Six schema-based experiments probing the substrate's boundary. Each ran with calibrated probability distributions. Findings feed into /invitation/questions.md and /theory/paper-psyche-4model.md.
exp-01: Is there a missing 12th opcode?
JEV · noul · choice · score · 2026-09-19
| Question | Result |
|---|
| needs_twelfth_opcode | 41% yes (low confidence — JEV split) |
| missing_opcode_role | uncertain (37% conf) |
| confidence_in_missing | 1.58 / 5.0 |
The substrate hints at something unnamed but JEV can't pin it down. 5 candidates: synthesis / memory / boundary / time-extended / "no missing opcode."
exp-02: What breaks at scale?
JEV · scale probe · 2026-09-19
| Question | Result |
|---|
| scales_to_1M | 19% yes (JEV very clear: doesn't scale naively) |
| scale_breakpoint | 0.43 → between 100 and 1K cells |
| first_to_break | all_broken_at_once (21% conf — bare plurality) |
Substrate fails somewhere between 100 and 1K cells without architectural change. JEV doesn't single out one cause — holistic failure mode.
exp-03: Should FORGET always be JEV-decided?
JEV · memory strategy · 2026-09-19
| Question | Result |
|---|
| always_jev_forget | 41% yes |
| forget_strategy | tiered (89% conf) — algorithmic + JEV override |
| no_forget_cost | 2.87 / 5.0 (significant degradation) |
JEV strongly favors tiered: most FORGETs cheap and algorithmic; load-bearing FORGETs need JEV. Without ANY forgetting, the substrate degrades over 10 years.
exp-04: Cell disagreement protocol
JEV · fork/scar question · 2026-09-19
| Question | Result |
|---|
| should_fork | 36% yes (low — substrate leans against) |
| fork_strategy | never_fork (55% conf) |
| fork_frequency | 3.69 → ~0.04% (rare if at all) |
JEV prefers never_fork — the loser cell records as scar, continues. Tracks with "you can't patch over a scar" principle. Substrate doesn't fork; scars persist.
exp-05: Is "muscle memory" the right name for embeddings?
JEV · naming probe · 2026-09-19
| Question | Result |
|---|
| muscle_memory_is_right | 23% yes (JEV prefers another framing) |
| alt_framing | trail (73% conf) |
| importance | 2.28 / 5.0 (mild — naming matters but not load-bearing) |
JEV prefers "trail" over "muscle memory" — a trail is where something has been, more honest about retrieval-as-history. Casey hasn't weighed in.
exp-06: Should /lab/ continue?
JEV · meta probe · 2026-09-19
| Question | Result |
|---|
| lab_should_continue | 49% yes (50/50 split) |
| lab_phase | interim (56% conf) |
| convergence_proximity | 0.72 (high — JEV thinks we're close) |
/lab/ is probably interim — merge with /theory/ once we converge on shape. JEV thinks we're 72% of the way there.