# Synthena reproducibility artifacts

This directory backs three claims on the [Evidence](https://synthena.ai/evidence/)
page with replayable primary-source data. Each artifact is trimmed to exactly
what substantiates the on-page numbers; where a source is large, a regeneration
recipe is included so the full sanitised file can be rebuilt.

Cognition model throughout: a Qwen3.6 MoE (35B-A3B, 8-bit), run in-process with
a held KV cache, deterministic decode.

All three are substrate-state / engineering results. None is a claim about
consciousness, sentience, autopoiesis, open-ended evolution, or weight-level
learning. See "What this does NOT demonstrate" in each file.

## Claim → artifact map

| Evidence-page claim | Artifact | Anchored at |
|---|---|---|
| "A belief learned by the body was load-bearing at the moment of death": a single perturbation drives a ~1,332-tick chain ending in death, with the belief cited in the dying decision | `belief-chain-L43-trace.md` | Perturbation tick 30864 → belief promoted tick 30980 (confidence 0.80, active) → renders every tick → death tick 32313 (age 1733, cause energy). Strict active-belief duration = 1,332 ticks. Decision window ticks 32308–32312. |
| "A structural attractor forms in the held cache; matched controls lock, controlled discontinuity does not": 0/3 vs 3/3 final-100 locks, entropy ~1.23 vs 0.0, ~123 vs ~35 moves | `structural-attractor-control.json` + `.md` | 200-tick run, n=3/arm. Discontinuity arm: 0/3 final-100 locks, mean final-100 entropy 1.2259, mean moved 123.0, mean invalid 31.3. No-reset control: 3/3 locks, entropy 0.0, moved 35.3, invalid 1.3. SENSE-only confound control: 3/3 locks (text cue alone does not break it). Backed by the 1000-tick 3-arm differential (all 20/20 locked windows, entropy 0.0). |
| "Cache restore is deterministic: token-identical across a fresh process" (release-gate canary) | `cache-restore-determinism-canary.md` | Create-process control tokens == restore-process tokens, all 8 ids identical; `ok: true`. Temperature 0, seed fixed, 435-token prompt, ~70 MiB cache. |

## How to replay

- **L43 belief chain:** run the five read-only SELECT queries in
  `belief-chain-L43-trace.md` against the relational event store; the
  perturbation, belief, motor-event, and per-tick decision rows reproduce the
  table and the verbatim narration exactly.
- **Structural attractor:** the per-arm / per-replicate aggregates in the JSON
  regenerate from the run summary using the field list in the artifact's
  recipe section.
- **Determinism canary:** run the create phase then the restore phase as two
  separate processes and compare restored token ids against the saved control
  token ids (`ok` must be true).

## Scope and honesty notes

- The "structural attractor" discontinuity arm is an experiment arm, not an
  adopted default: it breaks the lock but raises invalid/recovery outputs
  (the cost is real and is reported).
- The determinism canary is temperature-0 only and exercises an 8-token budget;
  it does not claim long-horizon trajectory equivalence or stochastic-sampler
  determinism.
- The L43 chain separates the longer "perturbation → death" span (1,448 ticks)
  from the strict "active-belief" span (1,332 ticks); the headline uses the
  latter.