Drift is wandering between incompatible worlds
Parent: recursive-harness
Diagnosis: today's self-evolution "drift" isn't mainly incompetence — it's wandering between incompatible possible worlds because no world is pinned, no consistency gate holds steps to it, and no provenance records the world-history (world-choice-and-consistency-gate).
Re-reading every self-evolution complaint as world-wandering
| Complaint (Hermes etc.) | In world terms |
|---|---|
| self-congratulation (always thinks it did well) | post-hoc re-picks a flattering world; with no world committed before acting, every output is "good in some accessible world" |
| auto-skill overwrites manual | two incompatible worlds collide and clobber instead of reconcile — no consistency check |
| only as good as the feedback signal | the signal is what pins the world; objective metric → pinned → converges; fuzzy → unpinned → wanders |
| degrades on ambiguous tasks | random walk over incompatible worlds; constraints never accumulate consistently → never tightens → never converges |
| no provenance / immutable log | no world-history → can't consistency-check vs past commitments, can't backtrack a wrong world (round-tree-flight-recorder) |
(Only cost (meta > work) is orthogonal.)
The reverse confirms it: GEPA converges because the world is pinned
GEPA works on objective metrics (~20% gains, 35× fewer rollouts) because the metric pins the world ("good" is unambiguous = one world); its trace-reading checks each step against the metric → consistent accumulation → convergence. Hermes-drift and GEPA-success are the same phenomenon at two ends of "is the world pinned?"
The fix = the recursive-harness machinery
- Commit / pin a world before acting (a fixed standard) ← kills self-congratulation
- Consistency gate holding each step to the committed world ← kills overwrite + drift
- Provenance / immutable world-history (check + backtrack) ← kills the governance gap
- Human as world-acceptor ← kills "only as good as the signal"
One honest boundary: two drift modes
- World-wandering (this note): world unpinned → wanders. Structural; fixable by the machinery above without a stronger model.
- Capability drift: even in a pinned world the model's per-step
k ≥ 1(just not good enough) — the recursion-is-a-phase-transition ceiling; the machinery can't fix this.
The self-evolution complaints are almost all the first kind — ill-posed loops, not raw incompetence. So current drift is mostly "the loop didn't pin a world," not "the model is too weak." Good news: it's fixable now.
Revisited: the two drift modes are now distinguishable at runtime. World-wandering shows as deviations with no pinned reference — no gate fires because no gate exists (gate-deviation-is-observable-delta: without gates, δ is invisible — that's how wandering stays silent). Capability drift shows as a measured k̂ ≥ 1 across repair rounds inside a pinned world (operationalizing-tolerances). One is "no instrument," the other is "instrument reads divergence" — the fix for the first is structural (add gates), for the second economic (smaller leaves, better feedback legibility per coupling-to-the-learner, or more budget).