Observable Transaction Admissibility in Ethereum Graphs
Story Persistence, Thesis Convergence, Online Pattern Learning, and Price-Movement Hypotheses
V1 Theory Paper - not an empirical performance claim
Prepared for FutCo LLC / E3D.ai
May 2026
Version 1.1 — Implementation Update, May 2026. Adds §3.3, §7, §10, and Appendix C.
We propose Observable Transaction Admissibility (OTA), a theory of structural measurement for Ethereum transaction graphs. OTA extends price-path admissibility into the on-chain domain by defining stories as first-order, time-stamped observations of graph structure and theses as second-order, directional hypotheses generated from converging stories. The framework is designed to formalize the current E3D implementation while specifying testable predictions for future empirical validation. We define a composite transaction-integrity score T_t as a weighted combination of graph integrity, flow integrity, entity integrity, temporal integrity, and market-alignment integrity. We then define story admissibility A^S_t and thesis admissibility A^Theta_t as hysteretic state variables over these measurements. The central hypothesis is not that any single transaction pattern predicts price. Rather, we propose that price-relevant information emerges when heterogeneous stories persist, overlap in counterparties, and converge into theses with rising conviction, explicit entry conditions, and falsifiable invalidation signals. We further isolate thesis convergence as a proposed incremental structural term, allowing the base transaction statistic to be tested against an enhanced statistic that explicitly includes higher-order story aggregation. We distinguish the implemented E3D system - story generation, story references, graph evidence, group mode, thesis cards, conviction, status, candidate pipelines, and decision workflows - from the proposed theoretical layer, which adds formal admissibility thresholds, persistence spectra, overlap metrics, null models, online learning, and price-label testing. This paper intentionally reports no backtested performance. It specifies the measurement theory, implementation gap, and empirical protocol required to test whether E3D-style structural intelligence can improve price-movement prediction beyond price-only baselines.
Keywords: Ethereum, transaction graphs, E3D, stories, theses, admissibility, on-chain intelligence, graph persistence, price prediction, online learning
On-chain markets expose a level of behavioral observability that traditional markets do not. Every Ethereum transfer, approval, liquidity action, swap, contract call, and counterparty relation can be placed inside a transaction graph. Yet the raw graph is too large for direct human interpretation. E3D addresses this compression problem by turning detected on-chain patterns into readable stories and aggregating convergent stories into theses.
The E3D documentation defines stories as the narrative layer of the platform: lightweight, crawlable summaries of what happened on-chain, who was involved, and why it matters. Stories answer 'what happened?'; evidence answers 'how do I verify it?'; and the graph answers 'what does the structure look like?' The thesis documentation defines theses as structured hypotheses built from converging stories. If stories are observations, theses are the analyst memo explaining what the observations mean.
This paper proposes a formal measurement theory for that implemented product design. The purpose is not to retrofit fake empirical results. The purpose is to define the measurable objects, specify the equations, identify where E3D already implements the theory, describe where implementation can be enhanced, and state testable predictions linking story and thesis structure to subsequent price movement.
We use language similar to structural admissibility research in price trajectories, but we shift the observable from price paths to transaction graphs. In the price-path setting, admissibility asks whether a market trajectory is structurally coherent. In the E3D setting, admissibility asks whether the observed on-chain transaction structure is coherent enough to support a story, whether several stories converge enough to support a thesis, and whether that thesis is sufficiently persistent and falsifiable to become price-relevant.
Story admissibility. We define stories as first-order graph-observation objects and specify a story-integrity score measuring entities, structure, impact, freshness, confidence, and evidence support.
Thesis convergence. We define theses as second-order directional objects formed from overlapping stories with shared tokens, wallets, contracts, counterparties, venues, time windows, or transaction motifs.
Persistence spectrum. We specify story and thesis persistence tests across multiple horizons, replacing isolated alerts with measurable temporal structure.
Price-movement hypotheses. We define the conditions under which story and thesis states are expected to contain incremental information for forward returns, volatility, drawdowns, liquidity, or regime shifts.
Online learning. We propose a feedback loop in which confirmed, invalidated, expired, and manually annotated theses update story weights, overlap weights, and conviction calibration over time.
Implementation alignment. We map the proposed theory to the OTA v1 implementation update, distinguishing already-materialized measurement instrumentation from remaining empirical validation requirements.
This is a v1 theory paper. We do not claim that E3D stories or theses have already been statistically proven to forecast returns. We propose a measurement system and an empirical protocol for testing that claim.
We do not claim that a single story is a trade instruction. The E3D thesis-use workflow explicitly treats theses as decision support: users should verify contributing stories, read risk factors, evaluate entry and invalidation signals, and monitor conviction over time.
We do not claim that price is fully determined by on-chain structure. Price is affected by liquidity, macro conditions, exchange flows, narratives, derivatives, regulatory events, and exogenous shocks. The claim is narrower: persistent and convergent on-chain structure may carry incremental information that price-only models fail to represent.
In the implemented E3D documentation, a story is a structured, time-stamped observation about on-chain behavior. It may contain a headline, main token or entity, score or priority, narrative summary, supporting addresses, graph or evidence hints, and a timestamp. The key design principle is reproducibility: the story is a compressed explanation of a detected pattern, but the claim should remain anchored to the underlying graph.
The documented story-reading workflow is evidence-first: read the title, read the summary, inspect the addresses, open the evidence summary, and drill into /links to inspect the raw transaction graph. This establishes the basic chain of observability: narrative claim -> evidence summary -> transaction graph.
| Story family | Examples in E3D docs | Interpretation |
|---|---|---|
| Rabbit forensics | Loops, clusters, mirrors, staging, funnels, rabbit thesis | Structure-first graph findings intended to reveal operational design, obfuscation, coordination, or campaign infrastructure. |
| Market/activity | Surge, whale stealth flow, ecosystem shift, bridge paths, discovery, hot links, category rotation | Activity and participation patterns that connect on-chain behavior to market behavior. |
| Risk/confirmation | Liquidity drain, exchange flow, momentum divergence, wash trade, concentration shift, token quality score, smart money, stealth accumulation, insider timing, spread widening | Signals used to separate real activity from manipulation, liquidity stress, thin markets, or deteriorating market quality. |
The documented thesis layer is downstream of stories but upstream of decisions. A thesis is a directional brief created when multiple stories converge strongly enough to justify a structured conclusion. It is not a prediction in the abstract; it is a hypothesis with explicit evidence and explicit failure conditions.
The implemented thesis object usually includes direction (LONG, SHORT, or AVOID), conviction score, supporting story IDs, token or entity address, entry signal, invalidation signal, risk factors, and status. Depending on UI state, it may also include a conviction-history sparkline, annotations, and trade-action links.
The thesis lifecycle is central to the theory. Theses are not static. Conviction can rise as confirming stories arrive, fall when market or structure moves against the thesis, shift status from active to confirmed, or become invalidated when a risk event appears. A thesis with rising conviction is structurally different from one whose headline score is high but fading.
Since the initial v1 theory draft, the E3D implementation has been updated to align the product more directly with OTA as a measurement system. The update does not change the theory; it closes the measurement gap by making scores decomposable, thresholds explicit, state transitions hysteretic, entry and invalidation conditions machine-checkable, and outcomes available for point-in-time validation.
The implementation update is organized around six measurement surfaces: story-integrity decomposition, overlap-vector and thesis-integrity scoring, structured entry/invalidation predicates, an outcome warehouse for Delta_E3D and Delta_Theta testing, an online calibration loop, and a methodology UI that exposes the equations and live counters. These changes move OTA from a conceptual hierarchy toward an auditable measurement instrument while preserving the paper's non-claim discipline: predictability remains a hypothesis until walk-forward validation produces positive residuals.
| Direction | Implemented meaning | Decision-support implication |
|---|---|---|
| LONG | Bullish structural signals such as accumulation, cluster formation, confirmation, or breakout support. | Potential candidate for watchlist, confirmation, or entry timing if risk factors are acceptable. |
| SHORT | Bearish structural signals such as distribution, thin liquidity, or counter-trend pressure. | Potential candidate for downside monitoring or short watchlist when the entity is tradeable. |
| AVOID | Danger signals such as rugs, loops, wash-like behavior, manipulation, or severe liquidity risk. | Risk overlay, exclusion signal, or short candidate if the market structure remains active. |
We propose a four-level observable hierarchy:
Transactions -> Stories -> Theses -> Actions / Price Hypotheses
At the transaction level, the observable is a time-indexed Ethereum graph G_t = (V_t, E_t), where vertices represent addresses, contracts, tokens, pools, bridges, exchanges, and entity clusters, and edges represent transfers, swaps, approvals, liquidity actions, contract calls, or other traceable interactions. A story S_t is a compressed observation produced from a subgraph and its supporting context. A thesis Theta_t is an aggregation over stories that converge around an entity, token, wallet cluster, or market condition.
$$ S_t = f_S(G_t,\; E_t,\; M_t) $$
$$ \Theta_t = A\bigl(\{S_i\}_{i=t-k:t},\; \Omega_t\bigr) $$
Here E_t denotes evidence metadata, M_t denotes market or activity context, and Omega_t denotes overlap structure across stories: shared counterparties, shared tokens, shared wallets, shared contracts, shared liquidity venues, shared motifs, or shared time windows.
We define a composite transaction-integrity measure T_t as a weighted sum of five pillar functions evaluated on the observed Ethereum transaction graph and surrounding context:
$$ T_t = w_G f_G(G_t) + w_F f_F(G_t) + w_E f_E(G_t) + w_T f_T(G_t) + w_M f_M(G_t, P_t) $$
where the weights satisfy w_G + w_F + w_E + w_T + w_M = 1. For a default v1 implementation, we propose equal or near-equal weights unless empirical calibration later justifies otherwise:
$$ (w_G,\; w_F,\; w_E,\; w_T,\; w_M) = (0.25,\; 0.20,\; 0.20,\; 0.20,\; 0.15) $$
| Pillar | Meaning | E3D evidence examples |
|---|---|---|
| Graph integrity f_G | Topology coherence | Clusters, funnels, loops, mirrors, staging complexes, density, reciprocity, hub structure |
| Flow integrity f_F | Capital movement coherence | Net flow, bridge paths, exchange inflow/outflow, liquidity drain, accumulation/distribution |
| Entity integrity f_E | Counterparty and actor coherence | Shared wallets, profitable wallets, holder concentration, repeated intermediaries, entity quality |
| Temporal integrity f_T | Persistence and sequencing | Dormancy breaks, repeated timing, burst duration, campaign recurrence, conviction trend |
| Market-alignment integrity f_M | Relation to price/liquidity context | Surge support, momentum divergence, spread widening, volume profile anomaly, breakout confirmation |
A story is admissible when its evidence is specific enough, fresh enough, and structurally coherent enough to justify investigation. We define a story-integrity score C^S_t:
$$ C^{S}_{t} = a_E E_t + a_R R_t + a_I I_t + a_F F_t + a_C C_t $$
where E_t measures entity specificity, R_t measures graph-structure reproducibility, I_t measures potential impact, F_t measures freshness, and C_t measures confidence. The weights satisfy sum(a_i)=1. A v1 default is:
$$ (a_E,\; a_R,\; a_I,\; a_F,\; a_C) = (0.20,\; 0.30,\; 0.20,\; 0.15,\; 0.15) $$
The relatively high weight on R_t reflects the E3D documentation principle that the graph must prove or disqualify the narrative. A story is not merely a label; it is a claim anchored to evidence.
$$ A^{S}_{t} = \mathbf{1}\!\left[C^{S}_{t} \geq \tau_S\right] $$
A thesis is not a larger story. It is a higher-order aggregation of stories whose overlap structure is strong enough to support a directional view. We define thesis integrity C^Theta_t as:
$$ C^{\Theta}_{t} = b_S Q^{S}_{t} + b_O O_t + b_D D_t + b_H H_t + b_X X_t - b_R R^{\Theta}_{t} $$
where Q^S_t is the aggregate quality of supporting stories, O_t is story-overlap strength, D_t is diversity of story types, H_t is conviction-history momentum, X_t is specificity of entry and invalidation signals, and R^Theta_t is a risk-factor penalty. The weights satisfy b_S + b_O + b_D + b_H + b_X + b_R = 1 when the penalty is treated as signed contribution. A v1 default is:
$$ (b_S,\; b_O,\; b_D,\; b_H,\; b_X,\; b_R) = (0.25,\; 0.25,\; 0.15,\; 0.15,\; 0.10,\; 0.10) $$
Overlap O_t is the central bridge between stories and theses. Without overlap, multiple stories may simply be unrelated alerts. With overlap, several observations become evidence for a persistent structure.
$$ O_t = \lambda_A\, O_{\text{addr}} + \lambda_K\, O_{\text{tok}} + \lambda_C\, O_{\text{contract}} + \lambda_V\, O_{\text{venue}} + \lambda_\tau\, O_{\text{time}} + \lambda_m\, O_{\text{motif}} $$
where lambda weights are nonnegative and sum to 1. This formulation formalizes the product intuition that theses are made from aggregate stories with overlapping counterparties and related infrastructure.
$$ A^{\Theta}_{t} = \mathbf{1}\!\left[C^{\Theta}_{t} \geq \tau_{\Theta}\right] $$
The thesis-integrity score above already contains overlap and conviction components. For empirical testing, however, it is useful to isolate thesis convergence as a separable candidate term rather than burying it inside the full thesis score. We therefore define a base structural statistic that excludes the incremental thesis-convergence term, and an enhanced statistic that includes it. This follows the measurement principle that a proposed conceptual improvement should be evaluated by whether it improves the statistic out of sample.
$$ T^{\text{base}}_{t} = w_G f_G(G_t) + w_F f_F(G_t) + w_E f_E(G_t) + w_T f_T(G_t) + w_M f_M(G_t, P_t) $$
$$ T^{\text{enhanced}}_{t} = (1 - w_{\Theta})\, T^{\text{base}}_{t} + w_{\Theta}\, f_{\Theta}(\Theta_t) $$
The proposed convergence term is:
$$ f_{\Theta}(\Theta_t) = c_O O_t + c_D D_t + c_N N_t + c_V V_t + c_K\, dK_t - c_Q Q_t $$
where O_t measures counterparty, token, contract, venue, motif, and time-window overlap; D_t measures diversity of supporting story families; N_t measures the normalized depth of confirming stories; V_t measures specificity of entry and invalidation conditions; dK_t measures conviction velocity; and Q_t penalizes contradictions, stale evidence, or risk factors. The coefficients c_i are nonnegative and normalized so that the positive terms sum to one before the signed penalty is applied.
The value of this term is not assumed. It is a testable proposition. The relevant statistic is the incremental contribution:
$$ \Delta_{\Theta} = \mathrm{Score}\bigl(T^{\text{enhanced}}\bigr) - \mathrm{Score}\bigl(T^{\text{base}}\bigr) $$
If Delta_Theta > 0 under walk-forward validation, surrogate controls, and liquidity-stratified testing, thesis convergence contributes information beyond the base transaction statistic. If Delta_Theta <= 0, then the current convergence formulation should be rejected or redesigned. This keeps the v1 theory falsifiable: thesis convergence is a proposed measurement term, not an asserted performance result.
Conceptually, f_Theta is the bridge between the implemented E3D thesis workflow and a statistical measurement system. It asks whether independent stories are collapsing into the same higher-order market hypothesis. This should improve predictability only when overlap, diversity, freshness, conviction velocity, and falsifiability separate coordinated on-chain structure from unrelated bursts of activity.
To avoid unstable oscillation around the decision boundary, we propose a hysteretic thesis state process:
$$ A^{\Theta}_{t} = 1 \quad\text{if}\quad A^{\Theta}_{t-1}=0 \;\text{ and }\; C^{\Theta}_{t} > \tau_{\text{up}} $$
$$ A^{\Theta}_{t} = 0 \quad\text{if}\quad A^{\Theta}_{t-1}=1 \;\text{ and }\; C^{\Theta}_{t} < \tau_{\text{down}} $$
$$ A^{\Theta}_{t} = A^{\Theta}_{t-1} \quad\text{otherwise,}\quad \tau_{\text{up}} > \tau_{\text{down}} $$
This converts thesis activation into a sticky state variable. It aligns with the E3D thesis lifecycle: active, confirmed, invalidated, or expired. A thesis should not appear and disappear merely because one weak story arrived or one weak story aged out. It should persist when the underlying evidence structure persists.
Proposition 1: Story non-degeneracy. If nearly every story passes, or nearly every story fails, story admissibility is uninformative. Meaningful testing requires a non-degenerate story-admissibility rate: epsilon <= Pr(A^S_t=1) <= 1-epsilon.
Proposition 2: Thesis convergence. Theses constructed from overlapping stories should exhibit stronger persistence and greater predictive relevance than isolated stories because overlap filters unrelated alerts into coherent campaigns.
Proposition 2a: Incremental convergence term. A thesis-convergence term improves the framework only if the enhanced statistic T^enhanced outperforms the base statistic T^base on out-of-sample price, volatility, liquidity, or regime labels. The proposed object of measurement is Delta_Theta, not the raw existence of theses.
Proposition 3: Persistence spectrum. If E3D stories and theses capture genuine structure, then admissibility states should cluster in time at multiple horizons. The measurable object is the persistence spectrum Z(l) over lags l in {1h, 6h, 24h, 7d}, not a single alert score.
Proposition 4: Price-alignment residual. If thesis convergence contains information beyond price-only baselines, then forward price labels should be better explained by models that include thesis variables than by OHLCV or volatility features alone.
Proposition 5: Conviction dynamics. The slope of thesis conviction should carry incremental information beyond the current conviction level. Rising conviction with fresh confirming stories should be more price-relevant than high but fading conviction.
Proposition 6: Online learnability. If some story combinations repeatedly confirm while others repeatedly invalidate, then an online learning system should improve calibration over time by updating story weights, overlap weights, and risk penalties.
We define price predictability conservatively as incremental information about future market states, not guaranteed directional accuracy. The target can be forward return, volatility, drawdown probability, liquidity deterioration, or breakout confirmation. For token i and horizon h:
$$ r_{i,t+h} = \log\!\bigl(P_{i,t+h}/P_{i,t}\bigr) $$
$$ Y^{\text{up}}_{i,t,h} = \mathbf{1}\!\left[r_{i,t+h} > \eta_h\right], \quad Y^{\text{down}}_{i,t,h} = \mathbf{1}\!\left[r_{i,t+h} < -\eta_h\right] $$
$$ Y^{\text{stress}}_{i,t,h} = \mathbf{1}\!\left[\max\,\mathrm{drawdown}_{t:t+h} > d_h \;\text{ or }\; \mathrm{liquidity\ deterioration} > q_h\right] $$
The baseline model uses price-only and market-only features X^P_t. The E3D model adds story and thesis observables X^E3D_t:
$$ \Pr(Y_{i,t,h}=1) = \sigma\!\bigl(\beta_0 + \beta_P\, X^{P}_{i,t} + \beta_E\, X^{\text{E3D}}_{i,t}\bigr) $$
Incremental information is measured by out-of-sample improvement in AUC, log loss, Brier score, precision at k, or calibration error. The structural residual is:
$$ \Delta_{\text{E3D}} = \mathrm{Score}\bigl(\text{Model}_{\text{price+E3D}}\bigr) - \mathrm{Score}\bigl(\text{Model}_{\text{price-only}}\bigr) $$
A positive Delta_E3D would support the hypothesis that story and thesis structure contains price-relevant information. A zero or negative Delta_E3D would indicate that the current formulation does not improve prediction beyond price-only or market-only baselines.
The thesis-specific ablation is separate from the full E3D residual. We compare Model_price+base against Model_price+base+thesis_convergence and define Delta_Theta as the incremental improvement attributable only to f_Theta. This isolates whether thesis convergence adds predictive information beyond raw stories, graph structure, and price features.
| E3D object | Testable price-market hypothesis | Relevant observables |
|---|---|---|
| LONG thesis | Positive forward return; breakout confirmation; reduced downside risk after entry signal | Entry specificity, conviction slope, supporting story diversity, accumulation/smart-money evidence |
| SHORT thesis | Negative forward return; failed bounce; drawdown or liquidity stress | Distribution, exchange inflow, momentum divergence, spread widening, liquidity drain |
| AVOID thesis | Higher drawdown, slippage, rug/manipulation risk, or poor execution quality | Loops, wash trade, concentration shift, thin liquidity, risk factors |
| Candidate persistence | Earlier detection of future thesis creation or price-relevant repricing | Repeated candidate across buckets, matched rule persistence, increasing overlap |
The E3D implementation now targets the measurement-closure layer required by OTA. The original product already contained stories, theses, conviction history, candidate grouping, and decision-support workflows. The OTA v1 implementation update adds the missing measurable objects: explicit story-integrity components, story admissibility thresholds, overlap vectors, thesis-integrity decompositions, hysteretic states, machine-checkable predicates, outcome labels, null-model scaffolding, and calibration surfaces. The remaining distinction is empirical rather than architectural: E3D can now compute the statistics required to test Delta_E3D and Delta_Theta, but the paper still does not assert that those statistics are positive until validated out of sample.
| Layer | OTA v1 implementation update | Measurement value |
|---|---|---|
| Stories | Stories now target a decomposed C^S_t with entity, reproducibility, impact, freshness, and confidence components, plus admissible flag, tau_S, and stored weights. | Turns opaque story scores into inspectable first-order measurements and enables non-degeneracy monitoring. |
| Story admissibility | Family-level thresholds and rollups monitor Pr(A^S_t=1) over time. | Prevents degenerate thresholds and makes persistence tests meaningful. |
| Theses | Theses now target C^Theta_t, C^Theta_base, f_Theta, and C^Theta_enhanced side by side, with conviction velocity and acceleration. | Makes base-vs-enhanced ablation possible without reconstructing history. |
| Thesis convergence | Overlap is promoted from hidden rule logic into a stored vector: address, token, contract, venue, time, and motif overlap. | Allows users and researchers to see what “converging stories” means quantitatively. |
| State machine | Thesis status uses tau_up/tau_down hysteresis rather than point-in-time threshold flicker. | Improves state stability and aligns thesis lifecycle with admissibility theory. |
| Predicates and outcomes | Entry and invalidation prose is paired with structured predicates; outcome tables label forward return, drawdown, volatility, liquidity change, and stress. | Makes theses falsifiable and supports Delta_E3D and Delta_Theta measurement. |
| Methodology surface | A methodology page exposes equations, live counters, admissibility rates, backtest status, weight versions, and limitations. | Makes E3D visibly auditable as a measurement system rather than a black-box narrative engine. |
The thesis-use workflow already implies a learning system: weekly review asks which theses confirmed, expired, or were invalidated; signal sourcing asks which thesis patterns historically confirm; annotations let analysts override or record context. We formalize that workflow as an online calibration loop.
$$ \theta_{t+1} = \theta_t + \eta\,\nabla L\!\bigl(y_t,\; \hat{y}_t(\theta_t)\bigr) $$
Here theta represents story weights, overlap weights, risk penalties, and conviction-calibration parameters. Labels y_t are derived from thesis outcomes: confirmed, invalidated, expired, profitable after entry, avoided drawdown, or false positive. The loss L may be log loss for classification, squared error for calibrated returns, or ranking loss for top-k signal sourcing.
| Label | Definition | Learning role |
|---|---|---|
| Confirmation label | Entry signal occurred and forward market behavior matched thesis direction. | Improves weights for contributing story types and overlap structures. |
| Invalidation label | Invalidation signal occurred or risk factor dominated. | Increases risk penalty and reduces weak story-combination weights. |
| Expiration label | Thesis aged out without confirmation or invalidation. | Improves decay curves and reduces stale-conviction bias. |
| Human annotation label | Analyst overrides, notes, or position outcome. | Adapts generic E3D signals to a strategy-specific research process. |
| Market label | Forward return, drawdown, volatility, slippage, or liquidity outcome. | Tests whether the thesis object was price-relevant, not merely narratively plausible. |
A future empirical paper should evaluate OTA across a universe of Ethereum tokens and time windows. The protocol should be point-in-time: only stories, theses, graph structures, and prices observable at time t may be used to predict labels after time t. This avoids look-ahead bias and protects the credibility of the result.
Construct a point-in-time story table containing story type, entity, score, timestamp, supporting addresses, evidence hints, and graph links.
Construct a point-in-time thesis table containing direction, conviction, supporting stories, entry signal, invalidation signal, risk factors, status, and conviction history.
Create price and liquidity labels at horizons h in {1h, 6h, 24h, 3d, 7d, 30d}.
Compare price-only baselines against story-only, thesis-only, and story+thesis models.
Use rolling or walk-forward validation, not random splitting, because temporal leakage is the primary risk.
Test robustness across token categories, liquidity tiers, market regimes, and story families.
Publish negative as well as positive results, including story families that do not predict price.
| Null / ablation | Construction | Question answered |
|---|---|---|
| Time-shuffled stories | Preserve story counts but destroy timing. | Tests whether timing, not mere frequency, matters. |
| Entity-shuffled stories | Reassign stories across tokens with similar liquidity. | Tests whether entity-specific evidence matters. |
| Direction-shuffled theses | Preserve thesis timing and conviction but randomize LONG/SHORT/AVOID. | Tests whether thesis direction carries information. |
| Price-only baseline | Use OHLCV, realized volatility, trend, and liquidity features only. | Tests incremental E3D value. |
| Story-family ablation | Remove one story family at a time. | Identifies which structural families drive any predictive lift. |
The following items were originally framed as enhancements expected to improve predictability. Under the OTA v1 implementation update, they become measurable product instrumentation. Their value is still tested, not assumed: each added term should be stored beside its base score so that the resulting delta can be measured through ablation, null models, and walk-forward validation.
Formal overlap scoring. Make shared counterparties, contracts, venues, motifs, and time-window overlap explicit. This should improve predictability by separating unrelated alert bursts from coordinated campaigns.
Thesis-convergence ablation. Store T^base, f_Theta, and T^enhanced separately for every token and time window. This allows E3D to test whether convergence improves the statistic instead of assuming that the thesis layer is predictive by definition.
Conviction velocity. Track not just conviction level but conviction slope and acceleration. A rising medium-conviction thesis may be more actionable than a high-conviction thesis that is already decaying.
Entry and invalidation event detection. Convert prose entry and invalidation signals into machine-checkable conditions. This turns theses into falsifiable objects that can be backtested without subjective interpretation.
Candidate survival tracking. Measure how long pre-thesis candidates persist before promotion, confirmation, or disappearance. Persistent candidates may provide earlier detection than published theses.
Story diversity weighting. Reward theses supported by different kinds of evidence: accumulation plus smart money plus liquidity improvement is stronger than three copies of the same weak story.
Risk-factor penalty calibration. Learn which risk factors are genuinely predictive of invalidation or drawdown. This should reduce false LONG signals and improve AVOID/SHORT quality.
Point-in-time outcome warehouse. Store every story, thesis, update, status change, and market label exactly as observed at the time. This is the foundation for credible performance measurement.
Analyst feedback loop. Use annotations and position tracking as supervised feedback, allowing E3D to learn which signals matter for different strategies.
The core strategic implication is that E3D should not be framed as a raw alert engine. Its defensible layer is the transformation from transaction graph to story, from story to thesis, and from thesis to falsifiable decision support. Raw transactions are public; structured convergence is not.
The OTA v1 implementation update strengthens this positioning by making the science visible in the product. A story or thesis is no longer only a narrative artifact; it is a scored, thresholded, stateful object with inspectable terms, stored weights, falsifiable predicates, and eventual outcome labels. This closes the loop between product usability and research measurement.
The theory also explains why E3D-style prediction may succeed where price-path-only models fail. Price-path models observe downstream behavior after market participants have already acted. On-chain transaction graphs may expose upstream behavior: accumulation, routing, liquidity withdrawal, exchange movement, shared counterparties, coordinated wallets, and campaign infrastructure before those behaviors are fully reflected in price.
The most important distinction is between activity and structure. High transaction volume alone is weak evidence. Repeated counterparty overlap, cross-story convergence, rising conviction, and falsifiable entry/invalidation conditions are stronger evidence. OTA therefore treats price predictability as a function of structured convergence, not event frequency.
A useful analogy is credit underwriting. A single signal rarely determines credit quality; a convergent profile does. In E3D, a single story may be noisy, but a thesis supported by heterogeneous stories, overlapping infrastructure, fresh evidence, and improving conviction may represent a higher-quality market state.
This paper does not report empirical performance. Any claim of prediction remains a hypothesis until tested with point-in-time data and out-of-sample validation.
Story generation may encode model bias if LLM summaries overstate weak evidence. The proposed theory therefore weights graph reproducibility and falsifiable evidence heavily.
Ethereum visibility is incomplete when activity occurs on centralized exchanges, private order flow, opaque bridges, or other chains not yet integrated into E3D.
Price labels can be noisy because tokens vary widely in liquidity, manipulation risk, venue quality, and market regime.
Online learning can overfit if labels are sparse or if market regimes shift. Walk-forward validation and regime-aware calibration are required.
The theory assumes that documented E3D story and thesis objects can be exported or materialized in a research warehouse with stable identifiers, timestamps, and status histories.
We propose Observable Transaction Admissibility as a formal theory of E3D-style on-chain intelligence. The theory defines stories as first-order observations of Ethereum transaction structure and theses as second-order directional hypotheses generated by converging stories. It introduces composite story and thesis integrity scores, hysteretic admissibility states, persistence spectra, overlap metrics, conviction dynamics, and price-movement hypotheses. The current E3D implementation already contains the core product hierarchy: stories, graph evidence, group mode, thesis cards, conviction, entry and invalidation signals, statuses, candidate pipelines, and workflows for pre-trade review, signal sourcing, risk monitoring, and weekly review. The theoretical layer adds measurable closure: thresholds, null models, outcome labels, online calibration, and empirical validation. If future testing shows that thesis convergence improves out-of-sample prediction beyond price-only baselines, E3D would have evidence for a defensible structural-intelligence layer: not a chart predictor, but a system for detecting price-relevant state transitions before they are fully visible in price.
| Parameter | Meaning | V1 recommendation |
|---|---|---|
| tau_S | Story admissibility threshold | Empirically chosen after non-degeneracy sweep; initialize near median story-integrity score. |
| tau_Theta | Thesis admissibility threshold | Higher than story threshold because thesis requires convergence, not merely observation. |
| Hysteresis gap | tau_up - tau_down | Prevents state oscillation around boundary. |
| Story horizon set | 1h, 6h, 24h, 7d | Captures short burst, session, daily, and campaign persistence. |
| Thesis horizon set | 24h, 3d, 7d, 30d | Captures decision-support and portfolio-review timescales. |
| Minimum support | >= 2 story families or >= 3 supporting stories | Discourages single-story promotion unless evidence quality is extreme. |
| Diversity bonus | Story-family entropy | Rewards heterogeneous evidence convergence. |
| Risk penalty | Risk-factor severity | Penalizes rugs, loops, thin liquidity, manipulation, or evidence contradiction. |
These defaults intentionally emphasize observable overlap, story-family diversity, and confirming-story depth. Conviction velocity is assigned a smaller initial weight because it may be more sensitive to update cadence and LLM summarization artifacts. The contradiction penalty should be calibrated against invalidated and expired theses once sufficient point-in-time outcomes exist.
For the incremental thesis-convergence term, a conservative v1 default is:
$$ (c_O,\; c_D,\; c_N,\; c_V,\; c_K,\; c_Q) = (0.25,\; 0.20,\; 0.20,\; 0.15,\; 0.10,\; 0.10) $$
E3D Story Guide, supplied as https://docs.e3d.ai/story-guide. Defines stories as structured, time-stamped observations; explains evidence, graph reading, story families, group mode, and investigation workflow.
E3D Story Reference, supplied as https://docs.e3d.ai/story-reference. Defines story families and story types including Rabbit Staging, Funnel, Cluster, Loop, Mirror, Surge, Whale Stealth Flow, Liquidity Drain, Exchange Flow, Momentum Divergence, Wash Trade, Concentration Shift, Token Quality Score, Smart Money, Stealth Accumulation, Insider Timing, Volume Profile Anomaly, and Spread Widening.
E3D Story Use Cases, supplied as https://docs.e3d.ai/story-use-cases. Defines workflows for pre-trade diligence, risk monitoring, discovery, graph investigation, hedge-fund edge stories, campaign detection, and evidence-first review.
E3D Thesis Guide, supplied as https://docs.e3d.ai/thesis-guide. Defines theses as structured hypotheses built from converging stories, including direction, conviction, supporting story IDs, entry and invalidation signals, risk factors, status, conviction history, candidate pipeline, annotations, and trade actions.
E3D Thesis Use Cases, supplied as https://docs.e3d.ai/thesis-use-cases. Defines workflows for pre-trade review, signal sourcing, risk monitoring, weekly review, candidate pipeline, portfolio-level use, and the chain from thesis to stories to /links to notes.
Structural Admissibility Revisited: Threshold Sensitivity, Persistence Universality, and GARCH Decomposition, Shawn Knopp, Aregis LLC, March 2026, supplied as https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6444419. Used as stylistic and conceptual inspiration for admissibility, persistence, hysteresis, threshold sensitivity, residual testing, and non-claim discipline.
OTA v1 Implementation Spec - Closing the Gap Between Paper and Product, supplied as e3d_ota_v1_implementation_spec.md. Defines the six-phase implementation update used to align E3D with the OTA theory: story-integrity decomposition, overlap-vector and thesis-integrity scoring, machine-checkable predicates, outcome warehouse/backtesting, online calibration, and methodology surfaces.
This appendix records the non-theoretical implementation update that aligns E3D with the OTA v1 paper without changing the paper's empirical status. The update closes measurement gaps by turning previously implicit product behavior into stored, inspectable, point-in-time measurements.
Phase 1 - Story Integrity Decomposition: Stories are scored with explicit entity, reproducibility, impact, freshness, and confidence pillars. The composite C^S_t, applied weights, tau_S, and admissible flag are stored so story quality can be audited and threshold sensitivity can be measured.
Phase 2 - Overlap Vector and Thesis Integrity: Convergence is represented as a first-class overlap vector across addresses, tokens, contracts, venues, time windows, and motifs. Thesis rows store C^Theta_base, f_Theta, and C^Theta_enhanced to support ablation.
Phase 3 - Machine-Checkable Entry and Invalidation: Human-readable thesis text is paired with structured predicates such as price thresholds, story-type observations, conviction cutoffs, liquidity changes, and concentration conditions. This makes confirmation and invalidation observable rather than subjective.
Phase 4 - Outcome Warehouse and Delta Testing: Forward return, drawdown, volatility, liquidity change, and stress labels are stored by horizon. This enables price-only baselines to be compared against price+E3D models and base-vs-convergence ablations.
Phase 5 - Online Calibration Loop: Rule weights and thesis calibration parameters are moved from hardcoded logic toward versioned weight tables updated from confirmed, invalidated, expired, and annotated outcomes.
Phase 6 - Methodology Surface: A public methodology interface exposes the hierarchy, equations, live admissibility counters, weight versions, backtest status, limitations, and paper link. The goal is to make E3D visibly scientific and auditable.
The implementation update preserves the central non-claim of OTA v1. The product may now be able to compute Delta_E3D and Delta_Theta, but public empirical claims should wait until point-in-time, walk-forward validation shows positive results with null-model robustness.