CARE: Conformal Risk Evaluation Framework

• CARE is a framework using conformal prediction to convert model outputs into calibrated risk assessments with finite-sample, distribution-free guarantees.
• It provides a post-hoc, model-agnostic safety layer, enabling calibrated omission and hallucination flags in applications like medical summarization and segmentation.
• CARE rigorously balances safety and efficiency by calibrating risk thresholds and controlling task-specific risks in diverse, high-stakes domains.

Conformal Assessment for Risk Evaluation (CARE) denotes the use of conformal prediction and conformal risk control to convert model outputs, score functions, or candidate decisions into calibrated risk statements with finite-sample, distribution-free guarantees. The designation is explicit in medical summarization, where CARE is a post-hoc, model-agnostic safety layer that overlays calibrated omission and hallucination flags onto summaries from any LLM without retraining (Bedi et al., 8 Jun 2026). The same label is also used more broadly as an organizing interpretation for conformal frameworks that assess segmentation false-negative risk, regulate policy deployment, certify decision suboptimality, or control risk among selectively trusted predictions (Shahid, 18 Jun 2026, Prinster et al., 2 Mar 2026, Zhou et al., 19 May 2025, Bai et al., 25 Mar 2026).

1. Conceptual scope

CARE departs from the classical use of conformal prediction as a coverage wrapper around a predictor. In the standard formulation, the target guarantee is miscoverage control, typically written as $\mathbb{P}(Y \in \Gamma(X)) \ge 1-\alpha$. In CARE-style systems, the object of interest is instead a task-specific risk functional: a document-level hallucination event, a fractional omission rate, a per-volume false-negative rate, a false discovery rate, a feasibility violation, a squared-error budget, or the probability that a candidate decision is suboptimal (Bedi et al., 8 Jun 2026, Hulsman et al., 2024, Zhou et al., 19 May 2025, Bai et al., 25 Mar 2026).

This broader scope has two complementary interpretations. In the first, conformal methods are used for risk control: a control parameter such as a threshold, set size, likelihood-ratio cap, or safety margin is calibrated so that expected risk is below a user-specified budget $\alpha$. In the second, conformal methods are used for risk assessment: given a fixed interval, prediction set, or decision, the procedure estimates or upper-bounds the probability that the object fails to meet a prescribed criterion. The latter viewpoint is explicit in inverse conformal methods for regression and multi-class classification, where conformal machinery is inverted to assess the failure probability of externally specified prediction objects (Singh et al., 2023, Ghandwani et al., 2024).

A recurring feature of CARE is that it is typically post-hoc and model-agnostic. The medical summarization CARE framework does not retrain or modify the summarizer; it calibrates sentence-level flags from judge scores (Bedi et al., 8 Jun 2026). Federated CRC for brain tumor segmentation calibrates thresholds on top of a fixed pre-trained SegResNet (Shahid, 18 Jun 2026). Conformalized policy control uses any safe reference policy as a probabilistic regulator for any optimized but untested policy (Prinster et al., 2 Mar 2026). CREDO similarly audits any candidate decision by combining inverse optimization geometry with conformal prediction and generative modeling (Zhou et al., 19 May 2025).

2. Statistical foundations

The common mathematical core is a monotone family of prediction objects indexed by a scalar or low-dimensional parameter and a bounded loss function. In segmentation CRC, the nested prediction set is

$C_\lambda(X) = \{ v : f(X)_v \ge 1-\lambda \},$

with per-volume false-negative loss

$\ell(C_\lambda,Y)=1-\frac{|C_\lambda\cap Y|}{|Y|},$

which is non-increasing in $\lambda$ and bounded in $[0,1]$ (Shahid, 18 Jun 2026). In the generic CRC theorem used by medical summarization CARE, if $L(\lambda;X,Y)\in[0,1]$ is non-increasing in $\lambda$, then

$$\lambda^* = \inf\left\{ \lambda \in \Lambda : \frac{1}{n+1} \left( \sum_{i=1}^{n} L(\lambda;X_i,Y_i) + 1 \right) \le \alpha \right\}$$

guarantees

$$\mathbb{E}\left[L(\lambda^*;X_{n+1},Y_{n+1})\right]\le \alpha$$

under exchangeability (Bedi et al., 8 Jun 2026).

CARE systems often require richer parameterizations than a single threshold. The omission controller in medical summarization is intrinsically two-dimensional because a source sentence is surfaced only if both

$\alpha$0

The omission loss is a fractional document-level loss,

$\alpha$1

with the convention $\alpha$2 when $\alpha$3, and calibration is performed jointly over the full $\alpha$4 grid via Learn-Then-Test fixed-sequence ordering (Bedi et al., 8 Jun 2026).

A second recurrent pattern is inverse assessment. In regression-based risk assessment, InvCP defines

$\alpha$5

for a user-specified interval $\alpha$6 and a conformal interval family $\alpha$7, and then estimates the global failure level by averaging $\alpha$8 over unlabeled inputs. This yields a conservative lower bound on $\alpha$9 under exchangeability or covariate shift, depending on the conformal backend (Singh et al., 2023). An analogous inversion appears in multi-class classification, where the smallest conformal level whose prediction set contains the model’s own top-$C_\lambda(X) = \{ v : f(X)_v \ge 1-\lambda \},$0 output is used as an instance-wise misclassification-risk estimate (Ghandwani et al., 2024).

Non-exchangeable and distribution-shifted variants extend the same template by weighting calibration examples. Non-exchangeable CRC replaces the empirical risk by a weighted empirical risk and yields

$C_\lambda(X) = \{ v : f(X)_v \ge 1-\lambda \},$1

so the guarantee degrades through a total-variation slack term rather than disappearing entirely (Farinhas et al., 2023).

3. Methodological patterns and representative systems

The current literature contains several distinct CARE instantiations that differ in their prediction objects, loss definitions, and guarantee types.

Instantiation	Calibrated object	Controlled quantity
Medical summarization CARE (Bedi et al., 8 Jun 2026)	Sentence-level omission and hallucination flags	Document-level unflagged hallucination probability; expected fraction of important omissions not surfaced
Federated CRC for segmentation (Shahid, 18 Jun 2026)	Site-specific segmentation thresholds	Expected false-negative rate under a site-mixture model
CREDO (Zhou et al., 19 May 2025)	Risk certificate for a candidate decision	Probability that the decision is suboptimal
SCoRE (Bai et al., 25 Mar 2026)	Binary trust decisions	Marginal Deployment Risk and Selective Deployment Risk
Conformal Policy Control (Prinster et al., 2 Mar 2026)	Likelihood-ratio cap between safe and optimized policies	Expected constraint violation under the deployed policy
Action-conditional risk-averse CP (Zhu et al., 4 Jun 2026)	Action-conditioned prediction sets	Safety guarantees conditioned explicitly on each action

These systems exhibit several recurring design motifs. One is the safety-layer architecture, in which an existing model is left unchanged and conformal calibration is applied to auxiliary scores. Medical summarization CARE uses judge-model support, importance, and coverage scores; the calibrated outputs are red and blue flags rather than revised summaries (Bedi et al., 8 Jun 2026). A closely related motif is selective trust, where the conformal layer decides when a black-box model may be used. SCoRE constructs generalized e-values whose product with the unknown risk has expectation no greater than one, and then passes them to hypothesis-testing procedures to obtain trust decisions with finite-sample error control (Bai et al., 25 Mar 2026).

A second motif is decision regulation. Conformal Policy Control constrains a new policy by clipping its likelihood ratio relative to a safe reference policy,

$C_\lambda(X) = \{ v : f(X)_v \ge 1-\lambda \},$2

and calibrates the largest admissible $C_\lambda(X) = \{ v : f(X)_v \ge 1-\lambda \},$3 from safe-policy data so that expected loss remains below the user’s risk tolerance (Prinster et al., 2 Mar 2026). In robotics, CRC is used to calibrate safety margins around Control Barrier Function predictions, yielding stepwise probabilistic guarantees on safety-constraint satisfaction (Gonzales et al., 11 Mar 2026).

A third motif is risk certificates for externally proposed objects. CREDO does not produce a decision by itself; it quantifies, for any candidate decision, a distribution-free upper bound on the probability that the decision is suboptimal by constructing conformal balls in outcome space and forcing them inside the inverse optimality region of the decision (Zhou et al., 19 May 2025). This certificate view is closely aligned with inverse conformal risk assessment in regression and classification (Singh et al., 2023, Ghandwani et al., 2024).

4. Marginal, group-specific, and action-conditional guarantees

A central issue in CARE is the distinction between guarantees that hold on average and guarantees that hold within groups, sites, or actions. The medical summarization CARE guarantees are explicitly marginal over the deployment distribution: hallucination risk is the expected indicator that a document contains an unflagged hallucinated sentence, and omission risk is the expected fraction of important omissions not surfaced in a future document (Bedi et al., 8 Jun 2026). The paper also states that these guarantees do not promise per-document worst-case control.

The federated segmentation setting makes this tension concrete. Under a site-mixture model, pooled CRC preserves marginal CRC guarantees, but nothing implies that each hospital satisfies the target risk individually. The paper distinguishes marginal (site-mixture) coverage from per-site coverage and notes that exact distribution-free group-conditional coverage is impossible without strong structural assumptions (Shahid, 18 Jun 2026). The empirical consequence is that the average hospital can be protected while vulnerable hospitals are not.

Several recent systems therefore move from marginal to more structured guarantees. SCoRE defines Marginal Deployment Risk

$C_\lambda(X) = \{ v : f(X)_v \ge 1-\lambda \},$4

and Selective Deployment Risk

$C_\lambda(X) = \{ v : f(X)_v \ge 1-\lambda \},$5

so the guarantee is about risk among positive, trusted cases rather than risk over the full population (Bai et al., 25 Mar 2026). Action-conditional conformal prediction goes further by requiring

$C_\lambda(X) = \{ v : f(X)_v \ge 1-\lambda \},$6

which yields safety guarantees conditioned explicitly on each action taken by the decision maker (Zhu et al., 4 Jun 2026).

This progression suggests a general CARE taxonomy. Marginal guarantees are often the easiest to obtain and the strongest in finite-sample distribution-free form. Group-, site-, trust-, or action-conditional guarantees are operationally closer to how risk is perceived in deployment, but they require additional structure, more refined calibration, or a narrower decision space. A plausible implication is that CARE systems should report both levels whenever possible: a formally guaranteed marginal statement and explicit diagnostics for the subpopulations induced by deployment.

5. Empirical behavior and trade-offs

Empirical studies consistently show that CARE systems expose safety–efficiency trade-offs that ordinary accuracy metrics conceal. In medical summarization, CARE satisfies the target risk bound at $C_\lambda(X) = \{ v : f(X)_v \ge 1-\lambda \},$7 with 95% confidence across 100 calibration/test resplits, using only approximately 100 labeled documents per domain. Joint calibration over the full $C_\lambda(X) = \{ v : f(X)_v \ge 1-\lambda \},$8 threshold space surfaces up to $C_\lambda(X) = \{ v : f(X)_v \ge 1-\lambda \},$9 fewer sentences than alternative calibrated baselines, and in a preliminary clinician study of 75 document reviews, calibrated flags improved omission detection by 28.6 percentage points on average (Bedi et al., 8 Jun 2026).

The federated segmentation study shows a sharper failure mode. On FeTS-2022, naive pooled CRC yields mean violations of $\ell(C_\lambda,Y)=1-\frac{|C_\lambda\cap Y|}{|Y|},$0 out of 20 sites, worst-site FNR $\ell(C_\lambda,Y)=1-\frac{|C_\lambda\cap Y|}{|Y|},$1, and stretch $\ell(C_\lambda,Y)=1-\frac{|C_\lambda\cap Y|}{|Y|},$2 at $\ell(C_\lambda,Y)=1-\frac{|C_\lambda\cap Y|}{|Y|},$3, whereas local per-site CRC reduces violations to $\ell(C_\lambda,Y)=1-\frac{|C_\lambda\cap Y|}{|Y|},$4 and worst FNR to $\ell(C_\lambda,Y)=1-\frac{|C_\lambda\cap Y|}{|Y|},$5 but inflates stretch to $\ell(C_\lambda,Y)=1-\frac{|C_\lambda\cap Y|}{|Y|},$6. Shrinkage-based federated CRC interpolates between these extremes: with $\ell(C_\lambda,Y)=1-\frac{|C_\lambda\cap Y|}{|Y|},$7, chosen by leave-one-site-out sensitivity analysis, it achieves $\ell(C_\lambda,Y)=1-\frac{|C_\lambda\cap Y|}{|Y|},$8 violations, worst FNR $\ell(C_\lambda,Y)=1-\frac{|C_\lambda\cap Y|}{|Y|},$9, and stretch $\lambda$0; removing the finite-sample correction triples violations (Shahid, 18 Jun 2026).

Risk-sensitive conformal prediction for catheter placement detection illustrates the same principle in a classification setting with asymmetric clinical stakes. The risk-sensitive scheme uses $\lambda$1 for critical findings and $\lambda$2 elsewhere, achieving 90.68% overall empirical coverage, 99.29% coverage for critical conditions, and zero high-risk mispredictions, at the cost of larger average prediction sets than standard conformal prediction (Hui, 28 May 2025).

These results indicate that CARE is often most informative when average performance looks acceptable. Pooled federated CRC “protects the average hospital” while failing vulnerable sites (Shahid, 18 Jun 2026); uncalibrated summarization error detectors produce scores with no formal missed-error guarantee (Bedi et al., 8 Jun 2026); naive decision pipelines can be brittle under multi-modal uncertainty, whereas CREDO produces interpretable risk certificates for candidate decisions (Zhou et al., 19 May 2025). CARE therefore functions less as a replacement for predictive performance evaluation than as a complementary layer that reveals where statistical reliability is actually concentrated.

6. Limitations, failure modes, and open directions

The dominant limitation is the usual conformal one: guarantees rely on calibration and deployment being sufficiently aligned. Medical summarization CARE states that CRC assumes calibration and deployment data are drawn from the same distribution and that domain shift can weaken guarantees, requiring periodic recalibration (Bedi et al., 8 Jun 2026). Non-exchangeable CRC formalizes this degradation through a total-variation slack term rather than an abrupt failure, but the required distances are not directly observable and must be mediated through weighting heuristics (Farinhas et al., 2023).

A second limitation is that risk definitions are only as good as their labeling interface. In medical summarization, omission and hallucination risk are defined relative to oracle labels from a strong LLM, not ground-truth clinical correctness (Bedi et al., 8 Jun 2026). In CLIPScore uncertainty calibration, the framework explicitly does not debias CLIPScore itself; it calibrates risk around the metric’s outputs (Gomes et al., 1 Apr 2025). In CREDO, validity does not require a perfect generative model, but tightness depends on the quality of the model and on the geometry of the conformal balls (Zhou et al., 19 May 2025).

A third limitation concerns conditionality and fairness. The federated CRC paper does not claim formal per-site guarantees; it empirically evaluates per-site FNRs as an approximate notion of group-conditional coverage (Shahid, 18 Jun 2026). SCoRE controls risk among trusted cases, but not arbitrary subgroup-conditional risk (Bai et al., 25 Mar 2026). The action-conditional framework explicitly points toward stronger forms of conditional safety, and briefly notes extensions to CVaR as future work (Zhu et al., 4 Jun 2026). This suggests that one open direction for CARE is a principled hierarchy of guarantees—marginal, group-conditional, action-conditional, and severity-weighted—matched to the operational stakes of the domain.

A final limitation is computational and communicative. Some CARE procedures are inexpensive, such as transmitting only a vector of $\lambda$3 scalars per site in federated segmentation (Shahid, 18 Jun 2026). Others require repeated generative sampling, semantic clustering, inverse-optimization geometry, or per-action pinball-loss calibration (Zhou et al., 19 May 2025, Wang et al., 2024, Zhu et al., 4 Jun 2026). A plausible implication is that CARE will remain most deployable when risk summaries are low-dimensional, auditable, and closely aligned with concrete operational decisions rather than with generic confidence scores alone.