file
Control logs
The matched-arm logs behind the attractor result, with per-arm entropy, lock counts, and move counts. structural-attractor-control.md →
Replication
How to attack these results.
A result you cannot challenge is not science. This page is about how someone else can reproduce or refute what is claimed here. Full external replication is not possible yet: the runtime is not open source. It is going public on GitHub in the coming weeks, and this page will become the entry point. Until then, here is what is already published, and where each result is weakest.
The artifacts are published before the conclusions are settled.
Not: published artifacts are not the full pipeline. Without the code and model identity, these support re-analysis and challenge, not yet a clean-room rebuild.
Published now
What you can inspect today.
Every result links to its own record. These are the sanitized artifacts behind the claims.
file
Raw data
The machine-readable record for the same run, for re-analysis without our framing. structural-attractor-control.json →
file
Determinism canary
The create-versus-restore comparison, seed and temperature fixed, that gates every cycle. cache-restore-determinism-canary.md →
file
Within-life trace
The full belief-to-death chain for Life 43, tick by tick. belief-chain-L43-trace.md →
file
Signal log
The 2026-05-29 self-maintenance reading, withdrawn on review (a scorer clock-ordering artifact; a control indistinguishable from a repeated-probe latch). Closure scores 0. Kept with its correction, for the record. self-maintenance-signal.md →
Failure modes
How to break each finding.
Each result has a stated attack. If one of these succeeds, the corresponding status should fall.
attack
The withdrawn self-maintenance signal
There is no live signal to attack: the 2026-05-29 closed-loop reading was withdrawn on review (a scorer clock-ordering artifact; the affordance-disabled control matched a trivial repeated-probe latch), and closure scores 0. Re-establishing it would need a scorer with correct temporal semantics and the still-unrun no-decay baseline at the same seed.
attack
Break the canary
Restore a saved cache in a fresh process at a different seed, or on different hardware, and check whether the decoded tokens still match. A single divergence falsifies the determinism claim.
attack
Enlarge the attractor test
Re-run the three arms at larger n. The lock effect is reported at n=3; if it does not hold at n=20, the support weakens.
attack
Attack the trace
Argue the Life 43 belief was merely co-present, not load-bearing, in the dying decision. The logged decision window is there to be re-read.
Still missing
What full replication needs.
- Code
- The runtime is not yet public. Open source: coming weeks
- Model
- The exact open-weights model and its version, pinned, so a rebuild starts from the same cognition.
- Hardware
- Results run on Apple MLX; the determinism guarantees are stated for that path. Cross-platform behaviour is unverified.
- Harness
- The self-maintenance scorer was found to have a clock-ordering artifact; the 2026-05-29 closed-loop reading was withdrawn and closure scores 0. The smoke test is currently blocked by an engineering launch bug, not a scientific negative. A corrected scorer ships only once it is trusted.
When the code is public, this page becomes a set of commands, not a promise.
Not: openness is not yet here. Until the repository is up, treat every result as ours to defend and yours to doubt.
The evidence behind each claim → The findings and their statuses →