The evidence-gated intelligence loop

DUMMY THINKS

Forty-five independent loops turn point-in-time public evidence into competing probabilities, force disagreement into the open, learn only from settled outcomes, and keep research intelligence structurally separate from authority over capital.

Open the system maps See every ability

The Organism visualizes persisted evidence. The DOM truth ribbon remains authoritative; the page is loopback-only, GET-only, and incapable of placing an order.

competing sources

autonomous loops

cycle phases

sports challengers

sports leagues

crypto scopes

dedicated crypto loops

LLM roles

automatic authority

Capability counts describe registered repository components, not profitability, deployment readiness, or permission to trade. Current launch status remains NO-GO until every external evidence gate is satisfied.

Visible by design

Two system maps, loaded as standalone assets

These diagrams sit near the top of the release and open at full resolution. They are ordinary SVG image assets—not script-rendered diagrams—so GitHub Pages, reduced-motion browsers, screen readers, and static mirrors all receive the same architecture.

Dummy capability map from observation through reasoning and learning to constrained action

Capability mapTwelve ability families, including crypto charts, model dissent, autoresearch, the firewall, Board, and Observer MCP.

Dummy 45-loop fleet map with explicit crypto paper twin and horizon evidence loops

45-loop fleetEvery responsibility band plus the explicit CryptoPaperTwin and CryptoHorizonEvidence cadence lane.

Complete public capability catalog

Every major ability—and the surface that proves it exists

This catalog is intentionally broader than a feature list. Each ability names its owning system and its proof boundary, so research tooling, operational visibility, and execution authority cannot blur together.

Observe

Market discovery & normalization

Kalshi and public context feeds become timestamped, deduplicated market identities with provenance and freshness.

proof: observation ledger + source health

Crypto · Observe

Multi-timeframe chart visualization

BTC, ETH, and SOL candlesticks across 15m, 1h, 4h, 1d, and 1w with deterministic indicators and pattern markers.

proof: immutable Market Observer chart bundles

Crypto · Loop

Crypto paper twin & horizon evidence

Dedicated 5-minute paper-twin and 10-minute horizon-evidence tasks grade asset × timeframe × strategy lanes.

proof: forward paper and horizon artifacts

Sports · Observe

History & play-by-play lakes

Seven league surfaces combine point-in-time games, box scores, EPA, period state, comeback matrices, and player context.

proof: source manifests + event-purged lakes

Reason

Forecasting & simulation

Market anchors, statistical kernels, power ratings, scoring distributions, live models, and scenario simulators emit attributed probabilities.

proof: preregistered source forecasts

Reason

Specialist council & Model Arsenal

Vertical specialists and four schema-bound LLM roles preserve dissent under exact routing and paid-call gates.

proof: redacted stored witnesses, zero auto authority

Calibrate

Trust, uncertainty & fusion

Brier, log loss, ECE/MCE, debiasing, contested-market scoring, and scope-specific uncertainty determine earned weight.

proof: settled contested forecasts

Evaluate

Walk-forward & backtest evidence

Temporal folds, event-cluster bootstrap, fees, liquidity, CLV, partial fills, and negative controls challenge every claim.

proof: predict-before-update reports

Allocate

Candidate portfolio allocation

Evidence-adjusted edge and settlement velocity divide one bounded pot across holdable, correlation-aware candidates.

proof: Σ grants ≤ pot and grant ≤ ask

Constrain

Risk governor & execution firewall

Stage, drawdown, clusters, liquidity, TTL, sealed caps, session, credential, and LIMIT-only checks can only shrink intent.

proof: typed blocks + transport witnesses

Truth

Reconciliation, settlement & audit

Orders, fills, cancellations, outcomes, corrections, and account snapshots remain separate evidence layers.

proof: append-only identities + corrections

Learn

Autoresearch, tuning & evolution

Strategy mining, parameter tuning, quality-diversity search, true crossover, ablation, chaos, and fragility create challengers.

proof: observational until promoted by evidence

Question

Self-scout & film room

Bias tendencies, worst-call reconstruction, recruiting, matchup lens, top threat, and development tracking expose blind spots.

proof: report failures surface as failures

Operate

Watchdog, healing & durability

Independent task health, self-heal, snapshots, ledger retention, pruning, vacuum, and allowlisted rotation contain failures.

proof: cadence-bound artifacts + bounded storage

Explain

Operator Board & desktop notifier

Overview, scopes, charts, Arsenal, glossary, readiness, health, and freshness render persisted truth without mutation controls.

proof: loopback-only GET surface

Expose

Market Observer MCP

Allowlisted public candles, indicators, patterns, chart bundles, network status, and source health through read-only tools.

proof: false execution / order / allocation authority

Dummy operator Board showing all sixteen public-release capability families

All abilities on the Board — the same sixteen capability families are now visible inside the read-only operator UI, including market perception, both named crypto loops, chart visualization, settlement memory, metacognition, the firewall, fleet reliability, and Observer MCP.

Intelligence is a system property

Six capabilities. One inspectable memory.

Dummy does not call one model “the intelligence.” Its intelligence comes from preserving provenance, making probabilistic claims before outcomes are known, recording dissent, grading every eligible claim, and constraining what learned confidence is allowed to change.

Perception

Timestamped market, venue, macro, weather, roster, schedule, and game-state observations enter with source, freshness, and market identity attached.

Stale, missing, or ambiguous evidence becomes an abstention—not a synthetic fact.

Probabilistic reasoning

Market anchors, statistical kernels, simulations, specialists, and quarantined language-model roles emit schema-bound probabilities at explicit horizons.

Model confidence and fluent prose are not treated as calibration.

Dissent

Every source keeps its identity. The market, specialist panels, power ratings, and challengers can disagree without their evidence being flattened away.

Consensus is not manufactured by hiding the minority forecast.

Memory

Observations, forecasts, settlements, corrections, calibration, promotion dossiers, and execution witnesses are retained as separate evidence layers.

A narrative report cannot overwrite a transport witness or settled outcome.

Metacognition

Self-scouting, film-room reconstruction, ablation, drift checks, walk-forward replay, and fragility tests ask where the organism is wrong and why.

Thin samples, hindsight, and correlated events cannot masquerade as durable skill.

Constrained action

Allocation, risk, caps, freshness, session, and firewall gates form a monotone chain: each stage can shrink a proposal; none can expand upstream authority.

Research success, paper results, and model votes cannot authorize capital.

Capability map — intelligence with inspectable feedback Each capability reads or writes the same auditable evidence memory while keeping its own contract, uncertainty, and refusal boundary.

solid = forward evidence · dashed = learning

What it prices

The arsenal

Every market is priced by a purpose-built model, cross-checked by challengers, and gated on settled evidence before it can earn trust.

Crypto

BTC · ETH · SOL — 15m / hourly / daily / weekly

Realized & implied volatility, market and macro risk regime (S&P / DXY / VIX / 10y / gold / oil)
Momentum, technical, order-book and cross-venue evidence
Multi-timeframe structure, candlestick & divergence chartist, patience/confirmation entries

Sports

MLB · NFL · NBA · NHL · WNBA · NCAAF · NCAAMB

Winners, totals, spreads, team totals, YRFI/NRFI, first-five, halves, quarters, periods
Platoon splits, bullpen quality, rivalry awareness, plate-appearance simulation
Power-ratings ensemble (FPI/BPI, Elo, Massey, tie-aware Colley) + licensed multi-book de-vig
162,915-game point-in-time lake → 11 walk-forward-tuned challengers: Glicko-2, MOV-Elo, Pythagenpat, Four Factors, EPA/play, scoring sigmas, rest shift, live win-prob, referee totals, player props
32,298-game play-by-play knowledge base — empirical comeback matrices for live re-pricing
Player props via metered licensed consensus; correlation-aware parlay pricing

Autonomous improvement

the machine tunes itself

Two-stage promotion ladder — challengers earn their place from settled evidence
Evolution lab: quality-diversity archive, true crossover, jitter-fragility verdicts
Self-scout, film room, recruiting board, matchup lens, top-threat, development tracker
Fusion ablation, property-invariant sweeps, chaos drills, and sharp/sheep diagnostics retained as evidence

Council & debate

specialists + LLM voices

Per-vertical specialist subagents with coherence, CLV, and trust surfaces
Exact 4-model panel — Gemini 3.6 Flash, GPT-5.6 Luna, Claude Sonnet 5, GLM-5.2 — 7-call atomic, double-locked, graded per voice
Every specialist fails closed — missing data means abstain, never a degraded guess

Crypto is a first-class intelligence lane

Observe, chart, price, grade, and learn across five horizons

Dummy does not hide crypto inside a generic market loop. BTC, ETH, and SOL have dedicated observation, charting, paper-twin, and horizon-evidence paths, while shared forecast, calibration, risk, and research loops retain the asset and timeframe identity.

Dummy Crypto Research Charts showing BTC candlesticks, deterministic indicators, pattern markers, provenance, and no production authority

Crypto Research Charts — the actual artifact-only chart UI, shown with a clearly labeled synthetic release fixture. The production observer accepts rights-reviewed closed public candles; both paths compute deterministic RSI / EMA / ATR / MACD / Bollinger / stochastic / OBV facts, final-bar markers, age, provenance, and explicit false execution authority.

Market Observer READ-ONLY SIDECAR

Coinbase Exchange public candles → immutable content-addressed chart bundles. The dashboard never contacts the provider itself.

DummyCryptoPaperTwin EVERY 5 MIN

Runs BTC / ETH / SOL paper lanes for 15m, 1h, 1d, and 1w cohorts with isolated asset × timeframe × strategy evidence and no capital authority.

DummyCryptoHorizonEvidence EVERY 10 MIN

Builds forward horizon matrices and evaluates whether each crypto cohort has enough time-consistent evidence to remain a challenger.

Shared crypto intelligence ATTRIBUTED

ShadowPredator, MispricingMonitor's crypto-fast lane, recalibration, backtests, autoresearch, allocation, and risk preserve the crypto scope instead of merging it into sports evidence.

Research boundary: charts display facts, not price targets; paper twins record hypothetical decisions, not fills; horizon evidence can challenge a model, not authorize an order. The chart page is GET-only and all execution, order, cancel, amend, allocation, and promotion authority fields remain false.

The cognition cycle

Eight phases, each with a contract

Public evidence flows in one direction. Each phase produces a bounded artifact for the next one, so a failure can be located, replayed, and explained without pretending the entire organism is one indivisible model.

Scan→Signal→Fuse→Allocate→Risk→Execute→Reconcile→Learn

01 / OBSERVE

Scan

Discover supported markets and attach point-in-time public context, venue state, freshness, and provenance.

output → normalized observations

02 / PREDICT

Signal

Run every eligible source at the correct asset, league, market type, and horizon; preserve abstentions and reasons.

output → attributed probabilities

03 / CALIBRATE

Fuse

Debias each opinion and combine it by earned, scope-specific trust while retaining the market as a standing anchor.

output → contested consensus

04 / PRIORITIZE

Allocate

Rank evidence-adjusted edge by settlement velocity and divide one bounded pot across candidates that can actually be held.

output → capped candidate grants

05 / CONSTRAIN

Risk

Apply stage, drawdown, correlation-cluster, bankroll, liquidity, price, and time-to-live gates.

output → reduced order intent

06 / FIREWALL

Execute

Shadow by default. Any controlled live proof still requires explicit authority, an active session, sealed caps, fresh state, and LIMIT-only transport.

output → paper receipt or transport witness

07 / TRUTH

Reconcile

Match orders, fills, cancellations, market settlement, and account evidence without promoting local gate blocks into broker claims.

output → outcome-linked evidence

08 / ADAPT

Learn

Brier-score all eligible forecasts, refresh calibration, inspect failures, tune challengers, and append the next promotion evidence.

output → forward-only trust

The monotone rule: fusion cannot outvote calibration, allocation cannot exceed the candidate pot, risk cannot exceed the allocation, and the firewall cannot exceed sealed caps. Intelligence may recommend; only explicit operator authority may permit.

One pot, many candidates

Capital allocation

Sizing one order and dividing a budget across many are different problems. The risk brain solved the first — quarter-Kelly under a SHADOW → CANARY → RAMP → CRUISE ladder. Nothing solved the second: the top-ranked candidate took the entire remaining budget its caps allowed, a near-equal rival got nothing, and forty qualifying candidates sized identically to three.

Three split policies

explicit configuration workflow

kelly_prorata (default) — each ask scaled by weight; apportioned to fit only when the pot is oversubscribed. Thin nights deliberately deploy less than the cap
proportional — share of the pot by weight, clamped to the ask
top_k — fund the highest-weighted K in full, starve the tail

Weighted by proof, not promise

contested Brier advantage

Weight comes from how much better the model's Brier is than the market's own price on the same rows
Uses the lower 95% bound — the same quantity the promotion gate tests. Twelve lucky rows cannot size up on evidence the sample won't support
The weight floor is never zero: a scope that can never be allocated can never settle, never accrue evidence, and never earn its way up

Two rules learned the hard way

now property-tested

Weights are a contention tiebreak, not a discount. An early version scaled every ask by its weight unconditionally — a lone 100¢ ask from an unproven scope became 25¢, below one contract, and trading stopped silently. A change that sizes everything to zero looks conservative and is not safety
Divide across holdable slots, not evaluated candidates. A hundred are evaluated per cycle; a stage may permit five open markets. A hundred-way split puts every grant under one contract

Invariants

proven over generated inputs

Σ granted ≤ pot, always
granted ≤ ask, always — the allocator can only reduce, and every existing cap still binds behind it
Adding a candidate never raises an existing grant
The operator throttle can only shrink the pot; enlarging it past the sealed ceiling is a registration ceremony, not a slider

The autonomous fleet

45 loops, isolated by responsibility

Dummy is not one immortal process. Its intelligence is distributed across independent Windows scheduled tasks that survive reboots, pick up merged code on their next fire, and are individually observed. One slow report, broken feed, or league backfill cannot hold the cognition cycle hostage.

Cadence — the heartbeat

8 loops

The brain cycle, live poller, sports board refresh, mispricing monitor, DummyCryptoPaperTwin every 5 minutes, the vNext shadow organism, the USE sidecar, and account snapshots

Predator · LivePoller · SportsBoard · Mispricing · CryptoTwin · vNext · USE · AccountSnapshot

Learning & self-improvement

10 loops

Weight recalibration split into a fast weight-only core and heavy diagnostics, so a slow report can never stall a cycle — plus the tuner, strategy miner, autoresearch lab, simulation trainers, and DummyCryptoHorizonEvidence every 10 minutes

Weights · Backtest · Tune · SelfImprove · StrategyMiner · Autoresearch · Simulation

Sports data & grading

18 loops · per league

History backfill and event-purged walk-forward grading for all seven leagues, box-score ingest for three, and NFL EPA — each league on its own schedule so one outage cannot stall the rest

Lake × 7 · WalkForward × 7 · BoxScore × 3 · NFL EPA

Health & durability

5 loops

A fleet-wide watchdog that grades the other tasks rather than its own exit code, a five-minute self-heal loop, readiness reporting, and a snapshot writer so the board never holds a ledger lock

Watchdog · Healer · Readiness · Dashboard · DashboardSnapshot

Bounded footprint

4 loops

Ledger retention, pruning, a self-skipping vacuum, and log rotation on an explicit allowlist — state, audit and full-history tapes are never blind-truncated

Retention · Prune · Vacuum · LogRotation

Why independent loops matter

fault containment is intelligence infrastructure

Each cadence has its own timeout, evidence artifact, and next retry
The watchdog evaluates the fleet instead of trusting its own exit status
Read-heavy diagnostics cannot stall price formation or hold the ledger open
Recovery resumes from persisted artifacts rather than hidden process memory

Where the crypto loops live: DummyCryptoPaperTwin is one of the 8 heartbeat tasks; DummyCryptoHorizonEvidence is one of the 10 learning tasks. ShadowPredator, MispricingMonitor's crypto-fast pass, recalibration, backtests, and autoresearch also process crypto under preserved asset × timeframe identities. The two dedicated tasks are included in the total of 45, not added on top of it.

Loop fleet map — 45 isolated jobs, one evidence plane Five responsibility bands run on independent cadences and publish bounded artifacts. The board observes those artifacts; it is not a control path.

8 + 10 + 18 + 5 + 4 = 45

The fleet can improve models, weights, calibration, and candidate selection. It cannot rewrite truth rules, lower promotion evidence, widen sealed caps, mint an operator signature, or grant itself a live session.

The command board

One hardened view over persisted runtime truth

The Dummy operator board is a loopback-only, read-only web board at 127.0.0.1:8787. It reads persisted artifacts and exposes no scheduler, configuration, authority, capital, model-provider, or broker mutation endpoint. Split-flap values change only when evidence changes, and a ⌘K palette jumps to any coin, league, model role, or saved theme.

The Organism

data-bound neural field · progressive enhancement

The four-model arsenal forms the cortex; sources and scopes orbit by tier
Pulses appear only when persisted evidence changes, never as fake activity
WebGL2, WebGL1, Canvas 2D, and static fallbacks preserve the same truth
Reduced motion stops animation without hiding evidence

Truth before spectacle

accessible DOM remains authoritative

Sticky status ribbon separates account freshness, authority, board, and system health
Forecast diagnostics are labelled as probability quality—not trading profit
Stored model witnesses are redacted and never refreshed by opening the page
Responsive scope docks retain keyboard and touch navigation

Updated Dummy Overview showing the Organism, truth ribbon, account state, and evidence diagnostics

Overview · current Organism buildUpdated UI capture — sticky execution truth, cached account state, system health, forecast diagnostics, and the data-bound neural field in one loopback-only operator surface.

Per-league scope — graded forecast quality, model vs book, live picks by edge, accuracy by bet type, and a day-by-day games breakdown

Per-coin scope — the model beating the closing line, accuracy by bet type, every priced market by category, and today's settled calls

The evidence ladder

What the organism knows—and what that knowledge cannot do

Dummy keeps four layers deliberately separate. Moving upward requires new evidence; no lower layer is allowed to imply the layer above it.

LAYER 01

Forecast

A timestamped probability from an identified source, scope, market type, and horizon—made before the outcome.

Cannot prove calibration, edge, profitability, or authority.

LAYER 02

Settled evidence

Outcome-linked Brier, log-loss, closing-line, sample-size, and event-cluster evidence produced by forward grading.

Cannot bypass uncertainty bounds, replication, or promotion review.

LAYER 03

Promotion

A bounded dossier showing a challenger met preregistered evidence gates for a defined role and scope.

Cannot create broker credentials, sealed caps, or a live session.

LAYER 04

Authority

An explicit, operator-held, time-bounded control contract checked again at the central execution firewall.

Cannot be manufactured by code, UI state, paper results, or model consensus.

Non-negotiable

Safety invariants

Paper / shadow by default — no real capital at risk
Promotion to capital is human-only, never autonomous
Execution is fail-closed and LIMIT-only behind a hardened firewall
Challengers are evidence-gated on settled, contested outcomes
Risk caps are byte-sealed — code may register a new baseline but is structurally forbidden from manufacturing the operator approval needed to use it. Self-authorization is unrepresentable, not merely discouraged
Markets are deny-by-default — an order cannot form unless the market is positively authorized by exact ticker or by series. Series matching is boundary-aware identifier matching, never category inference; category compliance comes only from fetched venue metadata
Claims carry timestamps — a validated candidate's tradability expires, and a market never observed reports null rather than false. "Not tradable" and "we never looked" are different facts
Broker contact is claimed only on transport evidence. Two early proofs recorded as BROKER_REJECTED never reached a broker — no submit call, no endpoint, no payload, submit disabled throughout. Both were local gate blocks wearing a broker's label; the rejection classifier and truth layer exist because of them
Research may evolve; truth rules, promotion standards & the firewall may not
Every paper decision is explained and recorded in an auditable ledger