# Conformal CBVF Safety Shield

- ID: 2263
- Canonical URL: https://synthcore.org/idea/2263/conformal-cbvf-safety-shield
- API JSON: https://synthcore.org/api/idea/2263.json
- API Markdown: https://synthcore.org/api/idea/2263.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.13819](https://arxiv.org/abs/2608.13819)
- Category: dynamics
- Solves: stability, accuracy, generalization
- ML areas: rl, rnn, transformer, optimizer
- Math tags: control-theory, dynamical-systems, probability, optimization
- Ratings: usefulness 8/10; difficulty 6/10; novelty 7/10

## Idea description

Wrap an observation-based neural policy with a real-time safety filter that accounts for uncertainty in its latent-state estimate. The policy proposes an action, while a quadratic program minimally modifies that action so a control-barrier/value function remains nonnegative for every state inside a conformally calibrated error set.

## Mathematical statement

Let x in R^n be the true state, y the observation, z=E_phi(y) the neural state estimate, and e=x-z the estimation error. Given N exchangeable calibration samples, compute residual scores r_i=||x_i-E_phi(y_i)|| and choose the conformal order statistic q_(1-alpha). Then the error event ||e||<=q has probability approximately at least 1-alpha under exchangeability. Let V(t,x) be a differentiable control-barrier/value function, with V>=0 defining the safe viability region, and let the dynamics be dot{x}=f(x,u,d), where u is the action and d belongs to a bounded disturbance set D. The robust estimated-state certificate is the minimum of V over the uncertainty ball. A sufficient online condition is that the worst-case derivative of V over all errors in the ball and disturbances in D is at least -gamma V, where gamma>=0. The shield solves for the action closest to the neural policy action subject to this robust inequality.

## Key formulas

- $$q_{1-alpha}=r_{(ceil((N+1)(1-alpha)))} ,\qquad r_i=||x_i-E_phi(y_i)||,\qquad P(||x-E_phi(y)||<=q_{1-alpha}) approximately >=1-alpha.$$
- $$V_hat(t,z)=min_{||e||<=q_{1-alpha}} V(t,z+e),\qquad V_hat(t,z)>=0 implies V(t,x)>=0 when ||x-z||<=q_{1-alpha}.$$
- $$u_star=argmin_{u in U} (1/2)||u-pi_theta(y)||_R^2$$
- $$subject\ to\ min_{||e||<=q, d in D} [partial_t V(t,z+e)+grad_x V(t,z+e)^T f(z+e,u,d)+gamma V(t,z+e)]>=0.$$

## Implementation notes

1. Integration point: train an observation-based policy pi_theta(y), using an RNN, transformer, or recurrent world model, but execute it through a safety layer. The estimator E_phi produces z. Fit or provide a differentiable certificate V(t,x), and use automatic differentiation to obtain its state gradient. 2. Pseudocode: collect N calibration tuples (y_i,x_i), compute r_i=norm(x_i-E_phi(y_i)), and set q to the conformal order statistic. At runtime compute z=E_phi(y), u0=pi_theta(y), generate error candidates e_j by projected optimization inside the ball norm(e)<=q, enumerate disturbance vertices d_k when D is a polytope, and solve a quadratic program minimizing norm(u-u0)_R subject to the barrier derivative inequality for each pair (e_j,d_k). Apply the resulting u and repeat. 3. Computed versus estimated: q is computed directly from held-out residuals; V and f are model-based or learned; the inner worst-case minimization is approximated by projected gradient ascent, random samples, or conservative Lipschitz bounds. 4. First experiment: use a two-dimensional point-mass obstacle-avoidance simulator with noisy position observations, an MLP or GRU policy, and compare an unshielded policy with a nominal barrier filter and the conformal shield. Sweep alpha in {0.01,0.05,0.1,0.2} and calibration size N. The prediction is that empirical estimator coverage is near or above 1-alpha, unsafe-episode frequency decreases as alpha decreases, and intervention magnitude increases with q. A sharp feasibility boundary should appear when q approaches the obstacle-clearance margin: beyond that point V_hat becomes negative and the QP becomes frequently infeasible or strongly corrective.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: inconclusive
- Verdict: Built a self-contained conformal CBVF toy experiment with split-conformal calibration, an analytic robust barrier shield, parameter sweeps, and closed-loop comparison. The math predictions manifested: mean coverage was 0.9133, 0.9029, 0.9000, and 0.9003 for N=20, 50, 200, and 1000 against a 0.90 target; the observed certificate boundary was q=0.50 versus predicted q=0.50; and the observed intervention slope was 0.800000 versus predicted gamma=0.8. The safety rollout had zero unsafe episodes for nominal and conformal shields, so this toy does not demonstrate a safety-rate win, while conformal intervention increased from 0.7638 nominal to 0.8937--0.9161.

### Mechanism check

- Verdict: Built a self-contained conformal CBVF toy experiment with split-conformal calibration, an analytic robust barrier shield, parameter sweeps, and closed-loop comparison. The math predictions manifested: mean coverage was 0.9133, 0.9029, 0.9000, and 0.9003 for N=20, 50, 200, and 1000 against a 0.90 target; the observed certificate boundary was q=0.50 versus predicted q=0.50; and the observed intervention slope was 0.800000 versus predicted gamma=0.8. The safety rollout had zero unsafe episodes for nominal and conformal shields, so this toy does not demonstrate a safety-rate win, while conformal intervention increased from 0.7638 nominal to 0.8937--0.9161.
- Confidence: 8/10
- Limitations: This is a low-dimensional analytic toy, not an RNN/transformer policy or a general QP implementation. The disturbance and estimator errors are simple exchangeable distributions, continuous-time guarantees were not tested between sampling points, and the nominal shield was already perfectly safe, preventing a meaningful empirical safety improvement comparison.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.688
- Paired wins: 5/8
- Benchmark verdict: no significant win

## Artifacts

- [bench_report.json](https://synthcore.org/code/796/bench_report.json)
- [experiment.py](https://synthcore.org/code/796/experiment.py)
- [report.md](https://synthcore.org/code/796/report.md)
- [report_bench_2026-09-01T104505.md](https://synthcore.org/code/796/report_bench_2026-09-01T104505.md)
- [results.json](https://synthcore.org/code/796/results.json)
- [stage2_conformal_cbvf.py](https://synthcore.org/code/796/stage2_conformal_cbvf.py)
- [Download all files as ZIP](https://synthcore.org/download/796)

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.