Question, hypotheses and gates.

HYPOTHESISThe object is a governance architecture, not the cognition of a model. A proposal recorded on 23 July 2026, still classified as a hypothesis in the archive, formulates: the Symbiontes architecture is a governance architecture, made of roles, gates and receipts. What is observed is the relationship: who authorises, who executes, what returns as evidence and what remains as memory.

RECORDED FACTThe central question, in the terms of the canonical thesis of September 2026: can a frugal architecture, with several AI vendors, produce repeated improvement in at least two operational dimensions of the host, without degrading the safeguards, and can it be withdrawn leaving the host better, documented and operational?

The canonical is the formulation of the thesis, not its proof.

Reference thesis, opening note (6 Sep 2026)

Four planes

RECORDED FACTThe four planes replaced, on 2 August 2026, an earlier two-level formulation (core × host instance), judged insufficient in the adversarial review between two AIs from different vendors. In the same review, “impossibility of dissociation” was replaced by “deep co-evolution with sovereign exit”.

Two versioning axes

Deep co-evolution with sovereign exit. Symbiontes must be able to withdraw leaving the host better, documented and operational.

Principle approved on 2 Aug 2026; kept in the canonical thesis

RECORDED FACTThe principle was born from a correction. The founder's original intuition was that symbiosis generated irreversible coupling, an “impossibility of dissociation”. The adversarial review reframed it: deep coupling can generate value, but irreversible dependence generates lock-in, reputational risk and endless artisanal service; the desirable symbiosis is continuity by value, not technical impossibility of separation.

RECORDED FACTA circularity was corrected between revisions 4 and 5 of the thesis: G0 became the baseline of the candidate kernel, and G3 came to require transfer plus a privacy gate before any promotion to validated. There is no shortcut: a pilot proves only its own slice.

1. H1 · reset plus a sanitised context package produces less leakage and drift than conventional context compaction.
2. H2 · a governed adaptive context package outperforms a fixed summary (the internal term was replaced by editorial policy).
3. H3 · control external to the prompt reduces undue instruction mutations.
4. H4 · replay restores internal state but does not undo external effects.
5. H5 · consensus between distinct vendors does not reduce error when failures are correlated.
6. H6 · the proposed protocol is only useful if it beats close alternatives (reversible contextual elasticity, capability flow control, hierarchical agent memory and published observability protocols).

HYPOTHESISNone of the six had been tried by the cut. The verification scale adopted has five orthogonal axes (access, independence, editorial status, relevance and adversarial review) and one hard rule: a claim only leaves the hypothesis state with full access to the text, verified source independence and at least one adversarial round; multiple addresses for the same paper never raise independence.

RECORDED FACTTwelve propositions of the programme were confronted with the literature in August 2026 and sorted into six boxes: KNOWN · ADAPTATION · SYNTHESIS · HYPOTHESIS · CONTRADICTION · POTENTIAL CONTRIBUTION. The result downgraded much of what had looked original.

P10 (elastoplasticity) downgraded to HYPOTHESIS — a term that does not exist in organisational theory, but whose semantic space is occupied by resilience/ambidexterity/antifragility; high risk of redundancy.

Prior-art review, round 1 (2 Aug 2026)

RECORDED FACTThe use axis was added on 28 September 2026 after a diagnosis by the founder: a great deal had stayed in draft “awaiting an acceptance” and never became valid or canonical, forming a graveyard of drafts. The inventory made the same day found 1,609 candidate or draft items; 34 documents were promoted to validated by human act on 28 Sep, with three left for a later stamp.