Learning that improves capability without granting itself authority.
H33-Tutu is a governed learning, tuning and evaluation capability that is designed to improve model capability and applies the same evidence discipline to governance adaptation, without authority to authorize, decrypt, promote its own results or modify its own trust boundaries.
Inside H33, H33-Tutu is called TUTU. It is being developed as a self-learning foundation model whose learning targets reasoning, mathematics, coding and security.
L100 is the complete AI system H33 is developing. Within it, TUTU is designed to improve the system’s model capabilities through governed self-learning — it is not an inference-optimization feature; Champion TUTU evaluates which learning opportunities are worth pursuing; and H33-HIVE orchestrates execution — it is not the model. TUTU and Champion TUTU are parts of L100, not separately sold products.
In every domain TUTU works in, it reads what actually happened, learns, and proposes. It does not decide. A learned improvement becomes eligible only through independent qualification, and nothing TUTU produces changes what HIVE may use until it has.
Governed learning, tuning and evaluation
TUTU trains candidate improvements and evaluates each candidate against sealed held-out sets it never trains on. A candidate that regresses is refused, and a stronger result in one area cannot outvote a regression in another. Retention of existing capability is designed to be measured against declared floors, not assumed; retention evaluation sets are being prepared, and no retention result has been established. Specialist areas — mathematics, code, security, science and agentic work — are planned; no specialist is active or qualified.
Historical experiments showed within-family learning gains; independently verified transfer to unfamiliar problem families remains unproven. A passing author-side test is never independent qualification.
Champion TUTU: what is worth learning
Champion TUTU coordinates learning across the specialist tranches. It is designed to identify capability gaps from verified evidence — including HIVE’s execution evidence — allocate learning effort to the experiments with the best expected gain, and track retention and capability over time.
Its experiment selection and learning allocation are implemented in research and run in shadow mode only. It recommends; it does not decide. Its recommendations grant no authority and never promote weights automatically, and there is no autonomous production learning. Background learning stays subordinate to foreground work: it must not take resources a live request needs.
Governance adaptation
A governance model that never changes eventually governs a system that no longer exists. One that changes itself is not a governance model at all. In the governance domain, TUTU occupies the narrow space between those failures: given replay-verified evidence, it proposes bounded governance improvements. It does not decide.
Proposing is not authorizing
The separation is the point. TUTU does not learn authority; in this domain it learns what bounded constraint additions might reduce future risk, given governed evidence. H33-Root decides what becomes reality, and H33-Stamption records the committed transition.
| Component | Role |
|---|---|
| H33-Tutu | proposes bounded governance transitions |
| H33-Root | owns identity, semantics, stability boundaries and commit authority |
| H33-Stamption | records the committed transition |
The adaptive governance learning invariant
One rule defines governance adaptation:
No governance mutation may originate from evidence that has not independently resolved to a governed control attribution.
And one distinction gives it force. A failing basis does not yield denied — it yields InsufficientLearningEvidence. The system has not proved the proposal wrong; it has proved it inadmissible. That line runs through the whole estate: a crash is not a failed control, absence of a ruling is not a ruling, and an unevidenced proposal is not a refuted one.
The attack class this closes is the one that does not require compromising Tutu at all. Poison the learning input and an adaptive system becomes a self-reinforcing error machine. With the invariant in place, Tutu learns only from measured, attributed, replayed failures.
What composes it
For governance adaptation, H33-Tutu composes Agent-008, H33-74, H33-Root and H33-Stamption. For capability learning it is built in L100 with H33-HIVE, which supplies execution evidence and uses only independently qualified results. Composition is an architectural relationship, not a readiness claim: composing evidenced parts is not itself evidence, and a capability does not inherit readiness from what it is built on.
Where TUTU stands today
| Behaviour | State |
|---|---|
| Learning and tuning pipeline with sealed held-out evaluation and regression refusal | Implemented and tested in research; of its scientific controls, only the contamination gate is independently qualified |
| Specialist areas (mathematics, code, security, science, agentic) | Planned; no specialist is active or qualified |
| Trained candidates | Shadow-only: none has been promoted, and none serves any task |
| Independent qualification and promotion of a successor | Not yet established |
| Champion TUTU experiment selection and learning allocation | Implemented in research; shadow mode only; independent qualification not passed |
| Governance-adaptation proposals and the learning invariant | Specified; no published implementation evidence |
| Closed loop with H33-HIVE execution evidence | Not yet qualified |
H33-Tutu is an admitted H33 identity with no published evidence surface. What does not exist in public evidence is a demo, a proof lab, a whitepaper, or a deployed instance.
This page therefore makes no claim about accuracy, learning quality, interfaces or throughput. For governance adaptation, one replay-verified adaptive transition — a learning-evidence record whose control attribution resolved, whose replay hash verifies, and whose verifier identity is established — is what moves it from stated architecture to published evidence. For capability learning, a successor that passes independent qualification is.
H33 states readiness explicitly rather than by omission. The distance between an identity that exists and a capability that has been evidenced is recorded, not hidden.