Dependency inventory: kb/articles/automated-software-houses-with-fixed-llms.md

Source commit: 50121b7178058782f7b264127bbd23d94f7eeff5

# link text target kind target title target description carrying sentence binding
1 the repository's GitHub Discussions page https://github.com/zby/commonplace/discussions external Comments and counterexamples are welcome > on the repository's GitHub Discussions page. onward
2 companion article kb/articles/the-software-house-as-the-unit-of-training.md article The Software House as the Unit of Training Doctrine that once an automated software house carries the production loop, project-specific continual learning runs over the whole house with the model as one update surface; training steps, write governance, explicit-… It does not argue that a fixed model is the best way to train such a house; a companion article takes that up. onward
3 software house kb/notes/definitions/software-house.md definition-note Software house Definition — a software house is the complete persistent system responsible for developing and evolving software for external users The automated software house conjecture. At least one automated software house (below, the house) capable of open-ended coherent software change is practically reachable with LLMs available by 2026-09-02. A
4 distributed-parametric state kb/notes/definitions/representational-form.md definition-note Representational form Definition - representational form classifies how an operative part is encoded and consumed: natural-language, symbolic, distributed-parametric, or mixed Together these are the distributed-parametric state. A
5 notes and code kb/notes/definitions/representational-form.md definition-note Representational form Definition - representational form classifies how an operative part is encoded and consumed: natural-language, symbolic, distributed-parametric, or mixed What gets trained is the house, through changes that computation produces and the house retains in its notes and code. A
6 definition kb/notes/definitions/software-house.md#scope definition-note Software house Definition — a software house is the complete persistent system responsible for developing and evolving software for external users An internal role is what the definition calls an internal production role: work the house depends on to produce the software, whoever performs it. A
7 Holding a program theory means sustaining coherent search under delayed feedback kb/notes/program-theory-sustains-search-under-delayed-feedback.md note Holding a program theory means sustaining coherent search under delayed feedback Holding a program's theory is tested by whether a partial, fallible account of what the program is for keeps modification search, backtracking, and recovery coherent until delayed evidence arrives, not by whether the fi… Holding a program theory means sustaining coherent search under delayed feedback: the multi-tenant choice may not show its consequences until the next three demands arrive. B
8 Naur's compiler case kb/sources/programming-as-theory-building.ingest.md source Ingest: Programming as Theory Building Naur's theory-building view makes maintainability depend on situated design understanding while bounding what retained rationale alone can transfer Naur's compiler case showed that full code, annotations, extensive design discussion, and personal advice did not give a successor team enough program-specific understanding. R
9 one set of documents and one way of reading them, bounded to its time kb/notes/naurs-compiler-case-tests-one-historically-bounded-documentation-and-consumption-system.md note Naur's compiler case tests one historically bounded documentation-and-consumption system Naur's compiler transfer failure rules out more documentation of the same kind, but tested one historically bounded package and consumption process rather than every possible rationale, indexing, retrieval, and activati… But it tested one set of documents and one way of reading them, bounded to its time. B
10 The distinction is between formal execution and explicitly formulated criteria kb/notes/naur-equates-machine-execution-with-formulated-criteria.md note Naur binds program theory to humans by equating machine execution with formulated criteria Naur argues program theory cannot be expressed as criteria, then concludes it is human-only; the bridge equates machine execution with formulated criteria, true for the programs of his day and separated since by trained… The distinction is between formal execution and explicitly formulated criteria. B
11 Governing a behaviour-changing write therefore requires selection, validation, authorization, and coordination kb/notes/continual-learning-requires-governing-behaviour-changing-writes.md note Continual learning requires governing behaviour-changing writes, not just storing content For deployed systems, persistence is insufficient; continual learning must select, validate, authorize, and coordinate behaviour-changing updates across the representational forms a system can change Governing a behaviour-changing write therefore requires selection, validation, authorization, and coordination. B
12 transition closure of the seed kb/articles/reachability-as-closure-under-the-seed-gate.md article Reachability as transition closure under the seed's successor relation Supplement to the reachability conjecture: an autonomous house's states form the transition closure of its seed under a state-dependent successor relation and declared input process; the Gödel-machine comparison differs… The companion article names this set the transition closure of the seed. A
13 Gödel machine kb/notes/goedel-machines-are-a-proof-governed-case-of-self-modification.md note Gödel machines are a proof-governed case of reflective self-modification The Gödel machine realizes reflective self-modification with a proof-gated acceptance rule, gaining model-relative rigor at the cost of excluding useful changes it cannot prove The fully formal case is AI researcher Jürgen Schmidhuber's Gödel machine, a proof-governed construction that rewrites its own code. A
14 Schmidhuber kb/sources/goedel-machines-schmidhuber.ingest.md source Ingest: Gödel Machines — Provably Optimal Self-Improvements Schmidhuber's Gödel machine — rewrite of any part of one's own code, proof searcher included, gated on a proof of higher axiomatized utility: the proof-governed case of reflective self-modification The paper states the price: the machine "must ignore those self-improvements whose effectiveness it cannot prove" (Schmidhuber, §2.4, verbatim). R
15 third article kb/articles/bootstrapping-the-first-automated-software-house.md article Bootstrapping the First Automated Software House Research program for reaching the first automated software house from a human-agent house: move internal roles out of human hands one class at a time as their premises, criteria, and checks are built; evidence of a move… It has no witness run and is scored as a design target, not as evidence; the third article in this series sets out the program for getting from a house of that kind to a witness. onward
16 companion map kb/articles/nearest-existing-constructions-to-a-reachability-witness.md article Nearest existing constructions to a reachability witness Comparison of reviewed self-improving systems against the reachability conjecture's four witness obligations, which do not fix which form carries the theory, with a separate stronger protocol for explicit retained theory The companion map compares twenty constructions and records the evidence behind each placement. onward
17 companion article kb/articles/the-software-house-as-the-unit-of-training.md#testable-consequences article The Software House as the Unit of Training Doctrine that once an automated software house carries the production loop, project-specific continual learning runs over the whole house with the model as one update surface; training steps, write governance, explicit-… The companion article states it as a hypothesis with its own test. onward
18 open-domain theory builder may itself become a software house kb/notes/an-open-domain-theory-builder-becomes-a-software-house-when-new-domains-require-production-machinery-changes.md note An open-domain theory builder becomes a software house when new domains require production-machinery changes A persistent automated theory builder for external users becomes a software house when genuinely new domains require it to revise the software that performs theory production rather than only the theories produced An open-domain theory builder may itself become a software house when new domains require changes to its production machinery. onward
19 companion article kb/articles/the-software-house-as-the-unit-of-training.md article The Software House as the Unit of Training Doctrine that once an automated software house carries the production loop, project-specific continual learning runs over the whole house with the model as one update surface; training steps, write governance, explicit-… It does not argue that a fixed model is the best way to train such a house; a companion article takes that up. onward

Terms introduced in emphasis

term carrying sentence binding
Open-ended software development with fixed LLMs Open-ended software development with fixed LLMs A
software house TL;DR. We use software house for whatever keeps changing a piece of software for its users, not only a company. body
The automated software house conjecture The automated software house conjecture. At least one automated software house (below, the house) capable of open-ended coherent software change is practically reachable with LLMs available by 2026-09-02. body
the house The automated software house conjecture. At least one automated software house (below, the house) capable of open-ended coherent software change is practically reachable with LLMs available by 2026-09-02. body
Practically reachable Practically reachable means success within a declared product scope, operating horizon (how long the house must keep running), and budget of compute, time, and cost. body
internal role An internal role is what the definition calls an internal production role: work the house depends on to produce the software, whoever performs it. body
witness A witness is one working house that makes the existential claim true, in the logician's sense of the word. body
Holding and application Holding and application. Given adequate project-specific state, the house sustains theory-guided proposal, evaluation, diagnosis, or recovery across novel changes, including cases whose correct handling is not stated verbatim in that state. body
Initial acquisition Initial acquisition. From permitted records, interaction, and participation in the work, the house acquires the capacity an adequate program theory provides instead of receiving the decisive project-specific understanding from a human. body
Successor acquisition Successor acquisition. When experience exposes an inadequacy, the house reaches a successor state that supports coherent later modification. body
Automated continuation Automated continuation. The house sustains these capacities across the declared scope and horizon with no human in an internal role. body
term definition note first sentence binding

No unlinked KB definition terms were detected.

These dependencies were not rows in the generated inventory. The staging procedure requires them to remain free until the operator supplies the missing definition or support; staging does not narrow or rewrite the body to remove them.

# kind term or claim carrying sentence binding
F1 term coherent software change / coherent later work At least one automated software house capable of open-ended coherent software change is practically reachable with LLMs available by 2026-09-02. free
F2 term available, as the model-eligibility criterion The eligible model versions, meaning those available by the cutoff date, stay fixed, and so do their weights and any other learned component, such as an embedding model. free
F3 term learned component The eligible model versions, meaning those available by the cutoff date, stay fixed, and so do their weights and any other learned component, such as an embedding model. free
F4 term relevant novelty / relevant demands Open-ended means that the process generating requests and their consequences is declared in advance and can produce relevant novelty; it is not a fixed list of benchmark tasks. free
F5 term adequate, as applied to state, theory, capacity, machinery, and successors Outgrowing the seed means that adequate human-free states are reached with usable probability and sustain coherent later work, not merely that one such state lies on a possible path. free
F6 term frontier model And it matches the regime every project built on a frontier model is already in. free
F7 term reviewed system / reviewed construction, including the boundary of the reviewed set No reviewed system is an empirical witness. free
F8 claim Agent harnesses with memory files and self-written rules are already moving this way. Agent harnesses with memory files and self-written rules are already moving this way. free
F9 claim Architecture decision records already carry part of this rationale. Architecture decision records already carry part of this rationale. free
F10 claim An LLM can interpret design rationale without a complete explicit rule. An LLM is still formal computation, but it can interpret a paragraph explaining why retry logic belongs in the caller without that judgment first being written out as a complete, explicit rule. free
F11 claim Fixed current LLMs supply the stated general linguistic, programming, and reasoning capacity. Fixed current LLMs supply the general linguistic, programming, and reasoning capacity that interprets project state and produces candidate changes. free
F12 claim Every project built on a frontier model is already in the fixed-model regime. And it matches the regime every project built on a frontier model is already in. free
F13 claim A project built on a frontier model cannot retrain that model and must retain project learning elsewhere. Such a project cannot retrain the model it uses, so whatever it learns about its own product has to be kept somewhere else. free
F14 claim Current LLMs can produce both notes and code. Current LLMs can produce both notes and code. free
F15 claim No reviewed system is an empirical witness. No reviewed system is an empirical witness. free
F16 claim Fluent and the OpenAI agent-first account come closest, but people retain the named internal roles. Factories with people in internal roles, such as Fluent and the OpenAI agent-first account, come closest. free
F17 claim Darwin and Huxley Gödel Machines and HyperAgents train software around frozen foundation models; Dynamic Cheatsheet and Voyager retain natural-language or executable artifacts. Darwin and Huxley Gödel Machines and HyperAgents train software around frozen foundation models; Dynamic Cheatsheet and Voyager retain natural-language or executable artifacts. free
F18 claim None of the named systems combines user-facing open-ended operation, program-theory acquisition and revision, and human-free continuation. None combines user-facing open-ended operation, acquisition and revision of program theory, and continuation without a human in an internal role. free
F19 claim The named harness practices instantiate the practice the article formalizes. The harnesses among them are the practice the article formalizes: memory files, self-written tests and tools, retained rules from past failures. free
F20 claim The named accounts do not identify whose learning produced an improvement. What their accounts do not say is whose learning an improvement was. free
F21 claim The people described performed credit assignment, and the accounts do not separate their contribution from the harness's. The person who noticed the recurring failure, diagnosed it, and chose the rule did the credit assignment, and the accounts do not separate that from what the harness did on its own. free
F22 claim Commonplace is the only reviewed construction built to attack holding and acquisition directly through retained explanatory theory. Commonplace, the knowledge base this article is written in, is the one reviewed construction built to attack holding and acquisition directly, through retained explanatory theory that later work loads and revises. free
F23 claim Commonplace has no witness run and is scored as a design target rather than evidence. It has no witness run and is scored as a design target, not as evidence; the third article in this series sets out the program for getting from a house of that kind to a witness. free