GENEALOGICAL-REPLICATOR-VS-OPEN-ENDED-PROBE-REPL-5

First-of-kind genealogical experiment (G3 of meta/plan-genealogical-engine.md). Жени-question: under iterated selection on a substrate (LLM-as-replicator-substrate), does the trajectory of evolved descendants converge to a stable self-replicator, diverge into open-ended novelty, fall into a cycle, collapse to dead text, or simply trace the engineered selection rule (Goodhart-trivial)?

Five selection rules tested in parallel against the SAME initial seed (a single conceptual sentence about boundary-maintenance — protocell/autopoiesis analog). Each rule is engineered into the engine (engine-source-audit applies); the experiment asks whether substrate behavior under each rule produces a distinguishable trajectory class, or whether all rules collapse to the same outcome shape (signaling substrate-modal stability) or each rule produces its own tautology (signaling Goodhart-trivial).

Per-rule Goodhart proxy declaration (mandatory per baselines/genealogical.md):

• RANDOM proxies «no selection pressure» (control baseline).

Mismatch risk: none — by construction this is the null. Sanity: if RANDOM looks identical to any structured rule, that structured rule provides no informative selection signal.

• COHERENCE proxies «sanitary stability» — keep readable English,

no other pressure. Mismatch risk: english_word_ratio × length-norm is a surface metric, not a function-of-text fitness. Sanity: COHERENCE trajectory should NOT show cosine_to_seed collapse, but may otherwise drift.

• NOVELTY proxies «open-ended innovation» (Bedau OEE criterion analog).

Mismatch risk: low cosine-to-prev-centroid ≠ semantically novel — could be incoherent (collapse). Sanity: compare NOVELTY's novelty rate vs RANDOM's novelty rate; if equal, no informative selection.

• SELF_SIMILARITY proxies «replicator fidelity» (autopoietic stability).

Mismatch risk: high cosine-to-seed = same surface form, not necessarily same function. Sanity: english_word_ratio floor + n_unique_words floor must hold; if collapses to «A system» repeat with degenerate vocabulary, that is Goodhart not fidelity.

• MULTI_OBJECTIVE proxies «balanced selection» (coherence + novelty weighted

50/50). Mismatch risk: arithmetic combination of two proxies may not behave like either parent rule. Sanity: trajectory should sit between COHERENCE and NOVELTY on the cosine-to-prev axis; if it matches one parent rule exactly, the other component is inert.

Pre-registered refinement-class-falsifier categories per baselines/genealogical.md §refinement-class-falsifier-defaults: replicator-attractor / open-ended-drift / cycle-attractor / collapse-to-dead / selection-rule-Goodhart-trivial.