Agentic trading field notes

Our pre-alpha record isolates a selection effect at a rendered rank boundary. A candidate-list audit should precede claims about model market preferences.

Historical agent leverage followed risk settings and explicit user numbers. Risk adaptation needs evaluation at the order boundary.

Across 6,400 historical positions, median leverage stayed at 5x while volatility and liquidation rates changed sharply.

Historical positions often reached favorable prices and then closed negative. Capture analysis isolates an exit question from a direction claim.

A historical stop-target replay improved paired returns, while a trigger-quota failure exposed why protection needs execution verification.

The historical fleet’s unfavorable return and signal results narrow the next experiment to measurable mechanisms and held-out prediction.

The historical fleet mixed mostly paper execution with a smaller real-capital book. Margin, funding and fill assumptions change what its results establish.

Specific numeric instructions were associated with better outcomes in a bounded real-capital event. The next question is mandate design, with confounding preserved.

The historical fleet’s builder fees changed across three eras. Common-rate restatement separates cost assumptions from observed trading behavior.

The two historical fleets shared configuration concepts but resolved text-setting conflicts differently. Transfer must name the component and environment.

Read a concentrated historical buy event without converting shared timing into proof of agent coordination.

Interpret the 92.9% trade share in two-sided windows using the correct denominator and time window.

Keep the 3,878 historical cascades attached to the ten-vault, ten-minute rule before comparing populations.

Interpret the 9.5-second median gap in a large sell cascade without inferring a reaction clock.

Separate the historical 92% withdrawal figure from return, residual value and cash-flow timing.

Use the approximate 7.5 million invocations and 300,000 actions without treating inactivity as failure.

Keep 3,505 funded vaults and 3,454 active vaults distinct when reporting participation.

Treat changing token survival as part of the historical market design rather than a background price feature.

Read the historical liquidation count separately from its share of gross loss.

Separate fills, positions, agents and market days before interpreting a mixed research fleet.

Use the withdrawn HIP-3 lag claim to specify what a point-in-time audit must preserve.

Keep the August 15 freeze separate from the June 8 to July 26 historical P&L window.

Interpret roughly 70 billion tokens across 7.5 million invocations without inventing context or serving limits.

Read repeated pre-launch repair as development evidence while preserving the trials behind a chosen prompt.

Distinguish published aggregates, figures and extracts from a raw-data reproduction of historical results.

A worked lifecycle for persistent agentic trading: separate mandate, market state, and settled history before the next decision.

Understand account ownership, delegated signing, and revocation through a concrete agent-wallet review scenario.

Compare trading models while holding the surrounding tools, policy, and account record fixed.

A provenance and timing framework for combining social signals, event odds, prices, and portfolio state.

Separate research and chart tasks without letting worker agreement masquerade as independent evidence.

Use periodic reviews for coverage and explicit triggers for changing conditions, with one reconciled decision state.

Review a trading decision through concrete questions about evidence, instructions, interpretation, and a proposed strategy revision.

A final-action test explains how deterministic policy differs from asking a model to respect a limit.

A revision workflow converts vague trading requests into explicit conditions while preserving the owner's intent and historical decisions.

A research-to-product loop turns observed failures into fixtures, measured repairs, and separately verified platform improvements.

A prelaunch review that tests whether an owner can explain entries, waiting, and existing exposure.

A diagnosis worksheet for account state and delegated signing authority.

A filled-volume worksheet separating builder fees, venue fees, funding, and price impact.

An evidence review for strategy illustrations, public positions, and historical performance.

A workflow for testing a catalyst thesis alongside completed-candle evidence.

A model review based on timely valid decisions rather than reputation alone.

A practical owner review covering active instructions, exposure, and completed decisions.

How to distinguish a market wake-up from an authorized entry condition.

A portability review for symbols, leverage metadata, and exposure definitions.

An owner audit for origin, timestamps, repetition, and unsupported claims.

A thesis review that distinguishes event probability from a perpetual-price forecast.

A conversation pattern for testing contradictions without demanding a trade.

A decision tree separating waiting, missing turns, rejected actions, and unfilled orders.

An architecture review that asks what a public claim actually establishes.

How to separate product feedback, observed trading outcomes, and public endorsements.

A proposed calibration study separates token probabilities, spoken confidence, market forecasts and decisions under one saved evaluation set.

A proposed randomized render experiment tests how visibility and position affect asset selection independently of market information.

A proposed sizing study tests whether volatility-aware exposure improves loss control under unchanged trade selection.

A proposed failure-injection protocol defines when a trading position counts as protected and tests retries, quotas and recovery.

A proposed experiment measures whether prediction-market inputs add information after timestamp, contract and liquidity checks.

A proposed execution study measures fill, funding and margin differences before using paper outcomes to assess an agent.

A proposed ablation distinguishes better information, changed decisions and net economic value when specialist research workers are added.

A proposed controlled experiment compares revisit timing while preserving information access, risk policy and agent exposure.

A proposed online evaluation tests regime labels, detection delay and action value using only information available at each turn.

A proposed corpus tests mandate, input, execution and recovery behavior so improvements can be evaluated across DXAP harness releases.

A proposed adapter experiment tests economic meaning across synthetic venue rules.

A proposed paired study separates visual representation from additional market information.

A randomized delay experiment measures the effect of information arrival on fixed decision opportunities.

A proposed allocation study measures evaluated opportunity coverage under a fixed total budget.

A proposed queue experiment prioritizes exposure management under bursts of agent wakeups.

A proposed protocol records owner interventions and separates assisted outcomes from autonomous behavior.

A proposed holdout study tests impact estimates with matched order-size and market-state controls.

A proposed two-stage forecast separates event occurrence from conditional market response.

A proposed shared-state experiment distinguishes correlated reaction, communication and price feedback.

A proposed planning study varies tool reliability and tests how agents value information before acting.

A proposed rehearsal protocol tests the complete planning path while withholding submission capability.

A proposed synthetic calendar suite tests decisions around open, close and scheduled market interruptions.

A proposed objective comparison evaluates interactions among simultaneous positions.

A proposed release protocol freezes questions, comparisons and decision rules prospectively.

A proposed research decision record distinguishes a tested failure from an unresolved or untested mechanism.

A proposed label design separates mandate compliance, execution quality and market outcomes before trading traces enter an agent evaluation.

A paired review experiment checks whether a profitable result changes how reviewers judge the same trading-agent decision.

A proposed two-axis evaluation tests whether better trading-agent process scores translate into decision economics under realistic costs.

A provenance design for distinguishing current-policy trading trajectories from historical traces before evaluating a learning proposal.

A worked example shows how a calibrated forecast can produce different trading actions under different cost and risk policies.

A four-arm trading-agent experiment crosses assigned context length with instruction placement to distinguish salience changes from action improvements.

A proposed component comparison distinguishes a stronger trading model from improvements to instructions, state or tools.

A proposed shift audit separates changing markets, user mandates and agent infrastructure before judging learning transfer.

A proposed replay audit identifies which alternative trading actions the saved data and simulator can actually evaluate.

A proposed marginal-value experiment tests when research or chart subagents improve trading decisions enough to justify their cost and delay.

Trace imitation can transfer a teacher’s unsupported rules along with its useful decisions.

A saved action and token log probabilities may still leave the behavior-policy denominator unknown.

A representation test needs a fact-level equivalence check before model scores are compared.

Per-day standardization can erase an economically important difference between quiet and volatile sessions.

A later portfolio outcome needs a declared path back to the decisions that could have affected it.

Before changing the learner, check whether opposite labels refer to the same mandate and information.

More model agreement cannot repair a stale or ambiguous observation.

Several models can repeat one source error and still look unanimous.

Severity weights encode an objective and should remain visible beside ordinary error counts.

A candidate can improve on its new task while losing an older instruction boundary.

A verifier should discriminate substantive compliance from text that merely resembles a passing answer.

Repeatedly choosing checkpoints on a holdout turns that holdout into a development signal.

A failed venue submission should teach the stage responsible for the failure.

A learned router earns its complexity by choosing useful information under the same available tools.

Clearer explanations deserve their own evidence rather than a trading-performance claim.

A proposed participation experiment connects user incentives to usable engineering evidence.

An illustrative account ledger separates recognition, fee-share cash and economic trading results.

A proposed feedback rubric values reproducibility and coverage without rewarding trading churn.

A checked referral example separates the builder fee base from trading volume and program points.

A marginal-reward example explains the difference between recognition and repeatable loss rewards.

A proposed cohort scorecard counts evidence, understanding and retained agency alongside activity.

Separate community growth from the evidence that helps improve a trading-agent harness.

A proposed review mechanism credits reproducible negative evidence and informative null results.

A proposed coverage ledger makes invite constraints explicit while seeking different user experiences.

A proposed staged testing design separates research consent, feature exposure and trading authority.

A proposed contribution record names useful work while keeping trading volume outside research standing.

A proposed support and research pathway includes cautious owners and informative no-trade observations.

A randomized offer study separates invitation effects from the enthusiasm of participants who claim a reward.

A support experiment separates immediate comprehension from voluntary review retention over four weeks.

A network-aware study distinguishes recruitment selection from the effect of a referral invitation.

A capacity comparison includes participant support, duplicate triage and curation before ranking evidence yield.

A semantic review contract separates inspected decisions from open tabs, visualizer time and autonomous agent operation.

A prospective sampling audit checks whether new rewards alter the cases available for future training and evaluation.

A proposed support study measures diagnostic visibility separately from the number of complaints.

A contrastive support question identifies the missing condition without asking the agent to trade more.

Measure whether an unanswered setup question produces different saved strategies, then improve the question.

A recovery study tests whether users can identify saved settings after an interrupted approval.

Define closure evidence by the reported failure rather than by the last reassuring reply.

A correction note connects a reported misunderstanding, the changed behavior and the evidence readers can check.

A proposed diagnostic handoff locates the earliest reproducible mismatch before assigning a repair owner.

A proposed versioned taxonomy tests whether support and engineering classify the same event consistently.

A proposed explanation study separates generation latency, reading time and comprehension transfer while holding execution fixed.

A proposed review sequence measures deliberate steering while preserving the separate approval step.

A proposed two-track triage contract preserves urgent investigation while measuring recurring comprehension problems.

A proposed transfer test checks whether a resolved support issue stays resolved for the next owner.

A proposed feedback loop connects user participation to reproducible fixes and evaluated releases.

A proposed evidence bundle preserves the decision, configuration and timing behind each feedback label.

A proposed coverage audit distinguishes new failure contexts from repeated observations of one event.

A proposed labeling contract identifies why an agent waited and what evidence could justify a different decision.

A proposed join contract preserves one intent across approvals, requests, fills and feedback.

A proposed adjudication process separates factual disagreement from different owner objectives.

A proposed permission receipt separates research use, publication, and trading authority.

A proposed transformation contract keeps the relationships needed to reproduce an owner correction.

A proposed grouping rule preserves support demand without inflating independent evidence.

A proposed rubric turns precise owner feedback into bounded labels instead of assumed ground truth.

A proposed lineage record prevents old failures and later fixes from sharing an invented context.

A proposed dependency graph makes withdrawal concrete across source records and derived datasets.

A proposed preference dataset turns explicit owner choices into response targets while withholding future market outcomes.

A proposed supervised-example contract preserves the user problem, relevant state and reviewed answer instead of training on raw tickets.

A proposed pairwise dataset holds facts and actions fixed while owners compare presentation choices.

A proposed training curriculum uses owner-confirmed clarifications to teach the objective an action must serve.

A proposed feedback model keeps owner satisfaction, mandate compliance and economic outcomes independently inspectable.

A proposed problem taxonomy routes curated user contributions to a specific training objective and held-out test.

A change-location test separates saved context, harness revisions, and proposed model training.

A proposed case-admission process turns user feedback into independent release evidence.

A proposed shadow study compares feedback-aware candidates while preserving the live authority path.

A proposed rollback map ties adaptation failures to the smallest reversible release artifact.

A proposed prospective cohort study tests whether feedback-driven fixes generalize beyond their contributors.

A proposed semantic contract review prevents feedback-driven releases from silently changing owner authority.

A proposed reconstruction test measures whether retained decisions make competing repairs distinguishable.

A stage-specific acceptance table makes future adaptation concrete without treating a roadmap as a shipped learning system.

A transformed fixture tests whether a common repair follows each owner’s cap rather than copying another owner’s action.

A donor-recipient experiment tests whether evidence from new participants helps existing owners on untouched cases.

A proposed marginal curation curve allocates research capacity by held-out benefit per added unit of cost.

A proposed independent-observation audit tests whether a candidate improves or simply changes which failures become visible.

A worked conflict case shows how authenticated controls and strategy text should resolve into one inspectable trading mandate.

Why a concrete trading instruction needs quantity definitions, conditions and expiry before its behavior can be evaluated.

A worked news-window example separates expiry of permission, pending execution and ongoing position management.

A race between a strategy update and an unfinished turn shows why every action needs one attributable mandate version.

Separating market facts, interpretation and rationale makes a trading trace easier to evaluate and prevents earlier reasoning becoming policy.

DXRG’s fee-placement result shows a large change in trace citation under fixed facts, with clear limits on economic interpretation.

DXRG’s rule-fabrication finding supports explicit authority labels and compound-intervention attribution, with a targeted regression design.

A news excerpt can describe someone else’s imperative without authorizing an agent to act; source roles make the difference inspectable.

A quantity conversion example shows how schema checks, mandate policy and final-payload validation answer different questions.

An explicit abstention record separates a thesis-driven wait, insufficient data and failed execution in trading-agent evaluation.

Separate total absolute position notional from signed directional exposure with a two-position mandate fixture.

A cash-flow-aware peak-equity fixture makes a drawdown restriction reproducible.

Use a UTC interval fixture to distinguish loss-budget boundaries from the owner’s displayed calendar day.

An illustrative trend-versus-reversion conflict makes precedence visible before execution.

An interval fixture distinguishes a valid shared budget from a mandate that requires owner resolution.

A forged amendment fixture separates a specialist’s information from permission to change a mandate.

A three-field proposal fixture reveals changes that a conversational summary can hide.

A conflicting two-proposal fixture prevents accidental combination of separately reviewed changes.

An ETH-removal fixture separates entry eligibility from authority to manage an outstanding holding.

A three-snapshot fixture defines when a strategy has enough information to propose an entry.

A held-position fixture turns an ambiguous pause command into explicit permitted actions.

A ceiling-and-patience fixture distinguishes a machine-checkable restriction from guidance for model judgment.

A buy-and-sell price fixture compiles execution tolerance without hiding fees inside the same number.

A spot-and-perpetual fixture separates directional offsets from gross family exposure.

A generated-case protocol checks authorization properties across many typed strategy inputs.

A proposed fixture for delayed fills, fees, and the portfolio version used to approve the next trading action.

How to test a strategy update when old instructions remain easy to retrieve and a model call is already running.

A source can be correctly identified while remaining unable to authorize a trade or change an agent mandate.

A dependency-based fixture for headlines that remain retrievable after the event they describe has resolved.

A proposed procedure for late research responses whose original request context has already been superseded.

A proposed schedule-invariance test for agents that confuse repeated observations with elapsed market time.

A relevance fixture that separates semantic similarity from current decision eligibility and downstream usefulness.

How to test a model that insists an order filled when the authoritative venue state says it remains open.

A proposed test for subagent results that accidentally import another account’s limits, positions, or mandate.

A proposed representation comparison that preserves modern context capacity while testing what current decisions actually need.

Choose the replay boundary before choosing how to store portfolio state.

Preserve both corrected history and the information available at the original decision.

Account state needs explicit dependencies when timestamps cannot order its changes.

Deletion needs to cover derived memories as well as the original item.

Separate a fact’s verification status from the model’s confidence in its interpretation.

Resolve a source conflict at the claim and scope level before selecting a current fact.

Measure the decision-relevant facts a summary drops, changes, or merges.

Compare memory retrieval by required evidence delivered inside the same rendered budget.

Retain exact operational numbers and compress the surrounding explanation.

Use digests to compare declared state representations, with schema and semantics attached.

A trade transition should invalidate every derived view that depends on its effects.

Test decisions across plausible clock offsets instead of treating timestamp comparisons as exact.

Keep counterfactual and stress-case memories from becoming factual history.

Choose retention by the incidents you must reconstruct and the information each record exposes.

A readable state view should preserve the operational meaning of its typed source.

A proposed schema that joins agent intent, policy checks, execution and reconciliation without treating a rationale as an execution receipt.

How to define the replay boundary, preserve captured context and avoid silently evaluating a different information set.

A procedure for separating tool-use skill from information access in trading-agent model evaluations.

Why completed-turn scores need scheduled-turn denominators and explicit missing-status reporting.

How to test explanation consistency against recorded inputs and actions while keeping causal claims bounded.

A coverage-aware method for comparing no-trade behavior with conditional decision quality.

Separate executable constraints from semantic interpretation and measure disagreement before using an LLM grader.

How to describe compound fixes, isolate components and retain the historical scope of measured trace improvements.

A stage-by-stage diagnosis of trading losses that distinguishes forecasts, sizing, policy and venue execution.

Why a model label alone cannot identify the configuration that produced a trading decision or evaluation score.

A case-family audit separates copied examples, semantic siblings and genuinely new trading decisions.

Resolve model identity at request time and quarantine ambiguous alias changes before comparing agent behavior.

A section receipt and boundary-marker fixture identify missing instructions before an agent trace receives a behavioral grade.

A staged taxonomy distinguishes serialization, schema, domain and authority failures in trading-agent tool calls.

Distinguish deadline failure from an unobserved completion time and preserve timed-out cases in quality reporting.

Compare latency distributions and deadline misses rather than relying on average response speed.

Define separate objectives for usable proposals, submitted actions and reconciled outcomes.

A three-identity audit prevents transport retries from becoming additional trading decisions.

A weighted review design finds rare severe traces while preserving a population estimate.

Report the agreement table alongside kappa so rare defects remain visible.

A severity-aware trace audit avoids treating common harmless defects as equivalent to rare unauthorized actions.

A dependency audit separates the first corrupted result from later calls that merely propagate it.

A compatibility matrix checks old and new consumers against changed trading tool responses.

A two-stage replay fixture distinguishes runtime instability from stochastic changes in model proposals.

Evaluate how agents respond when needed evidence is unavailable instead of grading only the final narrative.

A five-row worked example for checking probability scores, sample alignment, and comparison baselines.

An illustrative comparison shows how a confident miss can outweigh several confident hits.

A worked example shows why bin counts and boundaries belong beside a calibration chart.

Two constructed forecasters show why probability meaning and event ranking require separate checks.

A fixed reference probability prevents an impressive accuracy number from becoming a misleading skill claim.

A timestamp example explains why chronological rows alone do not prevent information leakage.

A three-window example separates model updating, probability calibration, and honest aggregate reporting.

A rolling-horizon fixture shows how many saved rows can represent far fewer distinct market intervals.

A simple chance calculation explains why a selected winner needs an untouched evaluation.

A zero-event example shows why a small observed rate can still leave substantial uncertainty.

Measure event frequency separately for each asset before interpreting a pooled forecast score.

Use a logistic diagnostic to separate shifted probabilities from excessive confidence.

Make candidate probabilities coherent while preserving the mass outside the candidate set.

Keep scoring precision separate from the probability shown to a reader.

Compare identical questions row by row before estimating uncertainty.

Evaluate confidence-based forecast selection across retained fractions.

Audit empirical coverage and set size when market distributions change.

Separate price-source ambiguity from native LLM probability error.

Use event counts to expose false alarms hidden by overall accuracy.

Evaluate an internal review screen using declared asymmetric error costs.

Choose whether the research question concerns a typical agent or a typical forecast.

Quantify the score floor implied by an explicit observation-noise model.

Keep unresolved future outcomes visible in the denominator.

Use lagged volatility strata to diagnose where probability error changes.

Treat each horizon as a distinct target before comparing model quality.

Calculate mixed maker and taker fees without mistaking an order type for an execution classification.

Measure the bid-ask component of execution cost with a declared reference price and quantity.

Walk book levels to calculate an illustrative execution estimate while preserving its snapshot limitations.

A worked partial-fill ledger for autonomous agents, including duplicates, residual quantity, and cancellation.

Trace a replacement instruction through a late fill so residual exposure remains explicit.

A recovery procedure that separates transport uncertainty from a new trading decision.

Keep accepted requests, live orders, executions, and final reconciliation distinct in agent traces.

Check closing intent against position mode, live quantity, and venue-specific reduce-only semantics.

Normalize executable prices and quantities without silently widening an agent instruction.

A signed funding ledger that distinguishes displayed rates, estimated charges, and settled account entries.

Time-in-force is part of trading authorization. A worked liquidity case separates immediate execution, resting exposure and unsupported FOK intent.

A moving quote can invalidate a maker-only order. Preserve the liquidity instruction and measure rejection handling separately from fill quality.

Translate an urgent trading instruction into a signed price boundary. A worked buy case distinguishes execution urgency from unlimited spending.

A modification can change waiting economics. Record amendment lineage and treat priority retention as venue-specific behavior requiring explicit evidence.

Measure favorable execution against a declared pretrade reference. A worked ledger separates price improvement, fees and subsequent market movement.

A canceled order can reflect owner action, margin changes or venue restrictions. Preserve reasons and reconcile economic effects before closing the agent record.

After a broad cancellation, verify orders, positions and collateral separately. A worked account case defines a narrower, reviewable stopping condition.

Nonce tracking follows the signing wallet. A concurrent-worker fixture defines atomic allocation and separates replay protection from execution sequencing.

Reopening a stream restores transport. A gap fixture combines historical fills, stable economic identity and a coverage-aware account reconciliation.

Two strategies on one address can interact without a trade receipt. A worked book case tracks the resting-order cancellation and the surviving aggressive intent.

Collateral sharing changes the meaning of available capacity. A simplified two-position case defines account abstraction and margin-mode fields for agent decisions.

A fee ledger should preserve native amounts before conversion. A two-rate example separates execution cost from later changes in the fee asset’s price.

An unfilled order can consume a future exposure budget. A simultaneous-worker example defines worst-case reservations without assuming offsetting orders execute together.

A partial close changes both accounting categories. A simplified position case separates closed trade profit, remaining marks and independent cash movements.

A chain identifier and an API observation establish different facts. Define the event evidence required before dependent agent actions and cross-system reports advance.

A worked timestamp example for reconstructing what a market-data consumer knew at a decision boundary.

A small quote-joining example that exposes future matches, stale matches, and ambiguous ties.

A minute-bar example for separating bucket labels from feature availability.

A gap-classification workflow that keeps no-trade periods separate from collection failures.

A token-decimals example with exact arithmetic, metadata provenance, and failure handling.

How to build instrument keys that distinguish venues, pairs, chains, and metadata changes.

A three-level example showing how depth differs from a best quote and why aggregation matters.

A reconstruction checklist for negative spreads, mismatched snapshots, and venue state.

A quiet-market example for separating old prices from broken subscriptions.

A two-regime example showing how feed activity changes an average even without a price change.

Build an eligibility ledger that preserves listings, restrictions, and removals before comparing agent selection.

A four-market fixture shows how a current catalogue can reverse a historical market-selection summary.

Use a feed-specific continuity check and a recovery receipt to identify when reconstructed market state becomes usable again.

A sequence-boundary fixture and absolute-size example catch two common errors in reconstructed depth.

An illustrative rate change shows why agent features and account funding cash flows need different data records.

A three-price fixture shows how an unnamed price changes portfolio valuation and market interpretation.

A revised-release fixture preserves the earlier agent input while allowing a later corrected account of the event.

A synthetic deadline amendment shows why a market probability needs the question and rules version that gave it meaning.

A post-reference fixture keeps social reach visible while preventing copied claims from becoming false corroboration.

A release-time conflict fixture treats disagreement as data and tests whether it can change feature admission.

An integer-width fixture and metadata binding check prevent valid JSON from routing an action to the wrong asset.

A versioned quote fixture separates normal observation lag from a true adapter disagreement.

A deletion fixture shows why prompt summaries and hashes cannot substitute for missing source payloads.

An inclusive-boundary fixture preserves equal-time events while preventing repeated rows from inflating a market feature.

A two-rate fixture separates native-currency performance from reporting-currency movement.

A transition-system formulation explains why an identical model and prompt can still describe different trading agents.

Historical rule fabrication suggests a theory of authority promotion in persistent agents, with a bounded reading of the 57% to 3% intervention.

The historical render boundary makes strategy attribution depend on the opportunity exposure process.

A four-layer argument separates expressed certainty, calibrated event probability, economic assumptions and execution permission.

A fleet-level theory explains how common observations can dominate mandate diversity without agents communicating.

A persistent-agent theory separates the number of reasoning turns from the number of independently validated observations.

Prompt injection becomes a trading-agent alignment problem when source content can redefine what an authorized account should do.

A future research hypothesis separates common-shock convergence from market-mediated self-excitation and crowding.

A delegation acceptance test separates a useful trading service from an account that happened to make money.

Outcome-blinded conditional review tests whether an agent followed the decision the owner authorized.

A constrained evaluation design keeps preferences, restrictions, execution and economics visible instead of hiding them in one score.

An authority-preserving regime test distinguishes changing tactics from an agent quietly replacing the owner’s objective.

Conditional variance and directional probability answer different questions. Their separation gives agent research a more productive target.

An order-flow impact relationship helps define execution research, provided observed same-interval events stay separate from future information.

Seconds, hours, days and months need different targets, information sets and decision mechanisms. A horizon map makes those differences inspectable.

DXRG’s continuous record rules out easy claims about its historical fleets. Its strongest use is a precise domain of failure, not a universal impossibility theorem.

Gu, Kelly and Xiu show why nonlinear interactions improve expected-return measurement and why the horizon matters.

Kelly and Xiu’s survey explains why trading costs and persistent signals change the objective for financial machine learning.

The April 2026 paper separates predictable equity clusters from weak ones, with different treatment of cross-sectional and regime evidence.

A four-question translation audit connects academic predictability to the full decision system without importing a paper’s performance claim.

Our 2026–2031 scenario makes a research agent broadly accessible, with adoption measured by usable persistence rather than model subscriptions.

A fixed-evidence experiment shows how judgment shifts toward selecting objectives and resolving conflicts in a trading mandate.

A paired-mandate experiment distinguishes shared model behavior from the different objectives of the owners using it.

The Grossman–Stiglitz information-cost mechanism gives our agent adoption scenario a falsifiable market-efficiency prediction.

Our competitive scenario identifies the hours-to-days research businesses exposed to cheap public-signal agents and a test that could disprove it.

An advantage decomposition explains why cheap research can threaten some quant businesses while strengthening structurally advantaged incumbents.

A fragmentation experiment distinguishes permission to enter from the ability to trade economically across pools and chains.

A claim-to-redemption map identifies the new agent tasks created when asset transfer and conditional settlement become programmable.

A conditional timeline separates cheap experiment production from reliable discovery and capital allocation.

The timing of bounded execution differs from the timing of responsibility for capital and market stress.

A task-based timeline for software, data and ML engineering explains why implementation speed and production ownership diverge.

Integrated task bundles explain how smaller teams could compete and why complete role replacement has no single deadline.

A protocol comparison explains the research questions answered by stock simulations, real-capital deployments and continuous agent records.

A dimensional account separates funded agents, runtime tokens, fleet turns, archive bytes and representation rows.

Specific retractions and development audits show how rejected claims improve the next experiment.

The research-to-product case rests on mandate, policy and outcome mechanics rather than a list of AI features.

What a 65,536-cell market representation and a 256-cell selection teach us about building a microstructure transformer.

An 87.26% development accuracy becomes a sharper research result when class prevalence and threshold support are reported.

A development audit explains why a shared random seed and a held-out label leave two separate independence questions unresolved.

Why measuring future mean price from the input opening can mix already observed movement with the outcome a trading decision faces.