An Adaptive, Self-Auditing Framework
for Equity Signal Generation

The Single Stock Intelligence Agent — internally designated CLAW VRT — is a continuously running analytical system that produces actionable swing-trading signals for a single equity instrument. At the time of this writing that instrument is NYSE:VRT (Vertiv Holdings). The agent operates as one node in the broader Agentic Womb fleet, sharing infrastructure conventions with sibling agents but specialized exclusively for the rigors of equity intelligence.

This document is a non-technical overview of how SSIA does what it does. It describes the loop, the ledger, the lifecycle of a hypothesis, and the risk discipline that wraps the whole system. It does not describe the exact composition of the composite scoring model, the specific entry triggers, or the parameter thresholds that turn data into decisions. Those remain under glass.

Most retail and even prosumer trading systems suffer from the same structural flaw: their model is fixed. A fixed model is a snapshot of the past masquerading as a window into the future. It works until it doesn't, and the moment it stops working is precisely the moment it appears most confident.

SSIA is built on the opposite premise. The model itself is not the product. The product is the process by which the model continuously revises itself. Every hypothesis we hold about VRT — about what makes it move, what predicts a pullback, what signals a regime shift — is treated as a candidate, not a conclusion. Each one earns its place through evidence. Each one is killed when evidence stops accumulating.

This is intelligence, not divination. The system is not designed to be right. It is designed to find out it is wrong faster than its competitors.

SSIA sits on a six-layer stack, each with a distinct responsibility:

• Data Ingestion — Live price tape, fundamental filings, regulatory documents, options chain telemetry, peer-cohort tickers, macro context (rates, vol, breadth), and selective unstructured signals (executive language, geopolitical events). Production pricing is sourced via Finnhub WebSocket and REST polling; backstop sources are layered for redundancy. Yahoo's v7 endpoint, formerly relied upon, is no longer trusted.
• Hypothesis Stack — The active set of testable beliefs about what predicts VRT's near-term price action. Currently 18 active, with 9 of an original 27 culled to date.
• Composite Engine — The transformation layer that takes weighted signals from the stack and produces a single integer score, 0–100, representing the aggregate read on VRT.
• Decision Layer — Translates the composite into one of four action verdicts: HOLD, WATCH, ACCUMULATE, CONVICTION, with position-sizing implications.
• Risk Discipline — A standing set of gates and protocols that override the decision layer when drawdown, volatility, or regime conditions demand it.
• Audit Layer — Every input, every signal fire, every kill, every score recalculation is logged with timestamp and rationale. Nothing the system does is opaque to itself.

The heartbeat of SSIA is a continuous loop that runs throughout the trading session and during select pre-market windows. We refer to each cycle as a Loop, indexed sequentially. The dashboard pictured at the time of this writing shows Loop #391.

INGEST pulls fresh data from all subscribed sources. Critically, this phase enforces a market-hours gate: signal-generating writes to the source-of-truth database are suppressed outside the 9:15 AM – 4:15 PM ET window, with decision logic gated more tightly to 9:30 AM – 4:00 PM ET. This was a hard-won lesson — early versions of the system produced false-positive signals overnight as stale percentage values triggered thresholds without active price discovery.

EVALUATE runs each active hypothesis against the new data and updates its individual signal value. This is the layer where domain knowledge lives — what counts as a "signal fire" for hypothesis H08 is structurally different from what counts for H01.

SCORE combines the individual signals into the composite. The weighting mechanism is informed by each hypothesis's running hit rate over a rolling 90-day window. VERDICT maps the composite to one of four action bands. The current verdict at writing — composite 52, WATCH band — implies "trim into strength to half-risk."

EXECUTE is human-supervised. SSIA does not place orders. It produces actionable instructions which are reviewed and routed manually. This is by design. AUDIT records the entire loop transaction: what fired, what didn't, what the verdict was, and what the trail of reasoning looked like. The audit log is the source of truth for the next loop's hit-rate computations.

Every hypothesis on the stack passes through a defined lifecycle. The visible dashboard surfaces these states explicitly.

• UNTESTED — A new candidate idea. It is being instrumented but has not yet accumulated enough fires to be statistically meaningful. H17 (arxiv DC-cooling velocity) and H18 (earnings drift residual) currently sit here.
• PROVISIONAL — The hypothesis has fired enough times to compute a hit rate but has not yet cleared the promotion gate. Provisional hypotheses contribute to the composite at reduced weight.
• BACKTESTED — The hypothesis has cleared a 30-fire minimum and demonstrated a hit rate above the 55% promotion threshold. It contributes at full weight. H01, H03, H05, H08 currently sit in this state.
• KILLED — A hypothesis whose hit rate has fallen below a kill threshold, whose mechanism has been definitively explained by another, or whose underlying premise has been falsified by observation. Killed hypotheses are removed from composite scoring but retained in the audit log forever — they are valuable as priors against re-introducing the same idea under a new name.

VRT does not trade in isolation. It moves inside a constellation of structurally related instruments, and many of the system's most productive hypotheses are relational rather than absolute.

• Direct competitor anchor — ETN (Eaton). The VRT/ETN ratio is one of the longest-running observed relationships in the system. Hypothesis H08, which fires when this ratio reaches a statistical extreme, has produced a 57.9% hit rate over 21 fires.
• Sector ETF anchor — XLI. VRT's beta to XLI is approximately 2.0×, meaning broad industrial moves are amplified at the VRT level. Cohort context informs regime detection.
• Hyperscaler-capex chain — Indirect upstream demand signals, principally from MSFT/AMZN/GOOGL capital-expenditure tape. H01 (AWS/Azure capex run-rate) currently leads the stack at 62.4% hit rate over 36 fires.
• Adjacent power-density names — Companies whose order books move on the same physical-infrastructure thesis as VRT.

The cohort layer exists for two reasons. First, relational signals are harder to fake — a single ticker can be moved by news, but a structural ratio between two tickers tends to revert to fundamental forces. Second, cohorts provide implicit cross-validation: a VRT signal that fires in isolation while every cohort peer is silent is treated more skeptically than one that fires with cohort confirmation. The peer tickers are in the model. They are not in the portfolio.

A signal that works in a risk-on environment may behave perversely in a risk-off one. SSIA addresses this with a regime-detection layer that runs in parallel with the hypothesis stack. The current regime read is RISK-ON · TIER 2, derived from a composite of:

• VIX quartile positioning
• ISM PMI level and trajectory
• 10-year treasury yield 5-day delta
• HYG option-adjusted spreads
• The VRT/ETN ratio's own standard deviation

When the regime shifts, hypothesis weights shift with it. A hypothesis with strong hit rate in expansionary regimes may have its weight de-emphasized — or zeroed — when the system reads contraction. The regime layer is not a switch; it is a continuous bias applied to the composite.

The most expensive part of any trading system is the part that decides when not to trade. SSIA implements three concurrent risk protocols that compose with each other and with the upstream signal layer.

• Drawdown Protocol — Peak-to-trough position drawdown is monitored continuously. A soft gate at −4% triggers an automatic position-size reduction to 0.5R. A hard gate at −8% triggers a flat-out exit and a 30-day lockout from re-entry. As of this writing the soft gate has tripped — current peak-to-trough is −4.2% — and the system has correctly cut to 0.5R.
• Volatility-Targeted Sizing — Position size is normalized to realized 21-day volatility. Higher realized vol forces smaller positions for the same dollar-risk envelope. VRT's current 21-day realized volatility is 62.8%, which is structurally elevated; sizing reflects that.
• Market-Hours Gating — Signal generation is windowed to live trading hours with a small write-buffer. Off-hours stale-data writes are explicitly suppressed. This was added after early versions produced false-positive overnight signals.

A signal can fire, the regime can be supportive, the cohort can confirm, and the protocol layer can still say no — and that vetoes everything upstream.

The system is observable through six interlocking views, each pictured in the production dashboard:

• 01 DASHBOARD — Composite score, verdict, regime, sizing, drawdown state, loop log. The trading-room cockpit.
• 02 GRAPHS — 90-day price tape with signal fires and kills marked, the VRT/ETN ratio z-score chart, hypothesis hit-rate bars, decision trail.
• 03 HYPOTHESES — The full active stack with state, signal, hit rate, fire count, last-fire date.
• 04 AUDIT LOG — Append-only loop-by-loop transaction record.
• 05 BACKTEST — Historical replay against any custom hypothesis specification, used for promoting candidates and re-validating live ones.
• 06 BRIEF — A natural-language pre- and post-market summary written by the system itself, condensing the state of the world into paragraphs a human can read in under sixty seconds.

The Brief is, for daily operational purposes, the most important view in the system. Everything else is the audit trail for what the Brief recommends.

VRT is at $284.52, six trading days off an all-time high of $297.10. The composite reads 52 — squarely in the WATCH band. The ratio mean-revert hypothesis (H08) fired at +3.3σ on April 20 and has since validated, with the ratio now compressed to +2.1σ. The soft drawdown gate has tripped. The verdict is: trim to 0.5R into strength.

This is the system doing exactly what it is built to do. It is not predicting the bottom of a pullback. It is reducing exposure as evidence accumulates that recent strength is unsustainable, while keeping skin in the game in case the broader thesis — capex-driven structural demand — reasserts itself.

The current configuration represents Beta 1.0: backtest-calibrated, with several originally proposed signals (S4, S9, S10) disabled after they failed to outperform random. Live signals include S1_LAG and the COMPOSITE, both at 67% historical hit rate over a +4.3% average 3-day return.

The immediate roadmap focuses on three threads:

• Tighter cohort instrumentation, particularly around hyperscaler quarterly disclosure cycles
• Re-introduction of options-skew telemetry once the underlying hypothesis (H14) clears its current 52.1% hit rate plateau
• Migration of select reporting components into a shared infrastructure layer with sibling CLAWs

Longer term, the framework itself — the loop, the lifecycle, the audit discipline — is intended to be portable. SSIA is the first instance, calibrated for VRT. Nothing about the architecture is specific to a single ticker. The model dies and revives every day; the system is what persists.