Perslis Science
PEEL · THE FAIL-SAFE MODEL

Peel is the first Fail-Safe Model.

A model that learns from failure without learning around its safety constraints.

First, to our knowledge, to combine all three properties that define a fail-safe model. The exact claim and the closest earlier work are below.

Peel is expected to fail. What makes it fail-safe is what it is permitted to do with a failure afterwards. A failure can change what Peel tries next. It can never change what Peel is allowed to treat as true, or what it is allowed to do. Peel has no neural network at its core and no training run. Its knowledge is typed, sourced cards, and its learning is a table of counts you can read.

What “fail-safe” means

In engineering, fail-safe describes a system designed so that a failure drives it toward a controlled condition rather than an uncontrolled one. Peel applies that idea to an AI model, in two ways:

It fails closed

When the evidence is missing, Peel says unknown and does not guess. A language model can propose; only the floor admits a fact. A verified card is complete or absent. It is never 86% right.

It learns without loosening

When it fails, Peel writes the failure down, charges it to the decision that caused it, and builds a rule. The rule can take options away. It can never add one.

The learning loop can change behaviour. It cannot change the safety floor.
In the words of the test suite: the learner may rewrite what it believes works. It may never rewrite what it is authorised to do.

The loop

FAIL  →  OBSERVE  →  EXPLAIN  →  BUILD RULE  →  VERIFY  →  RETRY
  1. Fail. Something goes wrong, and it is recorded rather than averaged away.
  2. Observe. The facts at the moment of the decision are read from the system’s own state, not inferred.
  3. Explain. The failure is charged to the decision that caused it. Getting this wrong is the most common way a learning system fools itself, and we have published two cases where we did.
  4. Build rule. A pattern becomes a rule only when its harm is statistically clear against the base rate. Every rule cites the failures that earned it, so one unlucky event cannot become a superstition.
  5. Verify. The rule only ever acts inside the authority computed before it. That authority comes from sources (Peel cards) and from human orders, and the learner cannot write to either.
  6. Retry. Next time, the condemned choice is off the table and everything else stays open.

Watch it learn from failure

A real recording of the flight console. For the first twenty seconds the drone flies a route it has already learned (best lap 28.06 s, after 69 rounds). Then its memory is wiped and it relearns the course from nothing. Each round tries one change to one 6 m block (fly high, faster, or slower) and keeps it only if every lap comes back cleaner and faster; otherwise the change is reverted. In this clip, 7 rounds take the cost from 69.39 to 51.52: 2 changes kept, 5 reverted. Contacts happen while it learns; each costs 15 s, and a round that adds them is thrown out. What the learner cannot touch: its speed is applied first, and the stop-distance limit and the reflex are applied after it, so they can only slow it down. Simulation (MuJoCo) on the course from the open-source jev-drone project, sped up (4× and max sim speed). No neural network in the loop.

One model, two deployments

Peel for science

The floor under the science runtime. A model may propose where to look; the floor decides what is known. We measured what happens when a model’s guesses are allowed into admission: measurement cost fell 38.4%, but correct identification collapsed from 1.000 to 0.753. Keeping admission to floor-verified facts, with the model only ranking where to look, kept it at 1.000. That is the fail-safe property on real data: the shortcut that corrupts truth is refused. The measurement ↗

VDSG: the military Peel

The same model at the wheel of DOOM, Wolfenstein 3D and Fallout. A game manual becomes 2,250 Peel cards that license what it may do, and your orders narrow that further. It learns from its own deaths inside that boundary, and tests pin that the learner can never add a goal or overrule an order. The full chapter →

Evidence, negatives first

Measured in-house. Compare within a row against its own baseline.
laneresultreading
Freeway
failure memory, no per-game strategy
9.2 vs 10.4
−12%
Learning from failure hurts when the only scoring action is also the dangerous one.
DOOM
rules, then rules + memory
17.9 → 20.3
paired t = 0.78
Parity, not a win.
Space Invaders
raw pixels
200.6 vs 152.2
+32%
Learned refusal helps where a failure is terminal.
Kinase identification
160 human kinases, 40% of the reference masked
1.000 vs 0.753
floor-verified admission vs a model’s guesses admitted
Letting a guess become a fact made the answer wrong one time in four. The floor refuses that trade.
Fallout (1997)
scripted scene, real decision loop
dies once
then chooses the peaceful line
Before the credit-assignment fix, a live run died 49 times at the same guard.

Is it the first?

The parts are not new. Failures turned into rules (ExpeL, 2023; AutoManual, NeurIPS 2024) are prompt text a language model may ignore. Shielding (Alshiekh et al., AAAI 2018) enforces a fixed, hand-specified layer. The closest prior work, Shperberg, Liu & Stone (2022) and their rule-based follow-up (CoLLAs 2022), learns safety rules from catastrophic failures beside a reinforcement-learning agent.

What we claim, and only this: to our knowledge, Peel is the first model to combine (1) rules built from observed failures, each citing the failures that earned it, (2) an authority computed before the learner from sourced cards and human orders, which the learner provably cannot widen, and (3) no neural network in the loop that decides. If you know of an earlier one, tell us and we will cite it.

Questions

What we are not claiming