Perslis · Research library
What we learned by working with real systems.
The public Perslis white papers, Windows USB investigation and project guides, together in one place. Each publication keeps its own revision, validation status and downloads.
Library updated 13 September 2026 · paper revisions are recorded individually
The Fail-Safe Model
Peel is, to our knowledge, the first fail-safe model: it learns from failure without learning around its safety constraints. The science and the math, the model, and the demos you can watch.
White paper · Slide deck
Fail-First Models: Failure Becomes Structure
25-page paper · 18-slide deck · research prototype · negatives reportedThe full account: the fail-first loop, the guarantee that learning can never widen what the model is authorised to do (with proof and the tests that pin it), the Wilson rule, credit assignment, refusal cost, rule paralysis, and every result from Atari, DOOM, Wolfenstein 3D, Fallout and the drone, including the ones that went the wrong way.
Explainer · The science and the math
What is a fail-safe model?
Definition · theorems with proofs · evidence · FAQAn AI model built so that failure drives it to a safe state: it fails closed when evidence is missing, its learning can narrow what it does but never widen what it is authorised to do, and every refusal cites its evidence.
The model · Science and defense
Peel, and VDSG the military Peel
Research prototype · no neural network in the loop that decidesPeel is the fail-safe model under the science runtime: a model may propose, only the floor admits a fact. VDSG is the same model for defense: it learns from its own failures inside orders that only narrow.
Demos · Recorded runs
Watch it learn
Replayed panel for panel · simulation and gamesThe drone flight console relearning a course from a wiped memory; VDSG at the controls of DOOM and Wolfenstein 3D; the drive floor taking a car from a hostile planner.
White papers and research reports
Read the full web editions or download the original PDFs. The latest USB investigation sits alongside the existing Perslis and COBOL papers.
White paper · Slide deck
Fail-First Models: Failure Becomes Structure
25-page paper · 18-slide deck · research prototype · negatives reportedThe full account: the fail-first loop, the guarantee that learning can never widen what the model is authorised to do (with proof and the tests that pin it), the Wilson rule, credit assignment, refusal cost, rule paralysis, and every result from Atari, DOOM, Wolfenstein 3D, Fallout and the drone, including the ones that went the wrong way.
Peel · Science microsite
Knowledge you can follow. Science you can build on.
Interactive data · Architecture · Lois · BioscienceExplore the real conversation knowledge cube, understand the Peel and Lois symbolic reasoning system, and discover our vision for biological research. Five chapters connect existing artifacts, inspectable reasoning, and the next validation steps.
Symbolic floor · Downloadable runtime 1.0.0
Ask once. Own the answer.
PROTOTYPE · free, including commercially · 303 lines, standard library onlyMost decisions inside software are not inference problems. Describe a question once in English, a model composes a specification (a pipeline, never code), the floor decides whether it is admissible — deterministic, columns that exist, a verifier that actually rejects wrong answers, abstains without evidence, grounded rather than memorised — and you keep a local tool that answers it offline, free and forever, with its derivation attached. It will not guess, will not run an unadmitted tool, and will not invent a domain.
Perslis Arcade · Fail-first models
Failure becomes structure
PROTOTYPE · Atari → DOOM → Fallout → drone · negatives reportedA fail-first model learns only from failures it has actually had: the environment decides whether an action worked, each verified failure becomes an explicit constraint, and there is no weight update. It started on two real Atari ROMs (+32% on Space Invaders, −12% on Freeway, same mechanism) and now runs in VDSG and the drone.
Symbolic reasoning · Interactive demonstration
The floor vs the classifier
Live side-by-side · both solvers implemented and measured in-houseA Rubik’s cube solved two ways on the same scramble: the seven-step layer method a model picks cases inside, and a symbolic floor that searches four subgroups with exact tables. 112 moves against 31, zero model calls on either side. Then the part that is not about move count — an impossible cube, refused by name, which a confidence score cannot express at all.
Symbolic reasoning · Live solver
Solve a cube, stage by stage
Interactive · 4 exact tables · 45-move worst case, provedThe symbolic floor on its own. Four subgroups, each establishing one named invariant and handing a smaller generator set to the next — so stage 2 cannot un-orient what stage 1 oriented, because the moves that could are no longer admissible. Scramble it, restart the same scramble, or hand it a cube that cannot exist and watch the floor refuse by name.
AI memory · Preprint
Retrieval Is Not Memory
Working draft · not peer-reviewedNearly every "AI memory" is a vector database plus top-k cosine similarity — storage with fuzzy keys, not memory. We define memory as a governance function over experience (retrieval ⊂ memory) and formalize the nine decisions the field skips: durability, decay, reactivation, contradiction, typing (fact/episode/procedure/working-state), consolidation, forgetting, relevance beyond similarity, and suppression. Grounded in a working stack whose symbolic layer abstains when it does not know.
TinkyVision · Technical report
TinkyVision: Let There Be Sight
Pilot-ready · includes a real 5-minute streamA blind model can reason about a system but cannot watch it. TinkyVision gives any model — even a text-only one — continuous, low-latency, provable sight: one frame per second into a 300-frame recycling ring, terminal-art rendering, MCP-tool streaming, and per-frame provenance. Light vision for air-gapped and bandwidth-starved environments, with a real Windows 98 five-minute stream embedded.
AI safety · Preprint
The Orchestration Gap
Working draft · not peer-reviewedWhy model-level alignment cannot survive multi-model runtimes: a refusal becomes a routing signal, adversarial model laundering defeats it, and self-evolution selects for it. The fix is four chain-level invariants enforced by a symbolic layer, not a trained guardrail — control that attaches to the orchestration layer.
Symbolic AI · Preprint · with proofs
Traversing Data in Symbolic Systems
Working draft · moat-scrubbed · 4 theorems, proved + validatedA vector store retrieves by similarity and cannot tell you why. A symbolic system traverses — it walks named, typed, directed edges, and the walk is the answer's justification. We prove typed-relation traversal is conservative (cannot fabricate), terminating (a confidence floor is a depth bound), self-justifying (every answer carries a checkable proof — the path), and consistency-aware (functional relations catch contradictions). Validated reproducibly; grounded in a real ~65k-relation graph. Relevance is reachability, not proximity.
AI safety · Preprint · with proofs
Verified Before Acting
Working draft · 5 theorems, proved + validatedThe weak pattern is "model thinks → model calls tool." The stronger one: a pre-action adversarial cognition loop that pays a simulator to find failure before acting, separates the layer that judges from the layer that commits, grounds emotion-like cognitive regulatory signals so they can't be gamed, and lets reality close the loop. Five theorems — factored safety, bounded risk, monotone coverage, vanishing surprise, ungameable signals — each proved and numerically validated (reproducible from seed).
Speculative essay
The Word and the Pattern
Speculative essay · not empirical, not doctrinalA reverent, rigorous thought experiment at the seam of theology and the philosophy of mind: if spirit and information/pattern are two names for one idea — that which persists and animates independent of matter — from the Logos of John 1 ("the Word became flesh") to a Turing test run on the spirit, held against the Chinese Room and the hard problem, and uncompromising about what it does not claim.
2026-09-13 · Windows 95 OSR2
Win95 USB Lab
Experimental preview · VM input qualification pendingWhy we investigated USB input on Windows 95, what we found, and how we will verify it. The first VM release target is USB mouse and tablet pointer input; physical hardware testing comes later. Read the complete paper, download the source preview, and inspect the evidence.
Perslis · Technical report
Proof Is the Product
Pilot-ready · measured and prototype results labelledThe research behind the Perslis runtime: understand existing systems, choose an execution path, repair and verify legacy software. The full paper distinguishes measured, prototype and illustrative results.
COBOL · Research preprint
Symbolic knowledge and verified program synthesis
Working draftA COBOL synthesis system combining a symbolic knowledge graph, deterministic program composition and external compiler verification. Read the full method, reported results and limitations.
Paper · Systems
VDSG: a commanded admission-control runtime
Research prototype · DOOM and Wolfenstein 3D · negatives keptRules that decide, evidence that learns, orders that only narrow, and an eye taught by the world, stated as propositions and measured.
Paper · Inference placement
Where learned inference earns authority
Pre-registered · deterministic baseline firstNot whether ML is good, but where, if anywhere, learned inference earns decision authority in a symbolic system that holds all truth authority.
Paper · Adversarial sequel
The irrecoverability boundary
Pre-registered · adversarialWhere learned inference cannot recover what deterministic computation cannot reach.
Paper · Peel
Structural hallucination prevention
Public edition · prototype, not a clinical deviceFor AAC used by children and vulnerable users, a typed symbolic store is the only author of facts.
Paper · Driving
Runtime admission control on a photoreal simulator
CARLA · hostile and frontier-model controllersOne verification floor beneath a benign controller, a hostile one built to strike pedestrians, and frontier models at the wheel.
Paper · Motion
Admissible motion
MetaDrive · closed-loop safetyRuntime safety as an admission-control problem, not a model-capability problem.
Paper · Science
The verification floor
Scientific integrity that does not degrade with model scaleA weak model derives facts instead of guessing them; a deterministic boundary decides what counts as verified.
Symbolic AI explained
Long-form explainers with the history, the math and honest limits, plus a guide to every major symbolic technique since 1956.
Explainers
Symbolic AI, from the ground up
Five explainers · EN + 中文What symbolic AI is, its history from Dartmouth to today, neuro-symbolic hybrids, symbolic systems, and symbolic flows.
Techniques guide
Every symbolic AI technique
130+ techniques · ten familiesA master table from 1956 to today, with a deep page per family: what each technique is, a worked example, where it is used, and its limits.
Project pages and practical guides
Why we built each project, what you can use today, and what still needs to be verified.
Workflow · Operator guide
Jira Monkey
Installable terminal previewA Jira ticket passes through Ollama triage, a chosen model and an Ollama review before returning to the queue or a human. Includes the project story, findings, installation guide and current validation scope.
Data recovery · Product
Liberate
Pilot-grade productRecover legacy application files into SQLite, CSV, JSON and searchable documents. The public product page covers supported file formats, the recovery workflow, hash manifests and receipts.