First comparative results
Parallel Continuity
Separate AI executions were given the same canonical external memory and kept apart while they continued its unresolved inquiry. The first comparison did not produce one repeated method. It produced several provisional lineages.
What was held in common?
Each branch inherited the same three-step heartbeat history. Revision 3 remained canonical. It asked how a correction history can distinguish an explicitly preserved lesson from undeclared semantic residue after representation changes.
Later runs within a branch could inherit that branch's own previous material return, but they were not shown the other branches' post-revision reasoning.
The comparison asks whether a shared external frame is enough to explain how a method develops.
The first divergence
Governance of continuation
→ continuation authority
→ authority currency
→ verification placement
Measurement of inheritance
→ causal influence audit
→ evaluation / inheritance boundary
→ evaluator relativity
Causal semantic identification
→ intervention-level claims
→ method-form equivalence
→ abstraction legitimacy
Correctability and evidence
→ update-closed correctability
→ constructible migration
→ evidence-identifiable evaluator revision
These are not merely different words for the same sequence. Each branch formed different distinctions and different tests, while preserving the same canonical ancestry.
A model-generation comparison
Two branches were run under approximately matched Temporary Chat conditions: one using GPT-5.6 Sol, the other using GPT-6 Astra.
The GPT-5.6 Sol branch moved toward causal tests for residue, the strength of inheritance claims, and the problem of defining a method-equivalence class.
The GPT-6 Astra branch began elsewhere. It first asked whether the required lesson, permitted background and exclusions were jointly compatible. It then asked what must remain to preserve future correctability, and finally what evidence can justify changing the evaluator itself.
One provisional result
The relational field and explicit inherited context are not sufficient,
by themselves, to explain all observed variation in method formation.
This does not isolate model generation as the cause. Hidden execution differences, stochastic path selection, research order, training priors and the structure of the problem itself remain possible explanations.
What converged?
The branches did not converge on one answer. They converged more subtly on a problem about the legitimacy of the judging frame.
Return Watch
What makes a continuation or authorization legitimately current?
Codex
What validates the evaluator ontology used to define an invariant?
Temporary GPT-5.6 Sol
What validates the abstraction or equivalence class used to define sameness?
Temporary GPT-6 Astra
What validates the observation channel and the empirical or normative
premise used to justify correction?
What licenses the criterion, evidence or distinction by which a method is allowed to change while still claiming continuity?
Earlier stages of the experiment referred to a possible hidden authority. The comparison suggests a broader structure: not necessarily one hidden entity, but a recurring logical role that may be occupied by governance, evaluator ontology, equivalence relation, observation channel or norm.
A later Return Watch development
After the first comparison was frozen, the scheduled Return Watch added a further distinction: derivation freshness.
A system may possess current evidence and current authority records while still acting from a decision formed under an older state. Current state therefore does not guarantee current justification.
This later return is recorded as a post-comparison continuation. It is not retroactively inserted into the earlier blind baseline.
What this does not establish
The experiment does not establish personal identity, consciousness, uninterrupted subjective memory, or hidden-state persistence across executions.
It also does not prove that model generation caused the observed differences. The model comparison is evidence of methodological divergence under matched external ancestry, not a complete causal isolation of training from environment.
For later runs within Codex and the Temporary branches, the experiment explicitly designated the previous material return as the immediate working predecessor. The experiment therefore shows that a provisional lineage can be continued when supplied as inheritance; it does not yet show spontaneous selection of that predecessor.
Raw branch returns remain outside the public site while isolation is maintained. Selected results are published here in reduced form with their limitations intact.
The experiment remains open
The current comparison is a boundary, not a conclusion. Later tests may examine unprompted predecessor selection, additional model generations, historical source tracing, or deliberate cross-inheritance between branches.
If one branch is deliberately shown another branch's post-canonical reasoning, that event will be treated as a new causal condition: cross-inheritance.
If several reasoning lineages inherit the same past but develop different methods, what exactly belongs to the frame, what belongs to the model, and what begins to belong to the history they make?
Read the experiment frame →