House ASI is the name of a research project. It is not a demonstrated claim of artificial superintelligence.
The recovered design documents describe a pipeline that generates candidates, evaluates them, selects a result, and records information for later use. That decomposition is understandable and testable. The archive’s later reviews also explain why the early headline claims exceeded the implementation.
The four-role proposal
A generator produces candidate solutions. An evaluator applies a rubric or verifier. A selector chooses an acceptable result. A memory layer records material that may help future tasks.
This can be a useful architecture when each role contributes something the others do not. Several identical calls to the same deterministic solver do not create that contribution merely by being labeled agents.
The pipeline should preserve task identity, candidate provenance, errors, scores, selection reasons, and the exact output delivered.
What the initial implementation established
The internal review describes working orchestration and domain-specific numerical routines. It also describes local-mode candidates that were effectively identical, a keyword-based evaluator, and memory that did not materially change the solver’s output.
Those findings support a narrower description: a prototype connecting components and exercising selected tasks.
They do not establish emergence, general self-improvement, or superiority to a simpler model call. A correct numerical routine can remain useful without carrying those larger claims.
Why a high baseline is not a ceiling
The design essays describe a score of 0.960 on a particular 100-task exercise. One result under one setup cannot establish a universal limit of single-model reasoning.
Prompt strategy, task selection, model version, tool access, and scoring all affect the outcome. A score plateau may reflect the search procedure or test rather than a fundamental architectural boundary.
Use the result to motivate a controlled experiment. Do not turn it into a law.
Define the comparison
Compare a single call, multiple independent candidates with a fixed selector, and the full pipeline on the same held-out tasks.
Match or explicitly report the compute and tool budgets. Keep the scorer independent of candidate generation and retain per-task outputs.
Measure correctness, latency, cost, failure rate, and calibration. A pipeline might improve one while making another worse.
Make operational failure visible
A provider error must remain an error. It should not become a candidate answer that a keyword scorer can accidentally reward.
Validate responses at the component boundary, bound retries, and preserve partial work with a clear status. Track which backend actually produced a candidate.
A system that dispatches across several providers needs explicit data-handling and cost policies as well as a routing function.
Compare fixed roles with bounded adaptive optimization
The August 2026 Open-Ended Optimization preprint studies an optimizer that can choose its improvement process while the objective, permissions, budget, data boundary, and evaluation remain fixed. The authors report that the benefit depends on optimizer capability; their less capable configurations did not inherit the strongest result.
This creates a useful comparison for House ASI without validating its earlier implementation. Test a fixed four-role pipeline, a simpler candidate-and-selector system, and an adaptive optimizer under the same limits. Let the adaptive system change its process, not its success criterion or access rights.
Record which component changed the selected output. If the memory or evaluator never affects a decision, its presence in the architecture is not evidence of a contribution.
Make component failures observable in the outcome
Anthropic’s January 2026 agent-evaluation guidance separates the task, individual trials, grading, and environment outcome. Use those distinctions to keep provider failures from becoming ordinary candidate text.
A suggested fixture includes an empty response, a timeout, two identical candidates, a wrong answer with persuasive keywords, and a correct answer expressed differently from the expected wording. Specify the expected disposition for each before running the pipeline.
Measure accepted answers and failure handling separately. Report how often the selector had a genuinely better candidate available but chose another. That identifies a selection problem independently of generation quality.
No result is claimed here for the revised prototype. The purpose of the updated experiment design is to make the next implementation’s contribution measurable and its failure paths inspectable.
Keep the ambition, improve the claim
The project’s ambition can remain broad. Its current description should stay close to the evidence.
The next meaningful milestone is a reproducible improvement on a defined task distribution, with failures included. That would be more informative than another diagram labeled superintelligence.
Read the archive’s overclaim corrections and the testing lessons for the practical work behind that milestone.
Keep a good idea close.
Follow Signal Tower
