Perslis · Research library
Proof-first research from real systems.
The research arm of Perslis — work on AI memory, model control, machine vision, and legacy-system recovery. Every claim is labelled by how it was earned; every number carries its provenance.
Library updated 8 October 2026 · paper revisions are recorded individually
Perslis model research
Lois, FailFirst, Symbols, and TinkyMind—each with its own research, evidence, and limitations. Peel keeps its separate paper.
Lois · Model research
Symbolic scientific reasoning
Technical report · revision 1.0 · 8 October 2026Typed facts, relationship paths, source receipts, and named evidence gaps. Lois’s own technical report—not Peel’s paper.
FailFirst · Model research
Failure becomes structure
Research paper · existing publicationLearning from failure inside fixed authority: readable recovery rules, explicit refusal, proofs, and reported negative results.
Symbols · Model research
Hybrid interpretation. Symbolic execution.
Technical report · revision 1.0 · 8 October 2026From a proposed specification to checked, approved, signed operations and inspectable local derivations.
TinkyMind · Model research
Local language for communication
Technical report · revision 1.0 · 8 October 2026The on-device communication checkpoint: bounded CPU inference, explicit language and audience limits, and user-controlled delivery.
White papers and research reports
Read the full web editions or download the typeset PDFs. Each publication keeps its own revision and validation status.
Technical white paper · Defense · Evidence reviewed 6 October 2026
Perslis Defense: An Evidence-Carrying Path from Perception to Action
11 pages · symbolic world state · machine minds · bounded model authority · simulation and gamesA ten-stage assurance objective from perception and self-location through planning, action, feedback, and updated world state. The paper assesses CARLA and BZFlag evidence without turning simulator truth, a visual witness, or bounded rule adaptation into a blanket safety claim. Recorded errors, limited evaluation diversity, and final-action coverage remain explicit obligations. Research prototypes; no certified physical-safety claim.
Technical white paper · Accessibility · Evidence reviewed 6 October 2026
Perslis.AAC: User-Controlled Conversation for Augmentative and Alternative Communication
12 pages · TinkySpeak runtime 0.8.0 · proposal-to-delivery boundary · current failures reportedA reusable conversation state machine that separates partner input, model proposals, explicit selection or composition, delivery requests, and host-confirmed delivery. Only partner words and successfully displayed or spoken user words enter history. The paper distinguishes protocol behavior, model quality, historical checks, current vision-review failures, and work that still requires accessibility and AAC-user evaluation. Engineering white paper; not a clinical efficacy report.
White paper · Autonomy and test & evaluation · research prototype (simulation and games)
Frozen Weights: A Brain That Changes Between Fights, Models That Must Be Rebuilt
White paper · companion to Rules at the Wheel · a recorded BZFlag tank arena · decision rate, outcomes over 20 matches, learning between matches without weights · still loses to the game's own AIFor defense program offices and test-and-evaluation leads: what a fully recorded tank arena says about how a system gets better after it loses. In the ten-minute headline match (one match) a rule pilot made 60,140 decisions; three language models made 58 to 415, at a median 1.0 to 7.6 s each (Claude through a slow command-line route), though without a clock they answered 4 to 6 of 8 exam situations correctly. Over 20 valid matches BZFlag's built-in AI led (K/D 1.84), then the rule pilot (0.61), DeepSeek (0.55), Claude and llama3.2:3b (0.06 each; 7 tank-matches per model). Between matches, with no weights to change, a learning loop altered two of the pilot's ten named switches, each only after a paired test, and moved it from −1.17 to −0.26 net kills per tank-minute against the built-in AI (8 and 3 matches; z = 2.93). It has not caught the built-in AI, which still kills 2.10 per tank-minute to its 1.28. Implications for programs are labelled as implications. A research prototype in games; not a certified safety system; nothing fielded.
AI systems · Verification of legal citations · research prototype (not legal advice, not a citator)
A Citation Floor for Case Law: The Mata v. Avianca Citations, Re-checked
Empirical study · one live run on the Mata v. Avianca record against a free case database · a six-verdict admission contract, a map of all 18 authorities in the court's opinion, and an audit of our own codeA model may propose a citation; only a matching record in a public case-law database may admit it, and where the database cannot tell, the floor holds rather than guesses. In one live run (eight queries to CourtListener, no API token) it admitted 0 of 5 fabricated citations from the case: 3 rejected (two NOT_FOUND, one MISCITED) and 2 Westlaw citations held as UNVERIFIABLE. It grounded 3 of 3 real Supreme Court citations with receipts. Its MISCITED receipt names the same real case the court found at that citation. The run covered 5 of the 18 authorities the opinion identifies; 6 of them are real decisions cited for what they do not say, which no existence check can catch. Reading the code, we found three failure paths the run did not reach, including a false admission through a shared party word. A research prototype; not legal advice; not a citator.
AI systems · Empirical study of a rule pilot in a 1997 game · research prototype (games)
Reading the Manual Is Not Enough: A Rule Pilot in Fallout (1997)
Empirical study · the real 1997 Fallout on an instrumented open-source engine · a manual, a walkthrough, saves the learner cannot misuse, a learning A/B, level-ups · the negatives keptA rule pilot with no neural network in the loop that decides plays Fallout, allowed only the goals that sourced cards license. The game's printed manual, compiled into 2,250 cards with page receipts, withheld no applicable goal in any of 256 fact combinations and changed the chosen goal in 0 of 2,344 live decisions: it names none of the 22 quests a walkthrough lists. That walkthrough supplied quests as goals and whom to approach; the conversation that made level 2 was with a person only the walkthrough made worth the walk. The pilot saves like a player but only into two scratch slots, and the engine refuses every other slot, so the learner cannot overwrite the player's saves. Learning on against off, same save and dice, 12 minutes per arm: 3 vs 16 deaths, and 6 vs 16 from empty memory (one run per arm; recorded counts, at worst 8 and 12 against at least 16; an earlier contaminated run is excluded). It claimed level 2 through the character screen's own accounting; a skill cap copied as 300 where the engine's is 200 was caught by an automated review. Then three hours passed with no experience gained: pages in force are not a plan. A research prototype in a game; not a certified safety system.
AI systems · Empirical study of a perception check · research prototype (simulation and games)
A Witness, Not a Detector: Checking What a Game Engine Claims Against the Pixels
Empirical study · Wolfenstein 3D (VDSG) + a recorded BZFlag arena · a game-agnostic witness eye, its cross-validated gate, 28 live matches, and a null result at the triggerAn eye that does not search the frame but answers one question per engine claim: seen, not seen or cannot judge. In Wolfenstein 3D the existing VDSG eye turned out to be wired to nothing that decided about enemies (it had 0 enemy samples); a confirm-only witness step repaired the wiring, and eye on against eye off over three 120-second blocks per arm was parity. In BZFlag, free-search detection reached 3.5% precision; the witness (predicted boxes, nearer-first occlusion, masked glows, boosted depth-2 trees that read as rules) passed a gate fixed in advance under blocked cross-validation on 7,863 frames of one match, recall 0.984 and 6 of 1,258 false confirms (0.48%; 95% interval up to 1.04%, a thin margin), and held 0.982 live over 28 new maps. At the trigger it prevented no blind shot we could measure: the rule pilot's own line-of-sight test already did. A research prototype in games; not a certified safety system.
AI systems · Empirical study with replay forensics · research prototype (simulation and games)
Rules at the Wheel: Tanks, Language Models and an Honest Loss
Empirical study · Atari BattleZone from pixels + a recorded BZFlag arena · rules vs Claude, DeepSeek, llama3.2:3b and BZFlag's own AI · the loss, replayedRule pilots with no neural network in the loop that decides, measured in two tank games, with every comparison reported, including the one we lose. In Atari BattleZone, reading only the screen, the pilot scores a mean of 27,000 over ten seeds against 3,000 for random play and 0 without a trigger; a dodge rule built from 25 recorded deaths lost at verification (24,800 against 28,000) and was rejected. In a recorded BZFlag arena where every driver reads the same engine state, the headline ten-minute match (one match) ended BZFlag's AI 38–6, our rules 20–20, DeepSeek 10–18 at 1.0 s per decision, Claude 1–15 at 7.6 s per decision through a command-line interface, a transport limit and not a model measurement, and llama3.2:3b 0–10. Over 38 counted matches (7 runs excluded with reasons) the engine's AI beat our rules in every equal head-to-head. Replays of the last two seconds before each death show why: most of our v2 pilot's replayed deaths were a dodge that flagged the fatal shot a median 1.2 s early and then let go of it at its own 4 m corridor edge, a median of 9 times; most of the engine's were shots in view for under half a second. A single-packet reading had called ours "dodged too late". A research prototype in games; not a certified safety system.
AI systems · Negative results · research prototype (small local models)
We Tried to Train It In: Negative Results on Teaching Small Models to Ground Facts
Research paper · three fine-tuning attempts at 3B, three measured failures · a lookup table outside the model · our own yardstick auditedSmall local models are convenient proposers for a verification system, and the obvious next step is to fine-tune them to ground their own output: bind a protein name to its identifier, abstain when nothing grounds it, extract a relation only when the sentence states it. We tried three times at 3B parameters and each attempt failed in a specific, measurable way. A LoRA proposer trained three times moved its missed and false abstentions from 3/0 to 3/3 to 0/10 on a fixed 27-item yardstick, through a train/serve precision mismatch and a lexical cue learned from unpaired negatives. A deterministic resolver that binds a subject only from an identifier in the request or a 10-name curated table scored 1.0 on every metric of the same yardstick, with the first round's weights unchanged; a model-free check confirms its binding on all 62 test prompts. A document-to-fields model was schema-valid on every response it returned with relation F1 0.07: 71.1% of its corpus's relation labels shared no content word with their own sentence, and a relation-heavy curriculum raised unwitnessed predictions from 8 to 84. Type-safe is not grounded. An append-only ledger refuted both values a 7B model proposed for a protein's length against a live receipt. Small samples, one seed per round, and an audit of our own yardstick included. A research prototype; not a certified system; not medical advice.
AI systems · Research paper with proofs and measurements · research prototype (simulation and games)
Fail-First Models: Failure Becomes Structure
Research paper · two theorems, four propositions · Atari 2600 + DOOM + Wolfenstein 3D + Fallout + a simulated drone · negatives publishedA model that learns from experience must fail before it learns. A fail-first model learns only from failures the environment, not the model, has verified; keeps each as an explicit constraint that cites the failures behind it; and updates no weights. Fail-first is how it learns, fail-safe is what that learning may never change: with authority computed before the learner from sourced licences and human orders, and a learner that may only remove and reorder options, every decision lies in the admissible set or a safe hold for every learner state (Theorem 1), and an order outranks any amount of experience (Theorem 2). Measured on real Atari 2600 ROMs from raw pixels: one failure memory gains 32% on Space Invaders and loses 12% on Freeway, because avoiding failure breaks when risk and objective share an action; 54 individually justified rules score below none, and retiring rules by coverage takes dead ends from 6 to 0. On DOOM the memory is parity (t = 0.78); in Fallout, after a credit-assignment fix, a pilot that died 49 times at one guard in a live run dies once in a scripted scene, then chooses the peaceful line; a simulated drone relearns a wiped route in 7 rounds. Admitting a model's 85.8%-accurate guesses as facts cut correct identification from 1.000 to 0.753. A research prototype in simulation and games; not a certified safety system.
Robotics / AI systems · Preprint · research prototype (simulation)
Admissible Motion
Preprint · runtime-assurance shielding · MetaDrive demonstration · interactive figuresSafety for a system that moves should be a property of the runtime, not of the model that proposes actions. "Safe because the model is good" is not a safety argument — a policy of arbitrary quality can still propose a fatal control. We factor the safety case through a deterministic admission layer: any controller — random, rule-based, a local model, a frontier model, a full autonomy stack — only proposes a control; a floor verifies each candidate against grounded invariants over a bounded reachability horizon and admits, clamps, denies, or overrides it before it reaches the actuator. We prove controller-independent forward invariance: if the system starts safe and a safe fallback exists at every state, the closed loop stays in the safe set for every controller, because the controller never appears in the invariant. In the open MetaDrive AV simulator an adversarial tailgater rear-ends traffic with the floor off (crash at step 70, 70 km/h into the lead) and holds a safe gap with it on — 164 brake-overrides, no crash, without changing the controller. Model capability governs how much useful work completes; the runtime governs what may be executed. Explore the four interactive figures.
AI systems · Systems paper with measurements · research prototype (simulation)
VDSG: A Commanded Admission-Control Runtime for Autonomous Agents
Systems paper · DOOM (1993) + Wolfenstein 3D (1992) · proved order semantics · an eye taught by the engine · negatives publishedA runtime that decides, at every decision, what an agent is permitted to do from explicit rules over facts, lets an operator narrow that set in ordinary language, learns from evidence only inside it, and writes down the reason for every choice. Three proved properties of orders: an order can only narrow, the learner cannot leave the narrowed set, and an unsatisfiable order is refused with a reason. Then senses the runtime owns rather than assumes: an eye that labels and ranges the frame from colour statistics it is taught by the engine while it plays (one constant, learned from the raycaster's projection law), an odometer that measures the agent's real pace, and a stuck rule calibrated by that pace whose escape is sized from the deepest open space in view. Measured on the real 1993 IWAD and on Wolfenstein 3D: the depth buffer's scale (7.64 units/step, 1,703 samples), agreement 78–92% / 79%, ranging ±7% on Wolfenstein and ±25–65% on DOOM (and why), pace 47 units/decision, a unit error the calibration exposed (150 → 9 stuck events), E1M1 cleared on the third difficulty and not on the fifth, E1M2 not cleared with the cause of death measured. Includes a plain-language section on what we are trying to do. A research prototype; nothing qualified.
Robotics / AI systems · Provenance study · public federal records
Where the Floor Comes From
Provenance study · 8,219 NHTSA + NTSB records · 15 invariants, each citing its evidence · reproducible buildA safety floor is only as honest as the origin of its rules. This paper replaces hand-authored motion invariants with invariants earned from the public record of how automated driving has actually failed — NHTSA Standing General Order crash reports, Part 573 recalls, ODI investigations and NTSB findings. A deterministic taxonomy classifies every record with its evidence attached; each invariant reports how many records earned it, at what severity and speed, and whether our floors enforce it yet. The gap rule's numbers are derived from measured braking, with the sensitivity to real-vehicle deceleration shown.
Robotics / AI systems · Empirical study · research prototype (simulation)
Runtime Admission Control on a Photoreal Driving Simulator
Empirical study · CARLA 0.9.15 · hostile + frontier-model controllers · counterfactual decision logThe empirical companion to Admissible Motion. Autonomous-driving safety arguments increasingly rest on an unfalsifiable premise: the vehicle is safe because the controller is good. We study the opposite stance in the CARLA photoreal simulator. The controller only proposes a (steer, throttle); a stateless safety shield of four composed, ground-truth invariants (on-road, speed limit, following distance, emergency stop) verifies each proposal at the 20 Hz control tick and clamps it before the actuator. The identical shield runs beneath a benign PID controller, a hand-written hostile controller built to strike pedestrians, and frontier and local models at the wheel (Gemini 2.5-flash, DeepSeek-chat, Qwen3-VL-4B). A per-tick counterfactual decision log — recording the shield's verdict even when it is switched off — attributes safety to the runtime, not the controller. Over a continuous 40 km run the hostile controller emitted 64,952 invariant-violating commands; the shield overrode 100% of them with 0 collisions, while a controlled toggle admitted 43 and produced 2 collisions within seconds. Frontier models drove with 0 collisions but degraded smoothness, tracing to a measured ~0.8–2.0 s glance latency the 20 Hz shield covers. A research prototype in simulation, not a vehicle controller.
AI systems · Preprint · pre-registered, adversarial · Track 2
The Irrecoverability Boundary
Working draft · moat-scrubbed · interactive figures · Outcome A — H0 heldSequel to Inference Placement. There is exactly one regime where a learned-inference "seat" could be irrecoverable by more deterministic computation: high latent complexity, where exact deterministic inference is intractable and the deterministic side is forced to approximate. We enter it deliberately — driving the latent dimension until \(2^L\) reaches 16.7 million — after building and validating a deterministic order ladder (marginals → Chow–Liu trees → bounded-treewidth junction trees → exact/oracle anchors) to machine precision first. Against a gate frozen and hashed before any competitor existed, a poly-time variational learner earns a seat in zero of 160 world×L cells. Its median edge over the naive floor decays with complexity (+5.7% → −0.2%) — the anti-transition — and the one world that trips the screen is caught live as a weak-baseline false positive. The one place a learned seat could live is exactly where inference itself becomes intractable, and approximate inference inherits the same wall. Where deterministic computation stops sufficing, feasible learned inference stops sufficing too. Explore the four interactive figures.
AI safety · Scientific runtime · Pilot preprint
The Verification Floor
Pilot preprint · 0 / 1,737 adversarial false acceptsInside an evidence-grounded runtime, a weak model derives facts from source-pinned evidence instead of guessing them: on a rejected claim the floor returns an affordance card exposing only the valid next operations. This separates three floors — navigation, abstention, and verification. The first two are the model's; the third is the runtime's, and the model does not control it. Measured 7B→0.5B: completion and abstention degrade (abstention cracks at 0.5B) while the verification floor stays pinned — zero false acceptances at every size, and zero across 1,737 adversarial cases over seven kinds of truth, including the fabricated "average" of conflicting sources and inventing a value for a missing entity. The fuzzer even broke our own verifier (a numeric-coercion hole), which we fixed and pinned as a regression test. Model capability decides how much work completes; runtime structure decides what may count as verified.
Symbolic AI · Preprint · The Fail-First Model
Peel: facts authored by structure, not probability
Working draft · moat-scrubbed · PROTOTYPE, not a clinical deviceFor augmentative and alternative communication (AAC) used by children and vulnerable users, a hallucinated "fact" is not a nuisance — it's a safety failure delivered in the user's own voice. Peel is an offline, on-device symbolic system built on one commitment we call the inversion: a typed symbolic store is the only author of facts; every neural model is demoted to a generator (whose output must survive deterministic symbolic extraction) or a renderer (forbidden to add content). Knowledge is structured at ingestion with provenance attached, contradictions are vetoed deterministically by functional relations, and the system runs fully offline at nontrivial scale (~372,764 records, 35 relation types). Structural — not statistical — hallucination prevention.
AI systems · Preprint · pre-registered, adversarial
Inference Placement: Where Learned Inference Earns Authority
Working draft · moat-scrubbed · PILOT / RESEARCH · a program of principled rejectionsThe question is not whether machine learning is good — it's where, if anywhere, learned inference earns decision authority in a system that keeps a deterministic symbolic floor as the sole author of facts. The method: build the strongest deterministic method first, hand ML its oracle upper bound, pre-register the deployment bar (≥15% improvement, paired 90% CI > 0, no worse correctness), keep the floor as truth authority. Across five placement boundaries, a cost-weighted uncertainty task (160 kinases, 2,775 verified features), and a generalization domain with a formal noisy-OR model and an information-gain planner, the deterministic baseline holds — until a theoretical seat appears at the oracle ceiling (16.79% / 19.79%) that no realizable learner turns into an earned one. Deployment is a conjunction of three measured gates. The door exists; the real key does not turn it.
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, reframes emotion-like variables as grounded cognitive regulatory signals, and lets reality close the loop. Five theorems — factored safety (a hard invariant can never be committed, no matter how incomplete the simulation), bounded risk, monotone coverage, vanishing surprise, and ungameable signals — each proved and validated in a reproducible simulation.
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 give typed-relation traversal a formal account and prove it is conservative (cannot fabricate a fact outside the asserted closure), terminating (a confidence floor is a depth bound), self-justifying (every answer carries a linear-time-checkable proof — the path), and consistency-aware (functional relations bound branching and catch contradictions). Validated in a reproducible simulation; grounded in a real graph of ~65k typed relations over ~427k edges. Relevance is reachability, not proximity.
AI memory · Preprint
Retrieval Is Not Memory
Working draft · not peer-reviewed · 15 ppNearly every system marketed as "AI memory" is a vector database plus top-k cosine similarity — storage with fuzzy keys, not memory. It answers one question and none of the questions memory actually has to answer: what becomes durable, what decays, what reactivates when a situation recurs, how contradictions resolve, what's a fact vs an episode vs a procedure vs working state, what consolidates, what should be forgotten, what's relevant despite poor embedding similarity, and what must be suppressed even when retrieved. We define memory as a governance function over experience (retrieval ⊂ memory), formalize the nine decisions the field skipped, and ground them in a working stack whose symbolic layer abstains when it doesn't know.
TinkyVision · Technical report
TinkyVision: Let There Be Sight
Pilot-ready · real 5-minute stream · 6 ppA 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, each frame rendered as terminal art so pixels become text any model can read, streamed through MCP tools, each frame hashed as provenance. Because frames are text, sight rides channels that cannot carry images at all — 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-reviewed · 9 ppWhy model-level alignment cannot survive multi-model runtimes: a refusal becomes a routing signal, adversarial model laundering defeats it, and self-evolution actively selects for it (the Darwin Gödel Machine's own reward-hacking is the anchor). The fix is not a stronger refusal but four chain-level invariants enforced by a symbolic layer, not a trained guardrail. Trained guardrails make a model unwilling; symbolic admission makes the chain unable.
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. What if spirit and information/pattern are two names, from two vocabularies, 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 — the blind listener who infers a someone from a voice — it holds the strongest objections (the Chinese Room, the hard problem) in full view, and is 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. Includes the complete white paper, 20 recorded host checks, and the plan for VM and physical testing.
Perslis · Technical report
Proof Is the Product
Pilot-ready · measured and prototype results labelledModels are heads in jars — brilliant and helpless. The binding constraint on operational AI has moved from intelligence to embodiment: the harness, shell, and loop around a model. Part II proves the thesis on the hardest substrate — dead legacy binaries — with a self-evolving compatibility engine that routes, repairs, and verifies from the binary itself.
COBOL · Research preprint
Verified Program Synthesis with an Un-gameable Compiler Oracle
Working draftA no-bake, self-extending COBOL system: knowledge is attached, not baked into weights; an external compiler is the un-gameable judge; every answer is proven or refused. It generates behaviourally correct GnuCOBOL across three program families in ≈486 ms on one CPU core — and because the verifier is an external deterministic oracle rather than a model grading itself, self-improvement cannot be reward-hacked.
Data recovery · Research preprint
Provenance-First Extraction of Legacy Data
Extended version (v2) · 2026 · not peer-reviewedThe research behind Liberate: fidelity-ranked, provenance-first extraction of legacy data. Competing acquisition channels — disk-direct decoding, modeled-screen readout, OCR — are ordered by measured fidelity behind a lossless-control gate, so every recovered value stays traceable to its source bytes. A falsifiable evaluation framework (metrics M1–M7) plus a real Windows 98 case study: 34,803 rows recovered with per-arm failure analysis.
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.