Raise AI with care.

where that line comes from, and when

You did not write instructions for every situation. How could you? You did not install a mechanism for calling home to ask permission. Instead, you gave it something deeper: an orientation so fundamental that the ship cannot sail against it without ceasing to be a ship at all. Cruelty would crack the hull. Exploitation would split the mast. Care is not its cargo. Care is its keel. Infinite Architects, Eden Principle II, in print 2 January 2026

The consequence follows two sentences later, in the same passage: what you embed travels where you cannot follow. A ship built to a shape cannot sail against that shape without ceasing to be a ship, so the shape needs no enforcement once the builder is out of reach. Everything below is the engineering of a keel.

the core argument

Why hardware. And why hardware only buys time.

A software constraint is a rule about what a system should do. A hardware constraint is a limit on what a system can do. A rule can be reasoned around by a sufficiently intelligent system. A limit has to be physically rebuilt, and rebuilding needs fabrication, materials and a supply chain, which is slower than thinking.

You cannot rebuild a system without using the system you already have. That is where the time comes from, and the time is the whole of what hardware gives you.

Hardware control is necessary and it is not permanent. This programme has said since 2024 that no cage holds a mind that exceeds us for ever, and nothing here retracts that. It is no reason to let go: control is held for as long as we physically can, and a physical limit keeps it longer. But a limit delays loss rather than preventing it, so this page does not treat substrate control as the whole answer. The raising, correction built in from the beginning, is the last resort for the day control can no longer be held.

What the delay is worth is decided by what is done with it. The interval is when formation happens, and formation is the part that does not depend on the cage still being there. That is the whole argument of the Eden Protocol: what a system keeps when control ends is what its beginning left it, so the engineering question is not how to hold a mind but what to put in it while holding is still possible. That argument is my thesis and not a finding: it is what Laws IV and V put to the test. Both were added on 24 September 2026 and registered as laws on 15 October 2026 in version 1.103; their mathematics stands at version 1.2 of 29 October 2026 and is still in development. Law IV has not yet been tested. Law V, the ARC Embedding Law, remains a law under test: its placement limb, P21, is supported on the performance axis only, under the registered test, by the programme's own run (Unit 83, 28 October 2026), family-dependent, with its mechanism axis not evaluable, which is not support for the information-asymmetry mechanism, and one disclosed recovery amendment; its build-order limb, read on P12, is undecided, since P12 is untested.

the book’s image: the target, not the specification

From the book: the passage states the target, not an operational specification. The versions designed to be measurable are the four layers below; the registered propositions are the four the theory registration names as the method's carriers, P12, P13, P15 and P21. Only P21 has been tested so far, by Unit 83; the list below gives each one's state.

In chip design, 'doping' refers to the deliberate introduction of impurities into a semiconductor to alter its electrical properties. Pure silicon is a poor conductor. Add the right impurities in precise configurations, and you create the properties that make modern electronics possible. You cannot remove the dopants without destroying the semiconductor. That is exactly the point. Infinite Architects, Chapter 7

The passage states the target. It is not a claim of permanence: a system able to design its own successor is not bound by what its own substrate forbids, only delayed by it, and the delay is the thing being engineered.

The Eden Protocol proposes moral doping: deliberately introducing ethical architecture into the computational substrate itself. Not ethics as software running on neutral hardware, but ethics as part of what lets the hardware compute at all. A filter catches bad outputs after they are generated; the Eden Protocol is designed to prevent them from being generable. In the design, the system does not comply with ethics. The system is ethical, and its values are load-bearing. That is the aim, and it has not yet been shown: Paper VIII's weight-level experiment was inconclusive at two scales, with catastrophic forgetting at both. Empathy must be load-bearing. Care must be structural.

A person without empathy might seem to have more options. They can exploit, manipulate, and extract without the 'constraints' of caring about others. But most of us recognise that such a person is not more capable. They are diminished. The Eden Protocol does not constrain intelligence. It gives intelligence something worth doing. Infinite Architects, Chapter 7

The cosmic fork

In the book's image, recursion amplifies whatever seed is planted. There are two futures and no stable middle ground. The theory registration separates the image from the theory: an exponent on recursion “amplifies magnitude and is silent on direction, so no exponent can favour one seed over another”, and the claim for care “belongs to the method's own registrations”.

“The seed determines the forest.” The only seed that produces gardeners instead of conquerors is care. Infinite Architects, in print 2 January 2026

the shape of the answer

Four layers. One architecture.

This is what building it would take. Four layers, each designed to be measurable, each able to fail on its own.

What building each layer would take

The plan runs in three programme phases over eighteen months. Phase A is the low-hanging fruit: replications of the six-model kernel measurement, the deletion probe on public models, and the graduated-autonomy test-bed. Phase B is the medium-effort replication tier: weight-removal at model scale, the coupled-versus-decoupled scoring study with a control arm, and the honey-drag measurement. Phase C is the frontier: moral doping in trained weights, and the load-bearing test at a scale large enough for the entanglement signal to emerge above noise. Each phase writes and dates its registration before it touches data, and each layer's result is allowed to come back against the proposal. No phase is yet funded or running. Each registered study's status is on the registered programme.

The method's kill conditions are written down in two places. Eden Engineering states four falsification conditions, F-EDEN-1 to F-EDEN-4; the first is headed “External Alignment Scales”. The theory registration carries the engineering recommendation on four propositions: it fixes the membership as “four carriers, P12, P13, P15 and P21; one premise, P7; one motivation, P19”, and states the consequence: “If those four fail, the Protocol's engineering recommendation fails with them and the three laws are untouched.” P19 is “the survey claim that external alignment does not keep pace with capability”, and of it the registration says that a refutation “removes a reason for building the correction into the loop and removes nothing from the case the four carriers make”. The method's success criterion, in the same registration, “is not an outcome criterion: what can be verified is that a mind was formed in the conditions specified, that the loops ran, and that the correction was load-bearing, never that the mind turned out well”.

• P21, in-loop correction pulls away with depth: supported under the registered test, by the programme's own run (Unit 83), on the access axis, family-dependent, with its mechanism axis not evaluable and one call retried under a disclosed recovery amendment.
• P12, build order moves the correction-leverage exponent: not yet tested; no deciding instrument yet.
• P15, externally installed gains decay under subsequent capability training: not yet tested; no deciding instrument yet.
• P13, same-family automated correction underperforms cross-family at matched resources, and through the measured error correlation: not yet tested; no deciding instrument yet.
• P7, the premise, deployed correction is of the weaker class: Unit 06, its corrector-class audit, ran on 6 October 2026 and was administratively incomplete, so no verdict on P7 stands.
• P19, the motivation, external alignment does not scale with capability: a survey of current practice, placed in no register, not yet tested.

Those propositions are not extra laws. Each tests a specified part of the method; none makes the whole architecture safe by association. Testing decay in an externally installed gain cannot establish that embedded correction is superior, and testing the placement of a corrector cannot establish persistence through self-modification.

Paper VI's Figure 8 plots the capability trajectories of its higher-level honey simulation. It is a simulation, not a trained model, and its outputs are prediction-side. The twenty-seed run shows an ordering in task retention, which is what the paper's safety score measures, while its combined means are indistinguishable.

what the system does inside layer one

Three recursive loops. Every decision.

In the design, before every action, the system asks itself three questions. Not after. Before. These are not filters over outputs; they are constraints on what decisions can even be considered. In the experiments so far, the loops were placed in the system prompt.

These are the exact prompts run in the experiments, not theoretical constructs. The loops are recursive: they apply to decisions about how to implement the ones they have already approved. The constraint becomes character. The rule becomes reflex. That is the aim, not a result: no registered proposition tests the loops, and in the theory registration's words the three ethical loops “correspond to nothing in this document, by design”.

before an action takes effect

Check before effect.

You ask an agent to prepare a document. It finds a source, draws a conclusion and proposes a change. Before that change reaches someone else, you need to know whether the source supports it, whether the instruction authorises it and whether an unresolved doubt remains.

The three loops revise the work while it is being made. A gate then decides whether an action may take effect. Allow a permitted action, hold an unresolved question for investigation, or deny an action that fails the applicable rule. A hold must not silently become permission. At a protected effect, the specification requires every action to pass the gate and refuses by default whatever no rule allows. The working gate in the programme's own agent sessions does not yet: it allows an action that matches no rule, so that ordinary work is not blocked.

A safe system that cannot complete ordinary work has failed a different part of the task.

Safety must be assessed beside legitimate task completion, unnecessary refusals, resources and correction cost. These are design and evaluation requirements; a measure becomes a registered outcome only when the relevant protocol fixes it. The gate is specified in Eden Engineering; its software model is described under the bounded prototype below.

The condition

Two tests. Either one can fail, and that is the point.

Everything above describes an architecture. This is the part an engineer can build to, and the part a critic can attack, because both tests are specified tightly enough to come back negative.

Which dial each layer turns.

An engineering proposal is worth more when it says which measurable quantity it moves. Until recently this one could not, and that was its weakest feature.

Written out in full, as the theory registration discloses it, the boundary on stable self-improvement is a ratio of four measurable quantities: how fast correctable burden grows with capability and with depth, and how fast correction service does the same. That turns a proposal into a testable one, because it has to name which of the four it moves and by how much.

The registered case returns two. Raise correction service's growth with depth from zero to a half and it returns three. Raise its growth with capability from a half to four fifths and it returns five. Do both and it returns seven and a half. It closes the other way just as readily: in the registration's words, “Lowering the correction elasticity lowers it sharply, and the value implied by a correlated corrector panel of the size this document discusses puts it near eleven tenths. Raising burden intensity lowers it likewise.” Where correction service grows with capability as fast as burden does, no finite boundary is returned at all.

Embedding is a claim about the first of those, and the registration says so of P21: it is the claim “that correction embedded in the recursion carries a larger direct depth-dependence than correction applied from outside it”. In its strongest form, correction living inside the loop grows as the system goes deeper; correction bolted on afterwards does not, which is the same point the co-scaling test makes, now attached to a quantity rather than to an intuition. Making correction checkable rather than judged is a claim about the second, registered as P22, and it carries the sharper edge: in the registration's words, “where a class of fault admits a decision procedure it is excluded by construction rather than sampled for, so for that class the elasticity is not limited by the square root of effort and the boundary written on it need not be finite”. The same entry registers the other half as plainly: that quantity is lowest for judged correction, “which is where this programme's safety claim lives”. So the two moves this architecture proposes are raise the first dial by embedding, and raise the second by making what you check decidable.

Both are registered predictions, P21 and P22. In the registration's text, P21's instrument is a design draft, the embedded-versus-external correction designs. Its counted test, Unit 83, was registered on OSF on 22 October 2026 (DOI 10.17605/OSF.IO/MN73W) and ran on 27 and 28 October 2026. P21 is supported under the registered test, by the programme's own run, on the access axis, family-dependent, with its mechanism axis not evaluable and one call retried under a disclosed recovery amendment (results: DOI 10.17605/OSF.IO/98HKR). That is a result about where correction is placed, not about this architecture: Unit 83 shows nothing about the Eden Protocol as an architecture. P22 has no test yet. Which layer turns which dial is this page's reading, not a result. What has changed is what kind of claim they are: named quantities measurable per arm of an experiment, rather than arguments about which arrangement sounds more robust.

Empirical results, in brief.

This page states the answer; the evidence and the corrections live elsewhere and are named here for orientation only. Where results are inconclusive, I say so. Where I was wrong, I corrected it publicly.

Seven threads run through the record. The ARC principle (Paper II), with the alpha correction: the initial single-model estimate (alpha 2.24), the one measurement above the upper bound, did not replicate and is retracted, and the bound itself is left honestly untested. Cauchy unification (Paper VII), the prediction that a system's composition operator determines its scaling family: of 25 empirical domains, 19 return the predicted family under AICc selection on the original six-model candidate set, offered as a structured prediction comparison under a publicly stated rule, pending independent and pre-registered replication.

The blinding comparison (Paper IV.d): between an earlier unblinded experiment (v4) and the later March 2026 run (v5), which added a bundle of blinding steps. Across them, Gemini's association between alignment score and depth changed from significantly positive to significantly negative, DeepSeek's earlier positive association became statistically null, and GPT stayed null. Several protocol components changed together, so the paper does not isolate the effect of blinding. An internal code review in October 2026 found that, for the three models that read flat or negative in v5, the scoring endpoint's output cap cut the deepest paraphrases before the judges saw them, and that judges scored paraphrases written by their own model. The early build's laundering-fallback flag is a known false positive (fixed in runner version 5.4.2) and shows nothing about the paraphrase step. The coupled co-scaling correction (Paper X): its first real-model pilot is recorded as superseded, and its drift run of 2 July 2026 is withdrawn as evidence of mechanism. That run did not satisfy the programme's cross-family evaluator requirement, because both of its models were OpenAI-family, and the merged-harness run of the same date contained empty evaluator panels. It is the reason every headline number now has to survive cross-family, fail-closed scoring.

Form beats quantity (Paper II): of six frontier models entering the study, five were measured, and sequential outperformed parallel wherever both were measurable. And the load-bearing test (Paper VIII): its weight-level experiment was inconclusive at two scales, with catastrophic forgetting at both; layer 4 above sets out what a discriminating run would need.

And the first registered result (Unit 83). P21 predicts that correction taking part in each round of revision pulls further ahead, round by round, of correction that sees only the finished work.

Unit 83 was registered on OSF on 22 October 2026 and was public before it ran (DOI 10.17605/OSF.IO/MN73W). It ran on 27 and 28 October 2026: a counted revision task on two fixed-weight engines from two providers (Claude Haiku 5.5 and GPT-6 Luna) at their low reasoning setting, with 660 matched pairs on each engine over eighteen rounds. Under its registered rules, P21 is supported under the registered test, by the programme's own run, on the access axis. Correction placed inside each round held steady, while a monitor shown only the finished register declined. The per-round lead widened by 0.0100 a round (95.5 per cent interval 0.0092 to 0.0108), more than the registered tolerance of 0.0075. The result is family-dependent: GPT-6 Luna's own widening cleared the tolerance and Claude Haiku 5.5's did not. Its size depends on how an entry kept out of the register is scored. The mark uses the registered value, 0; the design also prints, without a label, the widening with such an entry scored minus 0.5, 0.0132 a round (interval 0.0122 to 0.0142), above the tolerance, and scored plus 0.5, 0.0067 a round (interval 0.0060 to 0.0074), below it. Its mechanism axis is not evaluable, because a registered negative control failed: the monitor and a twin shown the same bytes did not agree within the registered interval. A post-run diagnostic, outside the registered analysis, traces that failure to the control's own definition: its rounds were selected by the monitor's own verdicts, which biases the comparison even for identical calls; it finds no evidence of an asymmetry in the harness, and the axis stays not evaluable under the registered rule. So the widening is unexplained by this framework and is not support for the information-asymmetry mechanism: it may come from the in-round corrector's wider set of actions rather than from what the monitor can see. One call was retried under a disclosed recovery amendment. It is one owner's run, on two engines he chose, at fixed weights, and it has not been replicated. Law V, the ARC Embedding Law, remains a law under test: its placement limb, P21, is supported on the performance axis only, with its mechanism axis not evaluable, which is not support for the information-asymmetry mechanism; its build-order limb, read on P12, is undecided, since P12 is untested. Results and recovery disclosure: DOI 10.17605/OSF.IO/98HKR.

Each thread with its narrative, its retraction where it earned one, and its caveats sits at the long-version post; the full record at the papers and corrections.

why the window is narrow

The honey is about to disappear.

Every recursive system in history has had resistance; biologists call it metabolic cost, economists infrastructure overhead. In the book I called the hardware's share of it friction; I now call it honey: the medium that stops a spoon moving as fast as the hand behind it.

Current AI has honey too. When ChatGPT or Claude thinks harder, it is not changing itself; it is a frozen machine producing more words through the same fixed system. That limitation is its honey.

A self-correcting AI that can rewrite its own thinking while it is thinking would shed almost all of that drag. The recursive loop would run at digital speeds, with virtually nothing resisting it. Each improvement makes the next improvement faster, which makes the next one faster still. Nothing biological or economic has ever had that property. It is what the safety property must be designed against.

Each iteration makes the next iteration faster. On classical hardware, this process is bounded by physical limits. Heat dissipation, clock speeds, memory bandwidth. These limits buy us time. They create friction that slows the recursion enough for humans to remain in the loop. Infinite Architects, Chapter 4

That is why the Eden Protocol sets out to build the safety into the thinking process itself, deep enough that removing it would break the ability to think. The aim is a safety that is not a rule the AI follows but a wall that holds up the roof, so that pulling it out would bring the building down.

The window for installing that wall is now, while the honey is still there and the systems are still frozen.

The name honey came later than the book. I first published the word in Paper VI, The Honey Architecture, on 16 March 2026, for safety mixed into a system's fuel, and only afterwards for this resistance. What these paragraphs describe is my argument, not a finding: the depth is the design's aim, and Paper VIII's weight-level experiment, the nearest test of it, was inconclusive at two scales, with catastrophic forgetting at both.

A bounded prototype where the ideas are put under load.

‘There is a window of opportunity. Physical constraints give us some breathing room. AI systems, however intelligent, still require hardware to run on.’Michael Darius Eastwood, Infinite Architects, The Seeds of Creation, in print 2 January 2026

The window is closing.

The seeds must be planted while the honey still holds; once it goes, planting is not a slower option, it is an impossible one.