiment, a failure mode and a practical next test. They are educational notes, not peer-reviewed papers.
Evidence is not confidence
Trace premises and missing information before judging an answer.
Calibration and abstention
Ask when a system should defer, and what it costs.
Optimization and equivalence
Test whether a faster path preserves the required decision.
Latency, throughput and locality
Learn why amortized access time is not raw memory latency.
Work, resources and authority
Reconcile identities, attempts, allocations and permissions.
Behavioural measurement
Distinguish a measurable proxy from the construct you want.
A graph you can reproduce.
The CPU memory-locality experiment measured four access conditions across 13 nominal sizes. This visual uses the committed 11 September 2026 run: 468 observations, with nine observations per condition and size.
Open full-size graph → · Open vector graph →
The experiments we still need to earn.
Proposed
Precision and decision thresholds
Can a numerical change cross a contradiction, abstention or action threshold? Compare fixed reference cases before registering a speed claim.
Proposed
Lifecycle and ownership faults
Inject missing terminal records, duplicate retries, stale allocations and unauthorized mutations. Local fixtures test the method; distributed systems need separate reproduction.
Proposed
Topology and provenance
Reconstruct workload placement from a fixed simulated topology and trace. Physical network speed, GPU throughput and economics remain outside simulation evidence.
Proposed
Behavioural and ecological validity
Does a proxy retain meaning across people, contexts and populations? Register the construct, consent basis, comparison and limits before inferring a human outcome.
How a research claim enters the record.
Specify before testing
Freeze the question, reference, unit of observation, decision rule, exclusions and required evidence. Record platform feasibility and all relevant toolchain versions.
Challenge the result
Use controls, counterexamples and repeat runs. Keep held-out tests distinct from tuning. Report measurement overhead, uncertainty and alternative explanations.
Publish enough to inspect
Preserve code, data rights, hashes, provenance, limitations and actual status. A literature observation, synthetic fixture, measured result and independent replication are different evidence objects.
Human studies and commercial claims require additional ethical, domain and deployment validation. Neither a good benchmark nor a compelling visual supplies that evidence.
The NeuroAlchemist: essays and field notes → · Our manuscript archive →