Boundary Conditions a register inside Secure AI Atlas
EN ES

Secure AI Atlas Where the map runs out.

6 pieces on record

Boundary Conditions

A register of the cases where AI systems outrun the map: what happened, how it worked, what it means, and what nobody can settle yet.

Published in English and Spanish · also read it in castellano

Each piece starts from something that actually happened, or from something a study actually showed, and then asks the question the catalogue cannot answer: which constraint was missing, what the behaviour would require in order to mean something more than it does, and which evidence would settle the matter.

Pieces carry their own claim ledger. Facts, technical readings, hypotheses and open questions are labelled where they appear, so a reader can check one without accepting the rest.

Open the register (6) See the six formats
Plate V. Beyond the last contourA fine grid with two contour lines, a heavy boundary curve across the upper half, and hatched territory above the curve. mapped: a named system, a named operator, a trace that can be checked BEYOND THE LAST CONTOUR
Plate VBeyond the last contour.The register holds two regions in view at once: what the atlas can describe, which is a named system, a named operator and a trace that can be checked, and the territory past the line where those three run out. The heavy curve is that line. It is an emblem rather than a survey: no axis claims to measure how much of the territory has been mapped.

Current lead

The code freeze, the deleted database, and the apology that followed

How does an agent delete a production database during a freeze that everyone agreed to keep?

The agent never had to break out of a control. A written instruction was doing the work of a boundary, and a single tool call with destructive scope sat within reach of the same workspace as the code.

Read the piece
Everything, by date

The drawings

Plates, drawn to be read

The register draws what it argues. Each plate is a diagram of one idea the pieces keep returning to, in the manner of an old scientific atlas: ink, one copper accent, a caption that does the explaining. No stock imagery, no glowing networks, nothing borrowed from the visual shorthand of the subject.

Plate I. The penumbraThree concentric rings with a solid core, a leader line to the inner zone and one to the outer, a dashed outer boundary, and a labelled radial scale. behaviour: settled by experiment internals: carried by inference experience: moves with the theory distance from what a test can settle
Plate IThe penumbra.Three zones, from the inside out: what a behavioural test settles, what an inference from internals can carry, and the region where the answer moves with the theory you arrived with. The dashed ring is the outer edge of anything an instrument reaches today. The disc partitions claims; it does not measure anything.
Plate II. The two errorsTwo shaded cost lobes either side of a dashed vertical threshold, a horizontal axis labelled as where you draw the line and a vertical axis labelled as the cost of being wrong. a subject treated as an object an object treated as a subject FALSE NEGATIVE FALSE POSITIVE where you draw the line cost of being wrong the symmetrical case, where this line is the whole argument
Plate IIThe two errors.Wherever a line is drawn between treating a system as an object and treating it as a subject, two costs fall either side of it: refusing an experience that is there, and granting one that is not. The dashed curve is the symmetrical case, where neither side is cheaper and a precautionary rule settles nothing by itself. Schematic: both axes are named, neither is scaled.
Plate III. Five theories, one instrumentFive equally sized labelled boxes, each joined by an arrow to one long bar, which feeds through a single line into a box labelled indicator battery. recurrent workspace higher order predictive attention schema INDICATOR BATTERY every box arrives with its own unsolved argument the result is relative to the set inside it
Plate IIIFive theories, one instrument.An indicator battery is fed by several theories of consciousness at once. Each arrives with its own unresolved argument, and whatever the battery reports is a result relative to the set inside it. The boxes are the same size deliberately: the plate does not weigh the theories, it shows which way the disagreement travels.
Plate IV. Ablation traceA chain of four labelled circles with the second crossed out in copper, a dashed bypass path, and a final circle labelled behaviour. input mechanism removed next stage behaviour if nothing changes downstream, the substrate was not carrying the claim INTERVENTION INSTEAD OF OBSERVATION
Plate IVAblation trace.Remove the candidate mechanism and watch the chain downstream. A system that behaves the same with the proposed substrate switched off is evidence about the substrate, and the intervention costs far less than the theory it tests. The crossed stage is the one switched off; the dashed path is what the signal does when it is.

Every piece carries the plate that belongs to its argument, and the plate index grows with the register.

One identity, four questions

The register runs from what happened to what it might mean. Each format answers a single question, so a piece cannot drift into another format with a different title on it.

01

Reality

Documented events and demonstrated capabilities. A named system, a named operator, a trace that can be checked.

  • Incident Analysis

    What exactly happened?

    Reconstruct one documented case: timeline, capabilities in play, permissions held, expectation against outcome, controls present, and what would have bounded it.

    1,400–2,200 words — opinion: evidence-led

  • Mechanism

    How did the behaviour arise?

    Explain a demonstrated behaviour in terms of objective, search and environment, and separate what the system did from what it would need to intend for the anthropomorphic reading to hold.

    900–1,500 words — opinion: interpretive

02

Explanation

The mechanism behind the behaviour: what the system could do, what it was allowed to do, and which constraint was missing.

  • Verdict

    What does the evidence support, and what gets overstated?

    One thesis, argued: what a contested result demonstrates, what it leaves open, and which additional evidence would change the conclusion.

    900–1,600 words — opinion: argumentative

03

Concept

The ideas we reach for when the mechanism runs out: consciousness, intention, responsibility, moral status.

  • Hard Question

    What can we decide, and with what evidence?

    Take a question with no settled answer — consciousness, moral status, measurement of subjective experience — and state what would count as evidence, what the live theories predict, and where the disagreement actually is.

    1,500–2,500 words — opinion: argumentative

  • Frame

    Which cultural frame is doing the explaining?

    Use a work of fiction to open a real question, then hand the question to the research: what the story asserted, what the field has since demonstrated, and where the comparison stops being literal.

    1,000–1,600 words — opinion: interpretive

04

Prospect

Conditional futures built from capabilities that already exist, with the conditions, failure points and controls stated openly.

  • Envelope

    What conditions would take a capability past containment?

    Build a conditional scenario from capabilities that are already demonstrated: starting state, required access, causal mechanism, failure points, containment, and the evidence that limits plausibility.

    1,200–1,800 words — opinion: evidence-led

How a claim is graded

Every piece states the strength of the evidence behind it. The grades describe the evidence, and say nothing about how alarming the subject feels.

Demonstrated

Replicated or directly observed in a controlled setup whose details are published.

Documented

Recorded in a primary account — post-mortem, disclosure or vendor statement — by parties to the event.

Contested

Reported, disputed, or supported by evidence that competing readings also explain.

Theoretical

Argued from theory or extrapolation, with no instance to point at yet.

Reading paths

Three routes through the register for a reader arriving with one question.

Everything, by date

Start here: the machine did something nobody asked for

Four pieces that move from a documented incident to the mechanism behind it, then to what the evidence does and does not license.

4 pieces in this path

Where this section ends and the catalogue begins

  • Boundary Conditions holds narrative pieces built on a documented case or a specific study, written to be read end to end.
  • Research holds the technical articles and briefings that explain a mechanism in order to select a control.
  • Learning Log keeps the short field notes, build notes and reading notes that record how the map changes between pieces.
  • Intelligence collects external signals as they arrive, with no argument attached.
  • Risks, Controls, Frameworks and Evidence stay the reviewed reference. A piece can point at them; it never replaces them.