Claim-Evidence Contract in Scientific Discovery

• Claim-Evidence Contract is a systematic process that binds scientific claims to predefined evidence thresholds, ensuring claims remain within evidential scope.
• It utilizes an explicit epistemic state to coordinate hypothesis formation, evidence acquisition, and local adjudication in open-ended discovery tasks.
• Empirical evaluations indicate that this framework balances investigation yield and evidential rigor, enhancing the overall quality of scientific discovery.

Searching arXiv for the primary paper and closely related claim–evidence contract work. arXiv search query: "StatefulDiscovery: Evidence-Calibrated Claim Formation in Open-Ended Scientific Discovery" StatefulDiscovery is a discovery framework for open-ended scientific discovery in which an agent must decide, across multiple rounds of exploration, which phenomena warrant investigation while avoiding overinterpretation, defined as situations in which emerging claims exceed the evidential scope of the analyses supporting them. The framework treats this as an evidence-calibration problem: the exploration trajectory must be coupled with claim status so that evidence can guide both what to investigate next and what can be claimed. Its central mechanism is an explicit epistemic state that coordinates frontier selection, evidence acquisition, and claim adjudication, so that the semantic strength of each claim never exceeds its empirical support (Chen et al., 10 Jun 2026).

1. Problem setting and conceptual scope

Open-ended scientific discovery differs from workflows that execute analyses for predefined questions. In the formulation associated with StatefulDiscovery, the core difficulty is not only selecting analyses, but also determining when an observed pattern merits escalation into an explicit claim. The paper identifies a failure mode of overinterpretation, in which investigation and claim formation drift apart: analyses accumulate locally, but the resulting claims are stated more strongly than the available evidence warrants (Chen et al., 10 Jun 2026).

StatefulDiscovery addresses this by externalizing discovery state rather than leaving exploratory progress implicit in prompt context or intermediate text generations. This explicit state is used to regulate three coupled processes: which unexplained patterns remain on the frontier, which structured hypotheses are under active consideration, and which candidate claims have achieved a status such as supported, weakened, refuted, or inconclusive. This design places claim formation under an explicit claim–evidence contract rather than treating claims as direct byproducts of whichever analysis happened to run most recently (Chen et al., 10 Jun 2026).

A common misconception is that open-ended discovery can be reduced to iterative data analysis plus post hoc summarization. StatefulDiscovery rejects that assumption. Its formulation suggests that discovery quality depends on whether exploration policy and claim adjudication are jointly controlled. A plausible implication is that exploratory breadth without adjudication can inflate discovery value while degrading evidential support, whereas adjudication without frontier management can improve local calibration but still misallocate budget across the search space.

2. Explicit epistemic state and round-level dynamics

At round $t$, StatefulDiscovery maintains an explicit epistemic state

$S_t = (H_t, E_t, F_t, C_t),$

where $H_t = \{h_{t,1},\dots,h_{t,n}\}$ is the set of structured hypotheses under active consideration, $E_t = \{e_1,\dots,e_m\}$ is the accumulated evidence pool, $F_t = \{p_1,\dots,p_k\}$ is the frontier of unexplained patterns or phenomena with priority scores $\pi(p_j)\in[0,1]$, and $C_t = \{c_1,\dots,c_\ell\}$ is the set of candidate claims, each tagged with a status and confidence (Chen et al., 10 Jun 2026).

After initialization, the agent seeds patterns via exploratory data analysis and sets

$S_0 = (\emptyset, E_0, F_0, \emptyset).$

The round structure is then organized as a controlled loop. The frontier-control policy $\pi(\cdot)$ chooses an action $a_t$ from create, deepen, switch, retire, or stop. If an investigation is active, the agent decomposes it into $S_t = (H_t, E_t, F_t, C_t),$0, issues executable queries to grow $S_t = (H_t, E_t, F_t, C_t),$1, and applies a local adjudication function $S_t = (H_t, E_t, F_t, C_t),$2 to update hypothesis statuses and append new claims to $S_t = (H_t, E_t, F_t, C_t),$3. The frontier $S_t = (H_t, E_t, F_t, C_t),$4 and pattern priorities are then revised based on red flags, unresolved gaps, and budget (Chen et al., 10 Jun 2026).

This round-level decomposition is significant because it turns discovery into a stateful control problem rather than a sequence of loosely coupled analyses. The state variables are not merely records. They determine what remains investigable, what remains claimable, and when stopping is justified. Related work in other domains uses analogous explicit state to separate claims, evidence, and revision scope. IMPACT-CYCLE stores a versioned semantic memory $S_t = (H_t, E_t, F_t, C_t),$5 with typed claims, a claim-dependency graph, and a provenance log for long-video understanding (Kong et al., 22 Apr 2026). PaperTrail likewise decomposes both source papers and answer text into discrete claims and evidence for scholarly QA provenance (Martin-Boyle et al., 24 Feb 2026). These parallels suggest that StatefulDiscovery belongs to a broader design pattern in which latent reasoning is replaced by auditable state transitions.

3. Structured hypotheses and evidential requirements

Each investigation in StatefulDiscovery is associated with a small hypothesis set

$S_t = (H_t, E_t, F_t, C_t),$6

so that the agent does not treat a salient pattern as a single undifferentiated proposition. A hypothesis may be expressed as a statistical model or logical formula. The paper gives, as an example, a main interaction hypothesis

$S_t = (H_t, E_t, F_t, C_t),$7

with test statement

$S_t = (H_t, E_t, F_t, C_t),$8

This representation embeds alternative explanations and artifact checks directly into the investigation object, rather than attaching them later as reviewer concerns (Chen et al., 10 Jun 2026).

Each hypothesis carries attached evidential requirements, including minimum sample size $S_t = (H_t, E_t, F_t, C_t),$9, significance threshold $H_t = \{h_{t,1},\dots,h_{t,n}\}$0, robustness checks such as bootstrap CI width $H_t = \{h_{t,1},\dots,h_{t,n}\}$1, and artifact-check constraints such as no group–page confounding at $H_t = \{h_{t,1},\dots,h_{t,n}\}$2. The summary gives a local support score

$H_t = \{h_{t,1},\dots,h_{t,n}\}$3

with $H_t = \{h_{t,1},\dots,h_{t,n}\}$4 required to mark $H_t = \{h_{t,1},\dots,h_{t,n}\}$5 as supported (Chen et al., 10 Jun 2026).

This structure clarifies what “support” means operationally. Support is not a generic confidence impression; it is thresholded against declared requirements. That makes the framework distinct from systems that score claim plausibility without explicitly binding claims to preregistered or locally declared evidence conditions. A related but stricter protocol-level formulation appears in Preregistered Belief Revision Contracts, where a non-fallback step is accepted only when it cites a preregistered trigger and provides a nonempty witness set of externally validated evidence tokens (Alqithami, 16 Apr 2026). In a different application area, VERI-DPO binds clinical summary statements to claim-level labels Supported, Not Supported, or Not Addressed, and aggregates them into a coverage-aware utility for preference mining (Liu et al., 11 Mar 2026). Across these systems, the common idea is that admissible claims or revisions are those that satisfy explicit evidence conditions rather than merely coherent generation.

4. Local adjudication and frontier control

After a query bundle is executed, the evidence-strength-judge computes

$H_t = \{h_{t,1},\dots,h_{t,n}\}$6

by checking evidence sanity, method fit, cross-hypothesis comparison, and confidence calibration. Confidence is assigned as $H_t = \{h_{t,1},\dots,h_{t,n}\}$7, capped by red-flag pressure (Chen et al., 10 Jun 2026).

The summary provides a simple decision rule. For each $H_t = \{h_{t,1},\dots,h_{t,n}\}$8, the procedure computes $H_t = \{h_{t,1},\dots,h_{t,n}\}$9-value $E_t = \{e_1,\dots,e_m\}$0 and sample size $E_t = \{e_1,\dots,e_m\}$1 from $E_t = \{e_1,\dots,e_m\}$2; if $E_t = \{e_1,\dots,e_m\}$3 or method check fails, status becomes inconclusive with $E_t = \{e_1,\dots,e_m\}$4; if $E_t = \{e_1,\dots,e_m\}$5, status becomes supported; if $E_t = \{e_1,\dots,e_m\}$6, status becomes refuted; otherwise it becomes weakened. The rubric uses thresholds like $E_t = \{e_1,\dots,e_m\}$7, $E_t = \{e_1,\dots,e_m\}$8, and caps confidence at $E_t = \{e_1,\dots,e_m\}$9 when red flags appear (Chen et al., 10 Jun 2026).

This local adjudication stage performs two roles simultaneously. First, it arbitrates among competing explanations within an investigation. Second, it prevents claims from being written unless their local evidential requirements are satisfied. The architecture therefore opposes a common failure mode in autonomous analysis systems: claims are not generated first and audited later, but are generated as outputs of adjudication. A useful comparison is EACon, which deconstructs a complex claim into atomic subclaims and verifies each against raw and abstracted evidence, setting overall veracity to false if any subclaim is false and overall confidence to the minimum of subclaim confidences (Gong et al., 2024). StatefulDiscovery applies a similar discipline earlier in the discovery cycle, before claims are finalized.

Frontier control then determines what the system does next. The level-1 strategist reads frontier priorities $F_t = \{p_1,\dots,p_k\}$0, active-investigation resolution signals, frontier saturation, red-flag pressure, and remaining budget, and chooses the next action by optimizing a trade-off between expected evidence gain and claim-overreach risk:

$F_t = \{p_1,\dots,p_k\}$1

Here $F_t = \{p_1,\dots,p_k\}$2 is the expected increase in aggregate evidence-support, $F_t = \{p_1,\dots,p_k\}$3 grows when current claims have low confidence or high red-flag pressure, and $F_t = \{p_1,\dots,p_k\}$4 trades off exploration against claim-conservatism (Chen et al., 10 Jun 2026).

Operationally, if the active investigation is unresolved and budget remains, $F_t = \{p_1,\dots,p_k\}$5 tends to deepen; if resolved with high confidence, $F_t = \{p_1,\dots,p_k\}$6 creates or switches to a new high-priority pattern; if frontier saturation is high or red flags dominate, $F_t = \{p_1,\dots,p_k\}$7 may retire or stop (Chen et al., 10 Jun 2026). This makes stopping behavior endogenous to evidential state rather than a hard round limit. A plausible implication is that the framework treats unresolvable ambiguity as a reason to curtail claim formation, not merely as a temporary inconvenience.

5. Empirical evaluation, ablations, and case study

StatefulDiscovery is evaluated across 40 real-data discovery tasks spanning biomedical, social-science, behavioral, and cross-domain settings. Over these tasks it produces 261 final claims, of which 64, or 24.5%, are high-quality claims defined as $F_t = \{p_1,\dots,p_k\}$8 and $F_t = \{p_1,\dots,p_k\}$9. The strongest baseline, SAGA, produces 52 high-quality claims out of 250, or 20.8% (Chen et al., 10 Jun 2026).

In blind pairwise LLM judging over the full claim sets, StatefulDiscovery is preferred over SAGA on 31/40 tasks and over all other baselines by at least 25/40 tasks (Chen et al., 10 Jun 2026). The article’s empirical claim is therefore not only that the framework raises the count of high-quality claims, but also that its claim sets are preferred under comparative evaluation.

The ablation study is especially important because it reveals a nontrivial trade-off between structure and calibration. Adding structured hypotheses to a raw agent increases average Discovery Value from 2.47 to 3.23, but lowers average Evidential Support from 4.22 to 3.18. Introducing L2 adjudication recovers Evidential Support to 3.63 and yields a high-quality rate of 26.8%. Finally adding L1 frontier control balances yield and quality, reaching $\pi(p_j)\in[0,1]$0, $\pi(p_j)\in[0,1]$1, and 24.5% high-quality claims (Chen et al., 10 Jun 2026).

These numbers directly counter a simplistic reading in which more structure automatically improves every metric. Instead, the evidence indicates a staged interaction: structured hypotheses increase yield and value, but without adjudication can lower evidential support; adjudication recovers support; frontier control then balances the portfolio of investigations. The empirical picture is therefore one of controlled compensation rather than monotonic gain.

The hurricane-data case study illustrates how this differs from surprise-driven updating. AutoDiscovery’s surprise-driven update often “rewards” hypotheses that simply fail under evidence, described as negative surprise, whereas StatefulDiscovery decomposes the female-name–death association into four linked investigations: gender causality, naming-convention artifact, time-period clustering, and wind-speed confounding. It only issues cautious claims after passing each artifact and robustness check (Chen et al., 10 Jun 2026). This case clarifies that the framework’s conservatism is not reluctance to discover; it is a requirement that explanatory decomposition and artifact analysis precede strong claim issuance.

6. Relation to adjacent claim–evidence contract frameworks

StatefulDiscovery is one instance of a broader family of claim–evidence contract systems, but its domain and control structure are distinctive. In long-video semantic memory correction, IMPACT-CYCLE reformulates understanding as iterative claim-level maintenance of a shared semantic memory with typed claims, a claim dependency graph, and a provenance log, and it restricts corrections to structurally dependent claims; on VidOR it improves downstream VQA from 0.71 to 0.79 and reduces human arbitration cost by 4.8x (Kong et al., 22 Apr 2026). In financial QA auditing, EvidenceLens uses a multimodal claim-evidence matrix whose cells carry support, partial, contradiction, or context labels, with a deterministic review-priority ranking that surfaces confidence–support gaps and modality imbalance (Gu et al., 19 Jun 2026). In scholarly QA, PaperTrail maps answer claims to paper claims and evidence and exposes supported assertions, unsupported claims, and omitted paper claims, while a within-subjects study with 26 researchers shows lower subjective trust without significant behavioral change (Martin-Boyle et al., 24 Feb 2026).

These systems share several structural commitments: decomposition of outputs into atomic or typed claims, explicit representation of evidence units, local verification or adjudication, and auditability through state, provenance, or alignment artifacts. VERI-DPO applies the same logic to clinical summarization by using a retrieval-augmented verifier that labels claim-evidence pairs as Supported, Not Supported, or Not Addressed, mines 1,513 preference pairs, and reduces Not Supported claim rates from 10.7% to 1.9% under a local verifier judge while maintaining informative length (Liu et al., 11 Mar 2026). EACon likewise addresses noisy evidence and multi-aspect claims through evidence abstraction and claim deconstruction, verifying subclaims separately rather than judging a complex claim holistically (Gong et al., 2024).

StatefulDiscovery differs from these systems in where the contract is enforced. Most related frameworks operate after a candidate answer, memory, or summary exists. StatefulDiscovery enforces the contract during discovery itself: frontier selection, evidence acquisition, and claim formation are jointly stateful. A plausible implication is that it moves claim–evidence alignment upstream, converting it from a post hoc auditing layer into a generative control principle. In that sense, it is closer in spirit to PBRC, where admissible epistemic change is constrained by public triggers and validated witnesses, than to interfaces that only visualize support after the fact (Alqithami, 16 Apr 2026).

The broader significance is methodological. By externalizing investigation state and binding claim status to local evidence thresholds and frontier policy, StatefulDiscovery offers a formal answer to a persistent problem in autonomous discovery systems: how to let exploration remain open-ended without allowing claims to outrun evidence. The results suggest that explicit discovery state can couple exploration with evidence-calibrated claim formation (Chen et al., 10 Jun 2026).