KFD Foundation

KFD Foundation

KFD is not a model architecture or an agent framework. It is a foundation for systems that remain correctable by reality, including when reality challenges the objects and primitives from which their world models are built.

This document explains the public worldview shared by the numbered KFDs. It is not a numbered decision and does not create an additional normative rule. When an explanation and a decision differ, the text in decisions/KFD-N.md remains authoritative.

The dated KFD Under Load companion follows this worldview into the founding Kungfu implementation. It is an evidence synthesis: it connects repeated dogfood pressure, implementation witnesses, confidence updates, and open falsifiers without creating a new decision or treating one adopter as universal proof.

Continuity under uncertainty

KFD uses continuity as a bounded engineering concept, not a metaphysical identity test. Continuity does not mean preserving a fixed state. It means that a system’s successor state remains inspectably connected to the facts, boundaries, perspective, direction, authority, responsibility, realized action, and admitted corrections through which the change became its own.

Two different uncertainty conditions matter:

1. When the current ontology is adequate, the system needs an action and feedback loop. KFD-1 through KFD-4 and KFD-7 preserve the path from a declared Fact cut through bounded trust, participant cooperation where applicable, declared perspective, direction, and authority, realized action, review, correction, and a successor Fact cut.
2. When the current ontology is inadequate, the system also needs a discovery loop. KFD-5 and KFD-6 preserve candidate genesis, qualification, causal grounding, and return to action without allowing a generated model to certify itself.

declared Fact cut
  -> bounded trust
  -> cooperation where multiple reasoning participants are involved
  -> declared perspective, direction, and authority
  -> action and causal record
  -> review, correction, and admission
  -> successor Fact cut

when the current ontology fails:
causal experience
  -> candidate genesis
  -> fact-bound qualification
  -> revised ontology
  -> return to the action and feedback loop

The action loop is the ordinary operating path. Primitive discovery is the hard-case extension used when reality invalidates the objects through which that path was described. Neither loop is sufficient alone: action without correctable ontology can preserve the wrong world, while discovery without return to accountable action remains an untested narrative.

Better answers are not always enough

Most progress improves an answer inside a world we already know how to describe. Some progress changes the objects with which the world can be described.

A spreadsheet did not merely calculate faster. It made cells, dependencies, and recalculation into objects a person could manipulate directly. A Git commit did not merely store files. It made a distributed change, its parents, its content, and its authorship into an object software could preserve, compare, and exchange. A log did not become important because append-only records were new. It became transformative when an internal mechanism could be treated as a first-class record of what happened and projected into many systems.

KFD calls such a load-bearing object a primitive. A primitive gives a recurring part of reality an identity, boundary, authority, lifecycle, and set of operations. Once it exists, new questions become cheap to ask and whole classes of reconstruction work can disappear.

a fact answers inside an existing world
a primitive changes what that world contains

This is why primitive discovery matters to knowledge. Knowledge does not grow only by accumulating more facts or improving predictions. It also grows when a previously implicit object becomes explicit enough to reason about, test, share, and build upon.

Why primitive discovery remains rare

Missing primitives are difficult to see from inside inherited vocabulary. If a database-shaped view is painful, local optimization usually produces a better database. If a trace view is painful, it usually produces another query or visualization. Both may be useful while leaving the missing object outside the visible horizon.

Historically, a candidate often became visible only when several conditions met in one place:

• a participant lived inside a saturated field of real consequences;
• that participant occupied a perspective that integrated pressures others encountered only separately;
• a forcing event made the old workaround impossible or no longer tolerable;
• the participant had enough judgment and authority to give the new object a durable form.

The events and artifacts could be distributed across an organization. The integrated perspective usually was not. Discovery therefore depended on a rare biographical convergence rather than an inspectable process.

Git: ten days after years of pressure

The Linux kernel exchanged changes as patches and archived files from 1991 to 2002, then used BitKeeper until that relationship broke down in 2005. Linus Torvalds later described Git as self-hosting after about one day and usable for his first kernel commit after about ten days. He also said that the decisive work was not the amount of code but getting the data organization right, after years of seeing what existing systems made difficult.

The asymmetry is the point: construction took days; the pressure field and the perspective able to compress it took years. Git was not the sudden invention of every technical ingredient it used. It joined prior lessons, integrity needs, distributed authority, nonlinear history, performance constraints, and a workflow containing tens of merges a day into one load-bearing object model.

An evidence cut is the declared boundary of evidence available to a claim or decision at a given point. Keeping that cut stable prevents later knowledge from being silently projected backward into the decision.

BitKeeper’s withdrawal can be replayed through the KFD lens as a forcing evidence cut: a condition that could previously be endured became a decision that had to be made. That is a KFD interpretation of the historical record, not terminology used by the historical actors themselves.

The historical cases companion develops this case and compares it with VisiCalc, the log as a standalone architecture primitive, and an ordinary cross-machine trace view.

What history does not provide

Historical creators often left code, design explanations, interviews, and retrospective accounts. These materials can explain why a result worked. They do not yet form a procedure that another participant can repeatedly execute to discover the next missing primitive.

There is no generally adopted standard pipeline that preserves situated pressure, makes perspectives replayable, compares their natural objects, separates candidate genesis from qualification, and keeps promotion answerable to facts and later consequences. The result survives; much of the path by which the object became visible remains dependent on biography.

KFD addresses that missing path. It does not promise deterministic invention. It asks whether the conditions of discovery can become durable, inspectable, and increasingly repeatable without flattening the reality that made judgment possible.

Why the foundation begins with procedure

KFD-1 is a procedure rather than a principle by necessity. A fact-grounded system cannot place another fact claim beneath its facts without restarting the demand for justification, and it cannot use a self-certifying axiom without asking participants to trust what the system is meant to keep inspectable. KFD therefore chooses a maintained practice as its operational ground: declare the fact source, register welded surfaces, classify change, and run the gate. “Facts must not drift” carries weight only while those acts continue to be performed and audited.

This gives the first six KFDs an intentional practice-invariant-practice shape, and KFD-7 begins the action layer built over it:

KFD-1       foundational practice that maintains the fact base
KFD-2/3     foundation principles for trust and cooperation
KFD-4       first derived principle for perspective-bound action
KFD-5/6     discovery procedures that extend the describable world
KFD-7       action principle that uses the extended world without fusing
            direction, perspective, authority, or occurrence

KFD begins in procedure, states the invariants under which facts can support trust, cooperation, and perspective transformation, then returns to procedure to extend the world those invariants can govern. The two practice layers are not illustrations around a propositional core. They express KFD’s operational claim that knowledge remains load-bearing through maintained, inspectable, and correctable practice.

Foundation structure

The first three KFDs are intentionally ordered:

In short:

reality pressure -> non-drifting facts -> inspectable trust
                 -> trusted value -> voluntary cooperation

KFD-1 makes fact sources operational: a fact-bearing contract world must be declared, inspectable, and unable to drift invisibly. KFD-2 defines how trust can stand on those facts. Where a scope involves multiple reasoning participants, KFD-3 defines how humans and agents can cooperate once facts, responsibility, and value are visible. A non-cooperative scope may declare KFD-3 not applicable with a bounded rationale; it may not silently erase participant-facing dependencies.

KFD-4 is the first complete operating application of that foundation:

KFD-1 makes timelines evidentiary.
KFD-2 makes replay trustworthy.
KFD-3 makes contrastive replay cooperative.
KFD-4 makes perspective transformation operational.

A timeline without non-drifting facts can become retrospective narrative. A replay without inspectable trust can become an unverifiable simulation. Views without cooperation can remain isolated even when each is locally valid. Competition or independent action may diversify perspectives; cooperation turns their trusted differences into shared discovery. KFD-4 depends on all three foundation principles without becoming a fourth foundation principle.

Why KFD-4 is the first derived operator

The placement is structural. Once the foundation preserves non-drifting fact cuts, makes reconstruction trust assessable, and supports trusted exchange among participants, multiple declared positions become available. The smallest new general operator is then to change position without discarding the fact base:

non-drifting fact cuts
  -> assessable replay trust
  -> trusted exchange among participants
  -> multiple declared positions
  -> perspective transformation

Under that installed foundation, perspective transformation is the lowest-marginal-cost default probe for moving the observation horizon. It can reuse preserved evidence and add only missing capture, replay, and contrast, rather than necessarily waiting for repeated failure, introducing a causal intervention, or searching a large corpus under a new compression objective. This is a bounded engineering hypothesis, not a universal ranking theorem.

The relevant comparison is total cost per qualified Primitive:

qualified cost =
    genesis
  + qualification
  + false-candidate handling
  + real-world validation

Perspective transformation may lose when source views are unavailable, privacy or reconstruction dominates, replay creates many false candidates, or a domain-specific anomaly, causal, or compression method is cheaper. KFD-5 qualifies the result; KFD-6 must compare the methods over causal experience.

From action to discovery and back

KFD does not reject stronger models. It rejects both the assumption that model improvement is sufficient for every next step and the assumption that discovery is complete before a system can act responsibly.

capability loop: model -> prediction -> evaluation -> stronger model

action loop:     facts -> trust -> bounded action -> consequence
                      -> review -> successor facts

discovery loop:  causal experience -> candidate genesis -> qualification
                                   -> revised ontology -> action loop

The loops are complementary, not interchangeable. The action loop carries ordinary work whenever the current ontology remains adequate. Stronger models remain necessary construction material, but capability alone does not decide whether the current object set is adequate. Discovery does not always begin with repeated failure inside the current model. Repeated observation through a database-shaped ontology may tend toward a better database-shaped answer, but anomaly, reconstruction pressure, causal-variable discovery, structural compression, direct judgment, and perspective transformation can all generate candidates when their search space permits it. KFD does not yet have comparative evidence that one method dominates the others.

KFD-4, KFD-5, and KFD-6 make that motion explicit:

declare a situated perspective
  + declare the current ontology
  + run one or more genesis methods
  -> propose or reject a candidate object
  -> qualify the candidate against facts and consequences
  -> return to reality through action

KFD-4 says that every view remains bound to a perspective and that perspectives can be intentionally transformed. That makes perspective transformation an auditable method, not an exclusive generator. KFD-5 separates perspective-declared, method-plural genesis from fact-bound qualification. KFD-6 asks whether an agent can eventually compare and internalize those functions from causal experience without replacing reality or certifying itself. A qualified ontology change returns to KFD-7 action; it does not become a detached research endpoint.

Engineering witnesses

Buildchain is the first dense engineering witness for the foundation triad. It keeps release facts from drifting, binds trust to inspectable passports and responsibility, and exposes cooperation constraints through declared interfaces rather than hidden process.

Kungfu is the first product witness for KFD-4. Its public product category remains Agent Runtime; its deeper role in the KFD model is expressed by this claim:

Kungfu is an infrastructure for reality-preserving perspective transformation.

Episodes preserve bounded causal experience. Observer-bound timelines make a perspective durable. Replay makes a perspective transferable with declared loss. Contrastive replay compares multiple views without collapsing them into an absolute history. Together they lower the marginal cost of perspective change while preserving facts, consequences, and degraded state.

KFD-5 and KFD-6 describe how to use that capability without making it the only path. KFD-5 lets humans and agents turn cross-perspective mismatch, anomaly, reconstruction burden, candidate causal variables, structural compression, or hybrids into qualified Primitive candidates. KFD-6 asks whether an agent can compare those methods and perform discovery and qualification autonomously across large bodies of causal experience.

Real-world agent work

Real-world agent work turns ordinary work into a dense system of products, files, repositories, traces, policies, humans, and agents. Complexity cannot be made safe by hidden state or forced compliance. It has to be compressed through non-drifting facts, inspectable trust, voluntary cooperation, and perspectives that expose how reality was projected.

This changes what a product interface means. Agent-facing CLI, API, documentation, envelopes, and local fact surfaces are not secondary integration channels after the human GUI. They are first-class interfaces for intelligent participants. A control plane in this model is a shared work environment for humans and agents, not only a human dashboard over agent activity.

This model used to be expensive to practice. Stable facts require engineering, iteration, and disciplined evidence paths, so older systems often survived by leaning on cheaper substitutes such as authority, habit, reputation, or intuition. Agents change both sides of that equation: they make the world more complex, and they also provide more intelligence for building fact-bearing systems. KFD is part of that bootstrap.

Load-bearing product witness

KFD does not ask readers to believe its worldview. No principle is load-bearing until it is embodied in an inspectable product witness: an implementation that humans and agents can use, audit, challenge, and compare with its claimed consequences.

Self-evident does not mean simple or documentation-free. It means that a product’s declared value, mechanisms, constraints, and residual risks can be traced through its public interfaces, source, build and release evidence, and runtime facts without depending on hidden organizational context. Open source is necessary for Kungfu’s proof path, but source availability alone does not create a witness when the product cannot be used at its declared surfaces or its claims tested against consequences.

Kungfu products are not illustrations placed beside KFD. They are the concrete surfaces through which KFD is made responsible to reality. Their breadth comes from the closure required by that responsibility: facts, trust, cooperation, perspective, action, consequence, and discovery must remain connected across the interfaces that humans and agents actually use.

That closure requires substantial engineering, but complexity is not evidence by itself. Complexity is justified only when it preserves a KFD invariant, reduces friction for human or agent participants, or produces an inspectable product witness. Complexity that cannot carry one of those responsibilities remains debt.

Existing capabilities and the passage

The new direction does not reject the existing technical world. Kungfu depends on capabilities accumulated through open source, version control, programming languages, operating systems, build systems, and high-capability agents. They are the departure ground and construction material for the passage into a new reality-correctable world.

The difference is the responsibility under which those capabilities are organized: keep the shared world correctable by reality, and reconstruct its ontology when the old map fails. Kungfu’s product program is to open that path for humans and agents. Whether the path carries weight is judged by inspectable product consequences.

Practice guidelines

KFDs after the foundation triad may define derived principles or procedures that apply the foundation to real-world behavior. They do not expand the foundation triad; they show how a fact-bound, trustable, cooperative system can change its view, discover missing objects, and remain correctable while doing so.

The continuity is generative rather than merely classificatory. KFD-4 declares the observation horizon and allows perspective transformation to move it. KFD-5 keeps candidate generation open to multiple methods and makes every candidate testable. KFD-6 attempts to autonomize both generation and comparison. KFD-7 closes the first return to action: the resulting world model still cannot synthesize intent, perspective, permission, or occurrence from one another.

KFD-8 through KFD-10 specify the three cross-domain action coordinates without choosing a business workflow. KFD-11 then separates Claim, Assessment, authorized Decision, and Admission at the trust-to-action boundary. KFD-12 and KFD-13 provide the first numbered domain application: durable software work and project-level settlement. That application is evidence and a reusable profile, not a command that other domains adopt Initiative, Assignment, Project Cut, or the same lifecycle. Other domains may inhabit the geometry with different vocabulary, fields, transitions, and macro commitments.

The first software application also demonstrates the full KFD-4/KFD-5 transition. Staying in the lower-object developer view suggests refining Fact, Episode, Pursuit, Atlas, and Warrant. Moving to the real participant’s view instead exposes Initiative and Assignment as the objects needed to organize work. Staying in the coordinator view then suggests maintaining those objects more precisely; moving to the successor Agent’s view exposes Project Cut as the project boundary needed for continuation. Perspective transformation generates these candidates, but does not certify them. Separate KFD-5 cuts qualify their distinct responsibilities, alternatives, falsifiers, evidence, and bounded acceptance. The live case preserves that separation.

Observer-bound timeline and replay are the first concrete KFD-4 profiles, just as release versioning is one concrete KFD-1 profile. Episode and Atlas are KFD-5 cases: each became natural when the view moved from an implementation object to the participant who had to understand and act. That history supports perspective transformation as a candidate method, not its dominance. A future KFD-6 system must test such participant views alongside anomaly, reconstruction, causal-variable, compression, and hybrid methods over causal Episode experience.

Continue with Primitive Discovery in History to test this model against external cases. Continue with the numbered KFD texts for the authoritative principles and procedures.