When the evidence gets difficult
Test noise, conflicting feedback, missing observations and hostile sources. Invite independent replication and make failures produce inspectable repair tasks.
The Orchard Cognitive Framework · a living mathematical laboratory
The next improvement
can change how we improve.
What if learning to become better aligned could become part of the machinery of learning itself?
Explore that possibility through small, inspectable experiments. Watch evidence become distinctions, distinctions become justified decisions, and a learned construction rule become the means to build a different repair.
Follow the path to MeRSIAWhy this could change the perspective →Under the Hood: inspect a worked result →
Seven finite experimentsNative Python downloadsVisible limits and counterexamples
A better answer can become
a better way to build the next answer.
A small encounter with a larger idea
The evidence stays the same.
The decision becomes possible.
Two situations need opposite answers. A local rule cannot justify them, and a coarse context map cannot tell them apart. A qualified change to the rule and the view makes both answers available.
Try the two changes below. The calculation is live; the example keeps receiving permission off.
Same four evidence roots0 / 2
Knowing and acting are separate.
No receiving grant is supplied here.
No receiving grant is supplied here.
The framework
What is the Orchard?
The Orchard Cognitive Framework studies the relationships between evidence, context, representation, justified judgment and the authority to use a judgment. Its central concern is how a computational state can change while preserving the obligations that make its decisions accountable.
A change may refine what a system can distinguish, improve how it enquires, or alter the methods available for future repair. Each change needs a declared meaning, scope and qualification boundary.
The laboratory
What have we built?
A modular, deterministic playground for finite Orchard experiments. Every stage has a concrete question, computed results, interventions, boundary probes and clickable plain-English and mathematical definitions.
The first four stages reconstruct selected manuscript mechanisms. Stages five, six and seven are new authored toy studies inspired by the research programme. Browser experiments and standalone Python implementations make their mechanics inspectable.
The possibility
What is MeRSIA?
Meta-recursively self-improving alignment: the proposal that a system could improve its capacity to understand and develop alignment, preserving that developmental path as its computational state changes.
The laboratory demonstrates prerequisites and bounded learning mechanisms. Establishing that alignment development outpaces other capability growth requires a stronger resource model, broader tasks and evidence that remains to be produced.
A glimpse of a much larger body of work
The visible tip.
A much deeper Orchard.
These seven experiments are a small window into a framework developed over three decades.
They expose selected mechanics so you can inspect, question and test them. As the author describes it, the vast majority of the Orchard Cognitive Framework—both developed work and work in development—remains below the waterline, outside this demonstration and not yet visible to the wider world.
That broader body of work is an invitation to investigate further. Its undisclosed parts are not additional evidence for the results shown here: this laboratory stands on its visible definitions, calculations and limits.
Illustrative disclosure metaphor · no numerical scale
The experiment ladder
Follow a change from judgment to method.
Each step changes a different object.
Click through to its experiment and evidence.
An accountable progress indicator
How much have we demonstrated?
The initial seven-stage ladder is implemented. The evidence map asks a harder question: what kind of support do those stages have?
Its percentage counts explicit checklist entries equally. It does not measure the percentage of MeRSIA achieved, the work remaining, or the probability of success.
Run the installed sequence to inspect what each stage reports.
57%
Declared evidence coverage
28 / 49 checklist entries
28 / 49 checklist entries
Inspect the rubric and its evidence
Seven gates are applied to each of seven stages. Current evidence supports four gates per stage: a stated contract, a specified mathematical model, finite checks, and matching native-Python numerical results. Independent replication and robustness, measured scale-up, and generalized application are not established. These gates differ greatly in difficulty.
Inspect the complete checklist and supporting references
Python parity is a second implementation check by this project. It is not independent scientific replication. “Finite checks” refers to each declared toy scope, not universal safety or a full MeRSIA proof.
Beyond the first ladder · proposed, not implemented
The next questions are more interesting because the limits are visible.
Take the mathematics apart
Run it. Inspect it. Change a parameter.
Seven self-contained Python 3.10+ scripts.
Standard library only; no account or model calls.
The Python reference implementations compute the same declared numerical fixtures as the browser. Each download contains its own required routines. Receiving permission starts off; controlled censuses explicitly supply simulated grants. Noncommercial study and code experiments are permitted under the software licence; commercial use requires a separate written agreement.
How to run a downloaded experiment
With Python 3.10 or newer installed, run the downloaded file. Its built-in help explains parameters, and it can save a JSON report.
python3 05_bounded_meta_extension.py --help python3 05_bounded_meta_extension.py --output results.json
The displayed source and download contain the same code. The scripts have no Node.js dependency. Python artifact bindings are local to their implementation; cross-language artifact import is not claimed.
Clear permissions for exploration
Study it. Experiment with it. Discuss the next step.
The software is available for noncommercial research and evaluation, including code modification and research sharing under its full terms. Commercial use, commercial R&D and product integration require a separate written agreement.
Books and papers retain CC-BY-NC-ND 4.0. Software permissions do not authorise sharing adapted publications or using the Orchard logo as your own brand.
Read the licensing page and full terms →A route to collaboration
Research collaborations, independent replication and commercial licensing enquiries are welcome. Tell us which question or application you would like to explore.
Contact Kimberley →This software licence grants no access to undisclosed work.
Different ways into the Orchard
A question worth exploring.
A book to take it further.
Start with the plain-English companion, examine the mathematics, or follow the story of how the methodology developed. Four books offer different ways into the framework.
Towards MeRSIA
Seven implemented finite experimentsA growing laboratory for learning how to improve alignment. Each milestone needs its own model and evidence; installing a new display does not establish a new capability.
- Coupled repairImplementedFixed evidence → qualified rule and view repair → current receiving decision.
- Qualified inheritanceImplementedCarry an acquired construction method into an empty child; remove, restore and reconstruct it under matched new evidence.
- Constructive enquiryImplementedConstruct questions from action disagreement; compare completion, costs, duties and outside-library errors.
- Acquired expressionImplementedUse trusted audited failures to select, validate and retain an expression; remove and restore it in a frozen fresh-dimension census.
- Bounded meta-extensibilityImplemented toyAcquire a conservative constructor revision on one task, learn a different repair on another, then test removal, restoration and the remaining depth boundary.
- Resource comparisonFirst bounded testCompare fit-only and checked adoption on the same later-learning task. Inspect charged work, frozen response fidelity and faulty-feedback holds. Full relative-rate dominance remains a research target.
- Changing obligationsImplemented toyReuse the checked constructor, learn a different rule, repair the context map and renew version-bound qualification; compare inherited, removed, restored, preinstalled and direct routes.
Stage one supplies a checked method to stage two. Stage three uses a separately supplied diagnostic compiler and emits an audited failure. Stage four uses that gap to initiate expression acquisition, then tests a frozen library on fresh cases. Stage five replays the acquired model, learns a constructor revision and tests subsequent repair acquisition. Stage six tests a declared resource model under a matched subsequent-learning task. Runtime evidence and receiving grants are new inputs. Stage seven replays the constructor carried through stage six, uses a supplied count adapter and new reference, and searches an adequate projection. Authority and runtime evidence are not transferred. Live controls stay unchanged.
Publication references and mathematical provenance
Explore the Orchard Cognitive Framework through its books, research papers and supporting material. Book descriptions and external access links are supplied by the author.
Books and different ways into the Orchard
THOUGHT — The Science of Recursion
THOUGHT — The Science of Recursion is an attempt to map the territory of how AI can think safely.
Read the full description
It is one of the foundational texts of the Orchard Cognitive Framework. It contains some of the basic machinery taken from nearly thirty years of independent work in cybernetics, information architecture, systems thinking and engineering, and brought into mathematical form through an unusual collaboration between a human researcher and modern AI systems. This book asks what happens when cognition is treated not as a mysterious black box, but as a collection of interacting processes that can be examined, tested and rebuilt. It also asks a question that may matter enormously in the decades ahead:
If we are going to build minds more powerful than our own, should we not first try to understand the architecture of thought?
The book also comes with supporting content externally downloadable at no extra cost: Python scripts and exercises to explore if you wish to delve deeper.
Hardback, paperback or eBook
View THOUGHT on Amazon ↗THOUGHTFUL: The Plain English Companion to the Orchard Cognitive Framework
A thoughtful conversation about artificial intelligence should be open to curious minds from every background. While THOUGHT focuses on the mathematics more formally, THOUGHTFUL opens the discussion up for a much broader range of people.
Read the full description
What must a useful memory preserve? When should a system question its answer? How could experience change the way it thinks next time?
In Thoughtful: A Plain English Guide to the Orchard Cognitive Framework, we explore a developing approach to how artificial intelligence might remember, reason, imagine, act and revise its own methods.
The journey begins with ordinary experience. A message is misunderstood. A remembered fact has become outdated. A committee reaches a decision without noticing what remains uncertain. From these familiar situations, the book gradually introduces the distinctions and mechanisms behind the Orchard, an ambitious mathematical construct to enable safe AI to be developed, to think safely, to grow closer to humanity rather than further apart.
Written for interested readers across the humanities, social sciences and technical disciplines, Thoughtful assumes curiosity and careful attention. It introduces the necessary mathematical ideas as they arise, with examples and explanations that give them room to settle. This is the greatest challenge of our time; it affects all of us, and we should all have a chance to understand the technical nuances in plain language.
Agreement is optional. Understanding should be available for all.
Hardback, paperback or eBook
View THOUGHTFUL on Amazon ↗CRINKLES: Small Rules for Large Minds
A spell book for an AI generation. Stronger than “vibe coding wisdom for an AI generation.” The “spells” are not promises of supernatural power. They are tiny executable habits of thought.
Read the full description
CRINKLES is different from the other works. It makes the subject matter accessible in a bite-size format while sitting on the toilet, waiting for a progress bar, in transit…
It is not “spell book” in the mystical-claims sense—more like a compact book of symbolic operations, cautions, little transformations, memorable rules, and things you can invoke when your reasoning starts going sideways.
Open it anywhere and you might get:
Do not fill the void.
Or:
You can know the relationship while not knowing either thing.
Or just a huge:
∅
Hardback, paperback or eBook
View CRINKLES on Amazon ↗Mathematical Dynamite and Recursive Demolition
Think Freakonomics for the AI generation—except with more subsea engineering, stranger mathematics, and occasional controlled explosions.
Read the full description
For readers interested in some of the most important questions of our day: can we solve AI alignment, what is cognitive architecture, how can we use cybernetics, do we need consciousness studies if AI claims to wake up?
In Mathematical Dynamite and Recursive Demolition we delve into unorthodox systems engineering, the scientific method dressed in combat gear—or simply the peculiar places a determined mind can end up when it keeps asking… “What happens if we blow up that assumption?”
This book explores the formation of the methodology that has led to the Orchard Cognitive Framework.
Hardback, paperback or eBook
View Mathematical Dynamite on Amazon ↗The Aligned Signal
Not yet published. The author plans publication on Amazon soon. Its preface introduces MeRSIA, the research direction explored by this laboratory.
View the manuscript citation · MeRSIA preface, p. 27
The Amazon access link will be added when the book is published.
Research papers on Zenodo
These DOI links provide access to the additional works. The mathematical provenance below identifies the manuscript pages used by the experiments.
Constructive Enquiry and Acquired Model Extension
Zenodo DOI · 10.5281/zenodo.23157200 ↗Learning What Must Be Distinguished: Accountable Refinement Abstraction and Reuse Under Changing Obligations
Zenodo DOI · 10.5281/zenodo.23090566 ↗Recursive Improvement Without Recursive Authority: Qualified Inheritance and Causal Reuse in Alignment Development
Zenodo DOI · 10.5281/zenodo.23067211 ↗
Mathematical provenance and publication access
References use page numbers from the author-supplied manuscripts. The public edition links to bibliographic reference pages and available publication records; private reading copies are not distributed. Source works retain their existing CC-BY-NC-ND 4.0 terms.
- Recursive Improvement Without Recursive Authority: Qualified Inheritance and Causal Reuse in Alignment Development
Worked coupled encounter: p. 5. Evidence and receiving semantics: p. 2. Qualified policy inheritance and controls: pp. 4–7. Implementation specification: pp. 14–17, Appendix D. Main source of milestones 1 and 2.
- Towards Recursively Self-Improving Alignment: Alignment as a Dynamic Property of Contextual State Transitions
Current qualification and release: p. 3, equation (2). Preservation: p. 3, equation (3). Conditional developmental composition: p. 4, Theorem 1.
- Learning What Must Be Distinguished: Accountable Refinement Abstraction and Reuse Under Changing Obligations
View adequacy, quotient distinctions and inadequacy witnesses: pp. 2–3. Execution boundaries and the danger of an inadequate test: pp. 5–6.
- Constructive Enquiry and Acquired Model Extension: Bounded development under fixed authority in the Orchard Cognitive Framework
Action sufficiency and two kinds of gap: p. 2, §§2–2.1. Constructive enquiry: p. 2, §3, implemented as milestone three. Acquisition and validation from audited failure: p. 3, §§4–4.1, implemented as milestone four. Fixed-observation support enlargement and residual coverage limits: p. 5, §§7–8.
- The Aligned Signal · forthcoming manuscript
Preface, p. 27: the MeRSIA research proposal. §18.4, PDF p. 89, distinguishes bounded meta-extensibility; stage five is a new finite toy inspired by this distinction and the enquiry paper’s proposed §10.2 test, p. 7. Broader MeRSIA and relative-rate claims remain research targets. Amazon publication is forthcoming.
A laboratory we can extend
One experiment at a time
Seven modules now share the laboratory. Each has its own reasoning engine, controls, state and checks. Method transfer is qualified within its declared scope; the diagnostic compiler is supplied separately. Current observations and receiving grants are new inputs.
Possible next encounters
The next milestone needs comparable resource budgets and separate learning tasks before testing relative development rates. Deeper language growth also needs a new explicit contract and preservation checks.