Pons / operational evidence

Fast decisions.
Accountable execution.

How observations become trades, how failures are recovered, and how new data can improve the model.

Built 09 Sep 2026 17:09:25 UTC · status: GMGN and Model C running; both admission sources healthy · model reference 179612d6d326

From chain event to managed position

1. ObserveSequencer events and chain logs arrive independently. Useful progress, committed rows, queue age and gaps are tracked separately from websocket pings.
2. ReconcileRecent data can be provisional. Aged ranges and block identities are checked, corrections receive revisions, and dependent caches are invalidated.
3. DecideSource health and the shared entry policy gate admissions. Features, model and wallet-snapshot identities are recorded at decision time.
4. Reserve and submitA durable attempt reserves token and exposure. Signing shares a nonce owner; the transaction hash is recorded before broadcast. Unknown outcomes remain reserved.
5. ManageConfirmed holdings enter management before optional notifications. Receipt failures trigger reconciliation. Existing exits continue when new entries pause.

Ingestion and position managers operate concurrently. Signing is serialized. Timed stages are elapsed spans—not six fixed RPC round trips. Latency must include detection, commit, decision, submission, receipt and management start, including failed attempts.

What learns from new data?

The reference classifier is a gradient-boosted tree ensemble with 44 inputs. New candidate observations change its inputs. Updated wallet histories change six wallet-derived inputs. Only an explicit training job changes the fitted trees.

NumPy and the pinned scientific dependencies are installed. The updated scorer loads a validated model bundle once in process, with bounded numerical thread use. Immutable snapshot versions record content, as-of cutoff and input history; a historical self-test does not overwrite live snapshots.

CollectRecord decision-time features for selected and rejected candidates, missing scores, controls and own-trading exposure. Let labels mature.
Fit challengersCompare bounded training alternatives. Choose calibration and thresholds on validation data; reserve later data for evaluation.
Shadow and evaluateScore identical inputs alongside the reference. Compare each cohort with its own base; report calibration, coverage and economic risk.
Promote deliberatelyA nightly fit does not imply a swap. Versioned evidence gates, explicit activation and atomic rollback preserve accountability.

A fixed validation-derived cutoff does not guarantee selecting 20% of every future day's population. Comparing a score with a payoff-derived break-even probability requires matched horizons and calibrated conditional returns.

Measured discrimination, with the right target

Live-ledger level-50 crossings in the past three days, within the five-point decision band, with the full eight-minute follow-up. Selection uses the model decision, irrespective of later fill state. Same-block graduations and missing scores are excluded. Graduation means a later block no more than 480 seconds after the crossing.

Crossing dayLedger modelSelected hits / n8-minute precision95% intervalOwn eligible baseLift
2026-09-0835d112eac93f40/11933.6%25.8–42.5%51/426 = 12.0%2.81×
2026-09-09179612d6d3269/3228.1%15.6–45.4%14/97 = 14.4%1.95×
2026-09-0935d112eac93f8/4916.3%8.5–29.0%14/203 = 6.9%2.37×

These are retrospective database observations, not a globally atomic or receipt-certified snapshot. Intervals are per-token Wilson intervals; time/deployer dependence and repeated examination can make inference less certain. Previously reviewed days are development evidence, not a fresh promotion test.

Holdbacks help assess own-trading effects, but assignment, actual exposure and completed fills differ. Prediction parity establishes behavior on fixtures. Neither parity nor graduation lift establishes profitable execution.

Changes in this release

AreaImplemented behavior
ExecutionDurable attempts, shared nonce ownership, receipt reconciliation and management before optional delivery
SpeedBounded optional work, receipt reuse, source-triggered wakeups with polling fallback and complete attempt timing
DataUseful-source health, durable reconciliation/gaps, revision-aware consumers and short raw-input commits
ModelIn-process scoring, retryable unavailability, immutable feature/snapshot provenance and versioned challengers
Exit researchNet curve reserves, current-holding pool quantities, consistent trailing and explicit unknown outcomes

Activation status comes from the release record. Source changes alone do not prove that a running process loaded them.

Node and feed claims require measurement

A dedicated node may improve reliability; it does not guarantee immunity to stale state, outages or indexing delay. Wallet-set coverage is not event completeness or identical entry decisions. Compare the union of source decisions with the same policy, including misses and tokens graduating inside the observation window.

Reading a few child-chain blocks behind the sequencer is soft observation. Durable research and recovery need explicit ancestry and reconciliation semantics. Arbitrum finality documentation.

What still needs prospective evidence

Profit labels require executable quantity, liquidity, costs and timing. A pool price alone is insufficient. Missing prices remain unknown; the research search stops on incomplete outcome coverage rather than dropping those tokens. Pool state and pinned quote evidence are retained for validation.

New held-out days, source-parity results, complete economic labels and shadow performance determine whether a challenger deserves promotion. Historical shortcomings remain visible; no release claims an optimal exit or guaranteed profit.