Adaptive Knowledge Production Architecture (AKPA) is an open working paper and executable blueprint, released under CC BY 4.0 with MIT-licensed code. It argues that most AI-assisted knowledge failures are created early and distributed across many small decisions that final review cannot catch. Its proposal is deliberately narrow: govern the move from initial request to committed working frame, and the move from an observed artifact defect to a change in the production method. It is a design proposition and research agenda — not a validated standard.
What AKPA proposes
Adaptive Knowledge Production Architecture (AKPA) is a risk-proportional framework for producing consequential knowledge artifacts — a thesis, a market report, an investment memorandum, a contract, a policy brief, a technical analysis. It applies either to one important artifact or to a recurring family of comparable ones.
The paper is published openly, with the full manuscript, a source registry, a claim ledger, an argument map, and a runnable market-intelligence blueprint that can be validated with the Python standard library.
Read the paper
The working paper, blueprint, and research documentation are public at https://github.com/saintaigo/adaptive-knowledge-production-architecture
What makes it worth reading is not novelty of primitives. Framing, provenance, review, versioning, and change control are all familiar. The proposal is the relation between them, applied to one artifact or one defined family, with two moments made inspectable.
Two transitions, governed on purpose
Ordinary prompting plus a final review leaves two moments implicit. AKPA makes both explicit and nothing else mandatory.
- From request to frame. An initial request is evidence about what someone wants, not a correct description of the problem. It becomes a committed working frame only after a challenge proportionate to ambiguity and consequence. The frame states purpose, definitions, material assumptions, exclusions, evidence requirements, decision criteria, and who may reopen them.
- From defect to method change. A defect in an artifact is repaired in the artifact by default. Only a critical or recurring signal justifies changing the production method — and then only as a diagnosed, reversible process hypothesis.
The first constraint limits premature commitment. The second limits process accumulation — the slow drift where every mistake produces another rule until the method becomes bureaucracy.
Why final review is the wrong place to catch these failures
A knowledge artifact can fail while every paragraph looks competent. The initial question defines the wrong market. Two analysts use different denominators and report apparently compatible estimates. A literature review repeats a claim whose source supports only a narrower statement. A technical report cites the right specification at the wrong revision.
None of these are writing problems. They arise from how purpose, evidence, definitions, calculations, decisions, tools, and reviewers were arranged during production. Final review is being asked to detect failures created much earlier and spread across many intermediate decisions.
Generative AI raises both output volume and the gap between fluency and evaluability. A polished argument reads as sound whether or not its key authority is current.
What the framework assumes
The paper is unusually direct about its own premises. These are the assumptions a reader should test before adopting anything from it.
| Assumption | What it means in practice |
|---|---|
| Process should be proportional to risk | Consequence, evaluability, complexity, and recurrence decide how much architecture is warranted. For low-stakes, easily checked, disposable, or primarily aesthetic work, AKPA is excessive and a template or direct expert review is better. |
| Failures originate upstream | Most material defects are created during framing, evidence selection, and definition, not at the end. Controls placed only at review arrive too late. |
| Fluency is not assurance | Output quality cannot be inferred from readability, and explanation does not automatically make a human–AI pair perform better. |
| The initial request is contestable | The person asking may be describing the wrong problem. Challenge is proportional — permanent reframing prevents completion. |
| Delegated output is a candidate, not truth | Model or agent output returns with its question, sources, exclusions, assumptions, uncertainty, and checks. It cannot promote itself into the committed working state. Agent count is not evaluator independence when agents share a model, prompt, sources, or incentives. |
| Process change is a hypothesis | Method changes are diagnosed, reversible, and attributable — because a later improvement may come from an easier task, a different executor, or a better model, not from the change. |
| Human accountability does not transfer | The architecture can expose sources, assumptions, calculations, disagreement, confidence, and reasons for abstention. It cannot substitute for legal, investment, scientific, engineering, or medical competence. |
| Recurrence creates path dependence | Stable definitions improve comparison but can hide change; templates can standardize a mistake; persistent memory can replay prior errors. Every recurring harness needs an owner, a version, a challenge cadence, and retirement criteria. |
Stated status
AKPA is a design proposition and research agenda — not a validated standard, a certification method, or a demonstrated superior process. The combined architecture remains untested.
The objections the paper concedes
Three, taken from the paper rather than added to it.
- It may rename good practice. Partly true. The name earns itself only if the relation among framing, authorized working state, checked production, and reversible change improves control relative to a simpler method. Reduced forms should be compared, not only the fully instrumented version.
- It may become bureaucracy. Scaffolds turn into forms and traces into proof theater. The answer is a four-relation minimum, measured maintenance cost, and a live option not to build it at all.
- False assurance is worse than visible absence of control. A complete trace of a poor inference is still a poor inference, and shared sources, models, and incentives can reproduce the same error in production and in review.
Who should read it
The framework is most plausible when errors matter, direct evaluation is difficult, evidence or calculations are complex, several contributors depend on shared definitions, or outputs must stay comparable over time.
- Leaders standardizing recurring analytical output — market intelligence, investment memoranda, due diligence, regulatory or quality reporting
- Teams where AI-assisted drafting has outrun the organization's ability to verify it
- Anyone building an internal knowledge or agent workflow who needs a defensible acceptance boundary
- Researchers looking for a framework that names its own failure modes and invites counterexamples
The repository explicitly welcomes empirical tests, counterexamples, simpler comparators, and evidence about setup and maintenance cost.
[ Author ]
Santiago Rueda
Santiago Rueda leads AI, operating systems, and transformation practice at Sinecta, working with organizations across South Florida, Mexico, and Colombia to turn adoption from experiment into infrastructure.