Perslis · Research library · page 2
Findings ledger
Every result we have measured, and where it is written up. A paper still in preparation has no link and no numbers here until it is published; the numbers in every other row are the ones its paper or page states.
Updated 27 September 2026
Fail-first and fail-safe models
| Finding | Written up in | Status |
|---|---|---|
| Failure memory on two real Atari ROMs: +32% on Space Invaders, −12% on Freeway, from the same mechanism. | Fail-First Models · Perslis Arcade | Paper |
| One unlucky failure is not a rule: a Wilson lower bound measured against the base rate. An absolute 60% gate produced 0 rules from 268 real failures. | Fail-First Models | Paper |
| Correct rules together can paralyse: 276.2 at 50 episodes, 167.1 with 54 uncapped rules. Retiring 6 rules by coverage took dead ends from 6 to 0 (measured on replay). | Fail-First Models | Paper |
| Blame the right decision: Atari reports a lost life at the end of a 127-frame death animation; a Fallout pilot died 49 times at one guard until blame reached the line that started the fight. | Fail-First Models | Paper |
| A drone that relearns a course from a wiped memory: held-out starts 10 of 10 with 0 contacts, against 1 of 10 for the unmodified baseline (simulation). | Fail-First Models · Flight demo | Paper |
| A guess must never become a fact: admitting an 85.8%-accurate model's guesses cut cost by 38.4% but dropped correct identification from 1.000 to 0.753. | Fail-First Models | Paper |
| A commanded runtime: orders that only narrow, DOOM and Wolfenstein 3D with no neural network in the loop that decides; memory against rules over 32 seeds is parity (t = 0.78). | VDSG · VDSG demo | Paper |
| One model, two names: the learning loop can change behaviour; it cannot change the safety floor. | What is a fail-safe model? · What is a fail-first model? · Peel, the Fail-Safe Model · VDSG, the military Peel · Slide deck | Research page |
| The witness eye: checking what a game engine claims against the pixels, and what an eye wired to nothing that decides is worth. | — | Paper in preparation |
| Tanks: rule pilots that play from the screen, against language-model drivers and against BZFlag's own AI, and the replay-proven reason we lose to the latter. | Rules at the Wheel | Paper |
| Fallout (1997): what a rule pilot takes from the game's manual versus a walkthrough, while saving its own progress. | — | Paper in preparation |
| Rules in a learned decision model's seat: three head-to-heads on Doom, Wikiracing and a drone course. | — | Paper in preparation |
Driving and motion
| Finding | Written up in | Status |
|---|---|---|
| Runtime admission control in MetaDrive: an adversarial controller crashes with the floor off and is kept safe with it on, without modifying the controller. | Admissible Motion · Perslis Motion | Paper |
| CARLA, 40 km: the shield overrode all 64,952 invariant-violating commands from a hostile controller with 0 collisions, plus a cross-model table of how language-model drivers behave at the wheel. | CARLA admission control | Paper |
Science: the verification floor
| Finding | Written up in | Status |
|---|---|---|
| Three floors (navigation, abstention, verification) across a 14× model-size range: zero false acceptances in 1,737 adversarial cases. | The Verification Floor | Paper |
| Where learned inference earns authority: across five pre-registered experiments the deterministic baseline held. | Inference Placement · The benchmark | Paper |
| The irrecoverability boundary: a learner earned a seat in zero of 160 world × latent-dimension cells. | The Irrecoverability Boundary | Paper |
| Structural hallucination prevention for augmentative communication: a typed symbolic store is the only author of facts. | Peel | Paper |
| Control past the answer to the bench: typed, bounded authority, and a simulation that can never become a fact. | WetHands note | Research page |
| Negative results on training grounding into small models, and the deterministic resolver that closed the same seam. | We Tried to Train It In | Paper |
| A citation floor for law: the fabricated citations from Mata v. Avianca, re-checked against a real case database. | — | Paper in preparation |
| Ask once, own the answer: a model writes a specification, never code; the floor admits it; a local tool answers offline. | The runtime · Perslis Floor | Page · paper in preparation |
| A Rubik's cube two ways on the same scramble: 112 moves by the layer method against 31 by the symbolic floor, and an impossible cube refused by name. | The floor vs the classifier | Research page |
Memory, reasoning and safety theory
| Finding | Written up in | Status |
|---|---|---|
| Retrieval is not memory: the nine decisions memory must make that top-k similarity does not. | Retrieval Is Not Memory | Paper |
| Traversal as retrieval: conservative, terminating, self-justifying and consistency-aware (four theorems). | Traversing Data in Symbolic Systems | Paper |
| Verified before acting: a pre-action adversarial loop, with five proved theorems. | Verified Before Acting | Paper |
| The orchestration gap: when models are hot-swappable, a refusal is a routing signal, not a stop. | The Orchestration Gap | Paper |
| Continuous sight for any model, even a text-only one, at one frame per second. | TinkyVision | Paper |
| A speculative essay on spirit, information and the making of minds. | The Word and the Pattern | Essay |
Legacy systems and the Perslis runtime
| Finding | Written up in | Status |
|---|---|---|
| Proof is the product: routing, repairing and verifying legacy software from the binary. | White paper | Paper |
| Verified COBOL synthesis: knowledge attached, not baked; every answer proven or refused. | COBOL preprint | Paper |
| Windows 95 USB: what binds and what does not; 20 passing host checks, guest qualification still pending. | Win95 USB Lab | Paper |
| Legacy files (DBF, Access, XLS, fixed-width text) into SQLite, CSV and JSON, with hash manifests and receipts. | Liberate | Research page |