RESEARCH · EVIDENCE ENGINEERING · AGENT OUTPUT QUALITY
AegisRho research is not market commentary. It is the technical work of building decision-grade evidence: multi-horizon microstructure, derivatives context, on-chain capital-flow priors, narrative risk stages, and a labeled trade corpus the agents distill into a living playbook. Better research → cleaner inputs → more selective, auditable agent output.
Each family is designed for a specific role in the decision stack: structure, positioning, chain prior, narrative risk, hard blackouts, and labeled memory. Together they form the evidence file the agents actually consume.
Closed-candle feature bundles across three horizons: trend structure, momentum state, and ATR-normalized volatility. Features are computed only on completed bars so the analyst never injects look-ahead into the evidence tensor.
Funding rate, open-interest delta, and order-book imbalance enter as conditional context around price structure — not as independent alpha. The research question is how positioning pressure modulates structure quality, not whether funding alone “predicts” direction.
Smart-money stablecoin netflow is modeled as a risk-appetite prior for BTC: net into stables = risk-off, net out = risk-on. Perp skew from profitable wallets is fail-soft and advisory; netflow is the gate. Missing or stale chain state degrades to neutral — never forces a trade.
Normalized news items with full article retention for auditability. Used twice: (1) entry-time narrative context for the decider, (2) adverse-risk pre-screen for the sentinel. Research focus is precision on emergency-close candidates, not headline sentiment noise.
Economic calendar events are hard constraints, not soft features. Research treats known volatility windows as stand-down regions so agents do not spend model capacity or capital through scheduled shocks.
Every cycle writes a full evidence file, model verdict, and realized outcome. Refused trades are shadow-replayed against subsequent path using the same ATR brackets. The corpus is the training substrate for playbook distillation — the research output agents actually read next cycle.
Public chain and venue data is abundant and noisy. Research work is deciding which series are gates, which are advisory, and how agents should degrade when a feed is missing — so output quality stays fail-closed.
Continuous trading hours, dense funding/OI history, and usable on-chain risk-appetite lenses make BTC the highest-signal environment for agent evaluation. Secondary assets are admitted only when they meet the same evidence-completeness bar.
We treat sophisticated stablecoin rotation as a slow prior on BTC risk appetite, not a tick-level signal. Netflow updates the evidence pack; it does not bypass structure or hard gates. Calibration goal: reduce false risk-on entries when capital is clearly hiding in stables.
Research separates gate features from advisory features. Netflow can neutralize; perp skew can only tint. If an endpoint 403s or returns empty, the agent continues with an explicit unavailable state — never imputed “average” flow that would silently bias the decider.
News is expensive and noisy. We stage it: cheap pre-screen on persisted items, escalate to deeper web search only under stress. The research objective is high precision on emergency_close — false positives cost opportunity; false negatives cost capital.
Agent quality is bounded by data completeness. BTC offers continuous liquidity, dense derivatives tape, and usable on-chain lenses. Multi-asset expansion is gated on the same completeness criteria — research rejects thin long-tail coverage that would degrade evidence quality.
Frontier models supply general reasoning. AegisRho research supplies the crypto-specific evidence, labels, and constraints that turn that reasoning into disciplined agent behavior.
Agents only decide as well as the evidence file allows. Research work — feature hygiene, stale-data policy, gate design, shadow labels — raises the signal-to-noise of that file. Better inputs yield more selective entries, fewer forced trades, and cleaner post-mortems for the next playbook cycle.
The objective is not raw PnL maximization inside the model. It is calibrated selectivity: honest SKIP rates, documented refusal quality via shadow replay, and a compact playbook of rules each backed by multiple labeled examples with win/loss counts the decider can weight.
We do not online-tune leverage, risk-per-trade, or blackout windows from outcomes. Those stay outside the learning loop. Research can make agents more careful; it is forbidden from making them more aggressive through silent parameter drift.
Working papers on how we build and evaluate the research stack that feeds the agents. Educational — not trade recommendations.
How we align multi-horizon TA, derivatives context, chain priors, and narrative state into a single snapshot with explicit missingness — so agent prompts never pretend completeness they don’t have.
Why netflow is treated as a hard-ish prior while perp skew stays fail-soft, how staleness thresholds neutralize rather than invent, and what failure modes we observed under API degradation.
Protocol for replaying blocked candidates against subsequent candle paths with ATR SL/TP, scoring good vs bad blocks, and converting graded refusals into source=shadow lessons without polluting the money path.
Cost–precision tradeoffs in sentinel design: cheap pre-screen on stored news, escalated web search under stress, and strict emergency_close criteria so narrative risk never becomes discretionary re-entry.
A completeness argument: continuous hours, dense funding/OI, and usable chain lenses make BTC the highest-ROI research surface until secondary markets meet the same evidence standard.
▸ Research materials describe system design and evaluation. They are not investment advice. Crypto is volatile; losses are possible. See risk disclosure.
See how the evidence stack becomes a fail-closed long / short / skip call — then a guarded position and a labeled lesson.