SPINE synthetic simulator
Analyze actual study data →Compare how a viability filter, a soft preference, and a task-only controller act on the same synthetic conditions. Freeze the parameters, inspect every decision, and export evidence for your study.
This is an experimental implementation of selected v7.4 mechanisms with declared synthetic dynamics. It is separate from the conversational C5 scores and does not constitute formal, clinical, or embodied validation.
Configure an experiment
Dynamics, uncertainty, and withdrawal parameters
Preview runs in this browser. Sign in to save and attach results to a study.
Frozen protocol and implementation choices
- version
- spine-lab-0.2.0
- paper
- Beyond Solipsistic Robotics v7.4
- sourceRevision
- 88189a8d4fd14bb49bbd0f2d4cf1d86150ceed0e
- sourceUrl
- https://github.com/Mommythegeek/protect-what-thinks
- status
- Experimental synthetic implementation, not an official upstream simulator or formal validation.
- state
- Three normalized reserve margins representing stress, autonomic activation, and movement capacity; all must be >= 0.
- horizon
- 1
- predictiveRule
- Independent Gaussian one-step marginal predictions; joint viability is the product of three Gaussian CDFs.
- varianceRule
- Eq. 7 is extended to three margins using their mean posterior variance; no resilience-adaptive Eq. 8 extension.
- dynamics
- reserve_next = 0.96 * reserve + action_delta + 0.015 + 0.045 * autonomy + recovery_drive + Gaussian process noise.
- taskObjective
- Declared discrete task-cost surrogate, not the full active-inference EFE of Eq. 1.
- withdrawal
- L2 permits autonomous action only if KL(current||previous) < deltaSat AND observed recovery success ratio >= rhoMin. Safety filtering still applies.
- resilience
- Ratio successes/(trials+1). A trial starts at an announced bounded perturbation. Recovery within 6 cycles counts only if observed margins recover without intervention intensity > 0.25.
- failSafe
- If normal L2 admission is empty, assist/stabilize may reduce expected negative reserve deficit relative to passive waiting, without worsening any coordinate. Choose minimum total deficit, then least intensity. This recovery extension is uncertified; passive fallback remains if no improvement exists. No hidden state is used.
- reference
- Before outcomes: true current margins and true autonomy, known one-step dynamics and process noise, fixed epsilonMax. Reference state is evaluator-only.
- se
- First cycle at which the last 10 realized transitions have AE >= 0.8; null if never reached.
- ss
- Lag from a margin crossing below 0.15 to a different selected action within 10 cycles. Undetected and right-censored events are reported separately.
- uncertainty
- Episode-level percentile bootstrap (500 resamples); independent episode streams assumed. Exogenous noise is paired across controllers within each episode.
- coverage
- Synthetic P1/P4/P5 diagnostics and a handoff stress test. No robot self-model or P2 bidirectional update; no clinical or embodied validation.
- expectation
- No target ranking or anticipated AE values are encoded as results.