VST: Verifiable Structured Transport for Auditable Agent-to-Agent Alpha Discovery

Yuqi Li; Siyuan Liu; Bingjun Liu

cs.AISep 7, 2026

VST: Verifiable Structured Transport for Auditable Agent-to-Agent Alpha Discovery

Authors: Yuqi Li, Siyuan Liu, Bingjun Liu

Organizations: Panda AI

Abstract

Agent-to-agent (A2A) alpha discovery is slowed by repeated feedback cycles between mining and evaluation agents, whose hand-offs, in contemporary LLM multi-agent systems, are free-form natural-language messages that carry no stable contract and cannot be replayed. We first restructure this communication as a structured agent-to-agent protocol of typed, causally addressable, unicast records, so that the committed stream forms a causal trajectory. On that trajectory a single predictor with four typed heads forecasts the accumulated guidance the two miners would receive several cycles ahead; a transactional verify--leap controller then commits a multi-cycle speculative outcome only when it passes a four-level gate, and otherwise rolls back to the exact prior state. Structure is the enabling contribution, and its value is not accuracy. A controlled ablation shows an equal-information free-text channel reaches the same predictor hit rate. What typing provides is a state that can be schema-checked, replayed deterministically, and prevented by construction from leaking a forecast to an evaluator: auditability by construction, not an empirically stress-tested guarantee. On a CSI~1000 out-of-sample holdout, our single run is the only one among eight methods (seven baselines and ours) to hold a positive median annualized return and Sharpe at the factor level, though the median return in excess of the benchmark stays negative for every method including ours; its development-selected top-20 portfolios reach a $0.71$ median holdout Sharpe, selected on a split inside the optimization horizon. We report these single-run results descriptively, gross of costs, and are explicit about their limits throughout; in particular we do not isolate the effect of the leap machinery from the inherited search substrate, which we leave to future work.

Figures & tables

Layer / field group	Helps transmission	Helps prediction
Envelope: identity	Deterministically identifies each transmission within a persisted round.	Records align into state-indexed trajectory observations $Z_{t}$ .
Envelope: routing	Receiver and its method argument are explicit.	Unrelated edges stay distinct; inaccessible fields cannot leak into a target.
Envelope: causality	Receivers reconstruct ancestry; the controller can replay a feedback cycle.	The two loops are time-aligned and the future target $t{+}k$ is well defined.
Envelope: contract	Gateway validates the round-file field against its receiver schema and access policy.	Type identity fixes the decoder, loss, and aggregation rule per field.
Native body $x_{f(m)}$	Retains operational keys ( rr_advise , failure_feedback , candidates , config ).	Prediction targets stay tied to the exact field paths and receiver arguments used at run time.
Provenance $p$	Evidence is externally inspectable and traceable to its source artifact; speculative state stays isolated until commit.	Only committed prefixes supervise future bundles; rollback traces stay out of the state trajectory.

Table 1: Field-by-field rationale for the structured communication plane. The left column is why a group helps ordinary transmission; the right column is why the same group helps later prediction.

Panel of Figure 1 — Figure 1: A2A-compatible three-layer protocol and deterministic route catalogue. The wrapper adds identity, routing, causality, visibility, and provenance while retaining the original round-file payload_path and receiver argument. Native round-JSON fields are wrapped, routed to existing method arguments, and retained as prediction targets.

Panel of Figure 2 — Figure 2: Predictor data pipeline. The structured gateway records committed packets and token-faithful model traces, aligns four miner-directed target channels by state, and composes them into two destination-specific guidance packets at each direct horizon. Rollback traces are stored separately for RL and error analysis, not inserted into the committed supervised trajectory.

Panel of Figure 3 — Figure 3: Four-level validation of a speculative outcome. The bundle is checked first; the actual miner outputs, blinded evaluator reports, and development-set evidence are checked next. Only an all-pass path atomically commits; any failure discards the scratch transaction and resumes one normal cycle from $S_{t}$ .

Panel of Figure 4 — Figure 4: Three transactional evolution paths. Predictor skills learn from commit/rollback outcomes. Evaluators evolve rubric and meta-rubric libraries from actual outputs only, never seeing the prediction, and their updates install only through the post-decision promotion path.

Method	$n_{\mathrm{valid}}$	$P(\mathrm{sh}{>}1)$	med. AnnRet	med. sh	95% CI (sh)
B1 Alpha101 [ 48 ]	14	0.000	$-0.078$	$-0.396$	$[-0.91,+0.51]$
B2 GP (gplearn) [ 31 ]	500	0.116	$-0.144$	$-0.593$	$[-0.64,-0.52]$
B3 DSO/DSR [ 32 ]	500	0.010	$-0.287$	$-1.648$	$[-1.71,-1.65]$
B4 AlphaGen [ 31 ]	200	0.188	$-0.001$	$-0.006$	$[-0.32,+0.13]$
B5 AlphaMCTS [ 49 ]	500	0.160	$-0.015$	$-0.056$	$[-0.30,+0.06]$
B6 AlphaAgent [ 34 ]	500	0.220	$-0.005$	$-0.021$	$[-0.21,+0.11]$

Table 2: Holdout comparison (CSI 1000, shared 10-group backtest; each method mined on train, filtered to valid on validation; no top-

K

post-selection).

n_{\mathrm{valid}}

is the validation-filtered pool size under a matched budget: every method contributes the valid factors it accumulates in mining order up to a common cap of

500

(B4 AlphaGen reaches only

200

and B1 provides

14

hand-crafted formulas, reported honestly). For OURS this is the first

500

valid factors of the continual run in mining order, not a Sharpe-ranked selection;

499

of them produced a scored holdout backtest (one dropped for missing holdout coverage), and the holdout statistics are computed over those

499

. This early segment is stronger than the full run (full-pool median Sharpe

+0.13

; see Section 4.1 ), so this row should be read with the order-agnostic top-

K

comparison of Table 4 . ∗ B7 is an automated Alpha-GPT approximation. The 95% CI is a bootstrap interval for the median Sharpe (

2000

resamples of each holdout pool); only our system’s interval lies entirely above zero, which bounds within-pool median stability and is not a between-method test (Section 4.2 ). The per-factor rank effect size

\hat{P}(\text{OURS}{>}\text{baseline})

is reported in text; per-factor units within a run are not independent, so we give a descriptive rank statistic rather than a

p

-value (Section 4.1 ).

Selection	med. Sharpe	med. AnnRet	$P(\mathrm{sh}{>}0)$	$P(\mathrm{sh}{>}1)$
dev-top 5	$+0.26$	$+6.6\%$	0.80	0.20
dev-top 10	$+0.48$	$+10.8\%$	0.80	0.40
dev-top 20	$\mathbf{+0.71}$	$\mathbf{+13.3\%}$	0.80	0.45
all 852 (reference)	$-0.23$	$-4.8\%$	0.42	0.13

Table 3: Portfolio-level holdout performance of development-selected portfolios (ranked by test-split Sharpe; holdout never used for selection). The selected set is profitable out of sample; the all-portfolio population median is negative, so the effect is due to development-guided selection.

	$K=50$		$K=100$		$K=200$
Method	med. sh	$P(\mathrm{sh}{>}1)$	med. sh	$P(\mathrm{sh}{>}1)$	med. sh	$P(\mathrm{sh}{>}1)$
B2 GP (gplearn)	$\mathbf{+0.576}$	0.320	$+0.423$	0.270	$-0.369$	0.170
B3 DSO/DSR	$-0.816$	0.080	$-1.199$	0.040	$-1.635$	0.025
B4 AlphaGen	$+0.205$	0.184	$-0.069$	0.186	$-0.006$	0.188
B5 AlphaMCTS	$+0.153$	0.020	$+0.251$	0.090	$+0.236$	0.140
B6 AlphaAgent	$+0.001$	0.180	$-0.033$	0.190	$-0.001$	0.215

Table 4: Same-footing comparison: each method’s top-

K

factors by validation Sharpe, evaluated on holdout. Median holdout Sharpe (and

P(\mathrm{sh}{>}1)

). Best median per column in bold. At

K=50

the genetic-programming baseline’s few strongest factors exceed ours on the median; the ordering reverses and our advantage widens as

K

grows.

Method	valid	test	holdout
B1 Alpha101	$-1.317$	$-0.385$	$-0.396$
B2 GP (gplearn)	$1.520$	$0.677$	$-0.593$
B3 DSO/DSR	$1.482$	$0.673$	$-1.648$
B4 AlphaGen	$1.457$	$0.509$	$-0.006$
B5 AlphaMCTS	$1.838$	$0.857$	$-0.056$
B6 AlphaAgent	$1.531$	$0.455$	$-0.021$

Table 5: Median backtest Sharpe across splits. OURS shows the smallest out-of-sample degradation and is the only method positive on holdout.

Test split
	Sharpe	AnnRet	ExcRet	IC	ICIR	IR
OURS	$+0.467$	$+0.055$	$-0.029$	$+0.016$	$+0.190$	$-0.432$
B4	$+0.509$	$+0.083$	$-0.014$	$+0.031$	$+0.234$	$-0.201$
B5	$+0.857$	$+0.158$	$+0.006$	$+0.042$	$+0.380$	$+0.086$
B2	$+0.677$	$+0.119$	$-0.016$	$+0.040$	$+0.288$	$-0.178$
Holdout split

Table 6: Matched-budget median of six metrics on test and holdout (the same

500

-factor pools as Table 2 ). Our system leads the return/risk-adjusted family on holdout but not the correlation family (IC/ICIR), and trails on test; we report all metrics on both splits. Best holdout value per metric in bold.

Metric	S-off (free-text)	S-full (typed)
Committed-leap hit rate	0.9749	0.9754
Mean committed horizon	2.37	2.47
LLM calls / round	10.80	10.89
Input-repr. alignment (offline)	$0.0547{\pm}0.0053$	$0.0528{\pm}0.0053$
Typed contract / schema check	no	yes
Deterministic replay	no	yes

Table 7: Structured vs. free-text transport ablation (only the predictor’s input representation differs; all else fixed). In this single paired run the two arms show no measurable difference on any predictive-effectiveness metric; we read this as no evidence of an accuracy advantage rather than as proven equivalence (single seed, no variance estimate). The value of structured transport is the correctness and auditability properties of Section 3.2 , which the free-text arm cannot provide at any hit rate.

Control (holdout median Sharpe)	S-off (free-text)	S-full (typed)
Same-round base ( $\leq 1100$ )	$+0.171$	$+0.098$
Top-100 (same-round, dev-ranked)	$+0.864$	$+0.274$
Top-200 (same-round, dev-ranked)	$+0.797$	$+0.221$
Equal-count ( $n{=}1608$ )	$+0.176$	$+0.098$

Table 8: Ablation under nested same-footing controls (median holdout Sharpe of in-database factors, real frozen-holdout backtests). The arms are close on the full same-round and equal-count pools; on development-ranked top-

K

the free-text arm is clearly ahead. Structured transport never leads on factor quality at matched rounds, matched top-

K

, or matched count.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Quantity	S-off (free-text)	S-full (typed)
Per-round wall-clock, median (IQR)	$385$ s $[363,414]$	$\mathbf{353}$ s $[335,373]$
Portfolio structure-gate pass	$0.985$	$\mathbf{0.991}$
Portfolio composite Sharpe, median	$2.24$	$\mathbf{2.28}$
Single-factor candidate pass	$\mathbf{0.246}$	$0.128$

Appendix

Table 9: Operational side effects of structured transport (same run logs, same period). The per-round wall-clock is observational (same machine and period, but not a controlled timing experiment); success-rate effects are small and mixed and are reported in both directions.

Multi-Agent Communication

Multi-agent communication studies how autonomous agents exchange information to coordinate decisions, share knowledge, and act together under constraints.

See all 380 papers

Jun 28, 2026

AI Trading's Alpha Singularity: Emergent Market Reasoning through Agent-to-Agent Self-Evolution

Automated alpha mining holds the scoring function fixed and varies the search algorithm over it. A search that converges against a fixed scorer overfits whatever the scorer cannot penalize, a primary cause of the out-of-sample generalization gap. We treat the scoring function as a search artifact alongside the alpha factors and study what conditions make this joint search admissible. Sealed Joint Search (SJS) is a framework: a set of structural conditions on information flow in an autonomous-discovery system that prevent joint search from collapsing into self-confirmation while keeping the evaluator sealed. Conditions cover role decomposition, typed inter-role communication, provenance-sealed reads, versioned stores, and substrate-local promotion. Agora tests SJS empirically: five LLM agent classes communicate via three channels, evolving eight skill libraries, with alpha libraries built on AlphaGen operators. Three evaluators write reports aggregated into one brief, carrying forward disagreement instead of voting. We run Agora for 100 rounds on CSI 1000 and evaluate on a 91-day 2026 holdout sealed from all LLM inputs. Agora achieves holdout Sharpe +1.87; best baseline +1.334 at favorable seed and -0.755 cross-seed mean. Pre-loading Agora's two metrics into a frozen-library ablation recovers only +0.40 of the +2.25 Sharpe gap, and adding PPO without library evolution worsens the gap. The two metrics emerge rather than being designed. Caveats: single-seed run, short-side concentrated signal, intended for long-short.

Yuqi Li, Siyuan Liu, Bingjun Liu

Panda AI

Finance Emergent Behavior in Multi-Agent Systems Multi-Agent Reasoning Automated Algorithm Discovery Algorithmic Trading Self-Evolving Agents+1

Ask PDF

Aug 4, 2026

AgonAlpha: Autonomous Alpha Discovery via Prompt Economy and Scalable Agentic Search

Language models can propose many plausible trading factors, but an autonomous research system must also allocate its evaluation budget, verify its own evidence, and preserve how each candidate was produced. We present AgonAlpha, an architecture that searches over frozen research artifacts---hypotheses, executable expressions, platform evidence, rationales, and review status---rather than formulas alone. To our knowledge, AgonAlpha is the first alpha-mining system to combine verified artifact search, a fresh-context adversarial reviewer with re-execution and veto authority, and pending-aware parallel budget allocation, together with a complete public evidence trail. Independent deployments on WorldQuant BRAIN produced SPECTACULAR-grade alphas across five users and six model backends, with Fitness reaching 9.50 and Sharpe reaching 3.48, while retaining prompt-to-expression provenance for every submission.

Weicheng Ye, Youran Sun, Xingyu Ren +3

Department of Physics, The Chinese University of Hong Kong, Hong Kong SAR, China · Department of Mathematics, University of Maryland, College Park, MD, USA · Department of Computer Science, University of Maryland, College Park, MD, USA

Quantitative Finance Automated Algorithm Discovery Agentic Search Finance Autonomous Research Agents Alpha Factor Discovery

Ask PDF

Sep 7, 2026

Scores Alone Do Not Prove Discovery: The Discovery Certification Protocol for Auditing AI Research Agents

AI research agents combine public information and experimental feedback to produce measurable results. The Discovery Certification Protocol (DCP) turns an outcome claim into an executable audit under a registered model, information boundary, and budget. Gate 1 validates useful improvement. Gate 2 tests recovery by matched agents given the starting information and observed Web content, with run history and new measurements withheld. Core requires adequate registered controls, zero recoveries, and a finite-sample recovery bound. Optional Gate 3 compares truthful and neutral feedback from a shared checkpoint; Evidence adds a supported effect and a null-policy equivalence check. Controlled SQLite and virtual catalyst audits pass both decision kernels. On real-data response surfaces, Yacht and Ionosphere pass the Core kernel after zero recoveries in 96 attempts, with an upper bound of 0.0468. Each target combines ten observed utilities and six predictions into a 16-entry data product. Yacht scores 0.7677 on reconstruction of all 32 switch effects, with utility-prediction MAE 0.0315 on its six unmeasured configurations. Fresh truthful continuations recover the target level in 9/30 and 16/30 trials, respectively, separating achieved utility from process repeatability. A deterministic verifier reproduces these local decisions from frozen records.

Jingjie Ning, Shanshan Zhong, Xiaochuan Li +1

School of Computer Science, Carnegie Mellon University

AI Agents for Scientific Discovery Algorithmic Auditing AI Agent Auditing AI Agent Evaluation

Ask PDF

Abstract

Multi-Agent Communication

Explore similar work

AI Trading's Alpha Singularity: Emergent Market Reasoning through Agent-to-Agent Self-Evolution

AgonAlpha: Autonomous Alpha Discovery via Prompt Economy and Scalable Agentic Search

Scores Alone Do Not Prove Discovery: The Discovery Certification Protocol for Auditing AI Research Agents