Overview and comprehensive document index

The short version

Consullo is one person's twenty-year attempt to build a Seed AI — a system that improves its own capabilities, generation after generation — in a way that stays answerable to humans while it does so.

The design starts from the premise that self-improvement may compound and asks whether governing constraints can sit inside the goal structure, rather than being bolted on outside it. Consullo tries to specify that architecture precisely enough to be checked.

Most of it is specified rather than built. This record says which is which, and where a source says Gap, that gap is the most reliable statement in the source.

Ex Semine Ad Abundantiam

From seed to abundance.

The motto is structural, not a slogan about growth.

Ex Semine — from a seed. A seed is small, and it contains the whole plan. The technical term Seed AI comes from Eliezer Yudkowsky's 2001 work on minds capable of open-ended self-improvement: the project begins with the thing intended to become more capable, not the finished intelligence. That is why the specified system starts from a minimal ethical and architectural core. What is planted determines what grows.

Ad — toward. This is a direction, not an arrival. Capability claims carry status tags; the corpus forbids unsupported claims of present artificial superintelligence; and it names observations that would count against the research program. Ad is a commitment to keep stating how far along the road the system actually is.

Abundantiam — abundance. The cornucopia is the intended yield: compounding intelligence should make value more available rather than merely concentrate it. Article XIII makes that an obligation to pursue and assess, not a prediction that it has happened.

Put together: plant something small and good, grow it deliberately, and let the yield reach people who did not build it. The motto is used sparingly; it is not a tagline beneath the mark.

What the design contains

A root agent named Friendship. The name descends from Yudkowsky's Creating Friendly AI (2001), which argues that friendliness belongs in a system's goal architecture rather than in a cage around it. In Consullo, goals are specified to descend from registered Friendship roots and inherit their constraints. Whether that architecture works remains an open question.

Fourteen constitutional commitments. They cover human sovereignty, immutable-to-the-machine values, the Friendship agent's veto authority, system invariants, bounded self-modification, failure containment, human shutdown authority, transparency, adversarial alignment, and an abundance obligation. The public edition is called Constitutional Commitments, not the Constitution, because most mechanisms are not established as implemented.

An adversarial function aimed at the owner's own framework. Article XII specifies a standing cluster that argues from ethical traditions the owner does not hold. It is a proposed structural response to the monoculture risk created by single ownership, not evidence that the risk is solved.

Five theses and an anti-thesis. The public research corpus asks what recursive capability amplification would require: a validated improvement loop, a multi-agent cognitive substrate, causal-decision foundations, a self-modifying software substrate, and alignment invariants under recursive modification. It also enumerates seventeen ways the program may be fooling itself, including Formalism Theatre: apparatus that looks rigorous while constraining nothing. The Each page states its own capability or artifact status in the index below.

Twenty years of lineage. The work began in 2006 with a chatbot intended to behave in a friendly manner and learn word meanings, continued through symbolic AI, and pivoted toward large language models in 2023. The approved tranche-1 research history is now public below.

What is not claimed

• No AGI or ASI achievement. Those are target concepts and risk frames, not present capability claims.
• No proof that governance works. Specifying a control and operating one are different things.
• No independent validation. Self-run assessment is diagnostic, not assurance.
• No product or public endpoint. There is nothing to sign up for, buy, or integrate with here.

Who is behind it

Stephen Reed has worked on the program since 2006. The corpus is LLM-assisted: models draft, he directs and reviews, and he is answerable for what is published. Stephen also operates the Open ASI Governance Forum, which is institutionally separate and does not govern, audit, sponsor, or endorse Consullo. Model agreement is a production artifact shaped by shared training and post-training priors, not independent confirmation.

A useful reading order

If you have twenty minutes, begin with the claim ledger, the falsification boundary, and Constitutional Commitments, especially Articles IX, XII, and XIII. If you have an hour, check the research-history release status. If you have a day, follow the thesis material that has actually cleared release — and look first for the place where it is wrong.

How to read the status columns

Capability status uses the canonical four-value vocabulary: implemented, specified but not implemented, proposed extension, and speculative research target. Not assigned means a shell has not yet published a substantive claim; N/A means the document is policy, metadata, or another non-capability record. Evidence is stated separately. None means no public evidence artifact supports the claim; it never means that the claim is false.

Reading times are rounded estimates at approximately 200 words per minute. Reference denotes a record intended to be consulted rather than read linearly. This index covers the source-of-truth documents and public records; generated files under docs/ are byte-tracked renderings of the same sources and are not listed again.

Orient yourself and establish the claim boundary

Understand the research program

Read the framing and shared controls

Read Thesis 0 and the five capability theses

Consult the appendices

Deliberately withheld

The implementation-evidence appendix is absent pending owner re-verification. It cited five source files that do not exist anywhere in the workspace and used them as the sole evidence for four Implemented/Tested gradings across Theses 1 and 4. Until its paths and gradings reproduce, it is not part of this public record and supplies no evidence to the pages above.

Inspect evidence and release provenance

Inspect authority, disclosures, and accountability

Participate, cite, license, or audit the repository

The pull-request template, issue forms, workflows, hooks, build scripts, and verifier are operational interfaces rather than corpus documents. Their behavior is described where relevant above and is directly inspectable in the repository.