Arch-AGI is Treeova's conviction analysis engine. It runs a structured 7-pass pipeline over an open trade or candidate position to produce a qualitative conviction score, scenario projections, regime-aware learning notes, and an adversarial stress test before surfacing a recommendation.

How does the 7-pass pipeline work?

Each pass takes the prior pass's output as structured input and adds a distinct analytical lens: edge framing, scenario projection, risk/reward sizing, regime context, macro tailwinds and headwinds, reinforcement-learning calibration against historical outcomes, and an adversarial stress challenge. The final output is a single coherent report with the conviction score and rationale.

How is conviction scored?

Conviction is expressed on a 0–100 scale derived from the agreement and quality of the seven passes. Higher scores require alignment across edge, scenario, risk/reward, regime, and stress-test passes. Treeova does not publish the exact weighting or threshold logic, but the pass outputs that drive each score are surfaced in the report so the user can audit the reasoning.

Where does the learning loop come from?

Every triggered position outcome is recorded as feedback against the conviction and momentum scores produced at report time, segmented by market regime. The reinforcement-learning calibration pass consumes these regime-segmented outcomes to bias future scoring up or down per regime. The system is meant to learn 'which conditions did I overrate or underrate?' rather than to memorize specific trades.

What are the limitations of Arch-AGI conviction scores?

Conviction is a model-derived estimate, not a guarantee. Scores depend on the quality of the input data, the regime classification, and the historical sample available for calibration. Arch-AGI is best used as a structured second opinion to a user's own thesis, not as a stand-alone trading signal.

Is the underlying logic proprietary?

Yes. Treeova publishes the architecture and pass-by-pass behavior of Arch-AGI but withholds the specific prompts used per pass, the model routing per pass, and the conviction-weighting math. This whitepaper is the canonical methodology reference.

Arch-AGI: 7-Pass Conviction Methodology Whitepaper

WP-01 — Arch-AGI: 7-Pass Conviction Methodology

Arch-AGI is Treeova's conviction analysis engine. It runs a structured 7-pass pipeline — edge, scenario, risk/reward, regime, macro, reinforcement-learning calibration, and adversarial stress — to produce a 0–100 conviction score, scenario projections, and a regime-aware learning note for an open or candidate options position.

Authored by Treeova Research· Research CollectiveUpdated 2026-04-18

#1. Overview

Arch-AGI is the conviction analysis layer of Treeova's agentic trading stack. Its job is to take a single open or candidate options position and return a structured second opinion: how strong is the edge, how likely are the base, best, and worst scenarios, what is the regime-adjusted risk/reward, and what does past behavior in similar regimes suggest about the score's reliability.

Arch-AGI does not place trades. It produces a report that the user (or a downstream agent acting under the user's authority) can act on. This separation — analysis as a first-class artifact, execution as a separate gated action — is fundamental to the methodology.

See the product surface for this methodology: Feature page for Arch-AGI.

#2. The 7-Pass Pipeline

A single Arch-AGI run executes seven passes in sequence. Each pass takes the prior pass's structured output as input and returns its own structured contribution. The final pass assembles the report.

Edge. Frames the trade thesis: what asymmetry is being captured, what is the trigger, and what would invalidate it.
Scenario. Projects base, best, and worst-case outcomes for the position with explicit price and time assumptions.
Risk / Reward. Translates the scenarios into a risk/reward ratio, breakeven, and probability-of-profit estimate.
Regime. Classifies the prevailing market regime and asks whether the trade thesis is congruent with that regime.
Macro. Surfaces near-term macro tailwinds and headwinds (earnings, rate decisions, key risk dates).
RL Calibration. Consumes regime-segmented historical outcomes for similar trades and biases the conviction score up or down based on what the system has consistently over- or under-estimated in this regime.
Adversarial Stress. A challenge pass that tries to break the thesis: what assumption is fragile, what is the blind spot, what would the strongest counter-argument be.

#3. Conviction Score Derivation

The final conviction score (0–100) reflects the agreement and quality of the seven passes. A high conviction score requires alignment across edge, scenario, risk/reward, regime, and stress test — disagreement at any pass pulls the score down.

The exact weighting math and the pass-level model routing are proprietary and intentionally withheld from this whitepaper. What the user sees in every report, however, is the per-pass output itself, so the conviction score is auditable from its components even when its arithmetic is not.

#4. Regime-Segmented Learning Loop

Trade outcomes are recorded against the conviction and momentum scores produced at report time, tagged with the market regime observed at that time. Over time this builds a regime-segmented calibration set: the system can detect that, for example, conviction was systematically too high in one regime and too low in another.

The RL calibration pass uses these signals to nudge subsequent conviction scores in the relevant regime. This is a calibration mechanism, not a memorization mechanism — Arch-AGI does not replay specific trades back into future analysis.

#5. What a Report Contains

The user-facing Arch-AGI report contains, at minimum:

The conviction score (0–100) and a one-paragraph rationale.
The base, best, and worst-case scenarios with assumptions.
Risk/reward ratio, breakeven, probability-in-the-money, and theta urgency.
The classified market regime and macro tailwinds / headwinds.
An RL calibration note explaining whether the score was adjusted up or down based on regime-segmented history.
The adversarial stress challenge and the conviction-after-stress score.
A self-correction directive when the stress pass invalidates the thesis.

#6. Limitations

Conviction scores are model-derived estimates, not predictions. They are intended as a structured second opinion, not as a stand-alone trading signal.
The RL calibration pass requires a sufficient regime-segmented sample to produce meaningful adjustments. In novel regimes the adjustment is small or zero by design.
The macro pass depends on the freshness and completeness of the calendar and news data available at report time. Late-breaking events that have not yet propagated will not influence the score.
Past performance — including past calibration accuracy — does not guarantee future results.