Ask the market. Every number is computed, not guessed.

Give it an entry, a stop and a target. Before any model reasons about your thesis, five deterministic checks run in a fixed order. A failure is a rejection with a machine-readable reason tag — not a softened suggestion, and not a silently resized position.

A stop floor no profile can relax

Your stop distance must be at least 1.5× ATR(14). That constant holds for intraday, swing, investor and F&O alike. A stop sitting inside normal market noise is rejected, not renegotiated.

Reward-to-risk floors that fit the session

1:2 for swing, investor and F&O. 1:1.3 for intraday — because a swing-calibrated floor makes a defensible intraday bracket arithmetically impossible and yields nothing but perpetual holds.

Then something argues against you

A Bear agent runs against your own proposal, hunting for overhead VWAP capping the move, call open-interest walls above your target, volume-profile gaps and unfavourable session timing. It returns a critique and nothing else — it has no authority to decide.

Rejections you can act on

Eight stable failure tags, not prose: missing levels, inconsistent direction, stop too tight, reward-to-risk too low, leg fraction out of range, target ordering inconsistent, breakeven out of range, blended reward-to-risk too low.

Written twice, on purpose

The validator exists in Rust and again in Python with identical constants and identical reason tags. The arithmetic is not something a language model can talk its way around.

Multi-leg exit plans are validated as a whole: leg fractions must sum correctly, targets must order correctly, the breakeven trigger must sit between entry and first target, and the blended reward-to-risk must still clear the floor.

Most terminals draw one forward line and let you assume it means something. Strat AI runs eight projection engines, four of them selectable directly on the chart, each using genuinely different mathematics. Switch model and the line changes — which is the entire point. A projection you cannot cross-examine is decoration.

OLS — unweighted least squares

A straight line fitted across a fifty-bar window, then re-anchored onto the last close. The baseline case, where every bar counts the same regardless of what traded.

VWLR — volume-weighted least squares

The same straight-line model, weighted by traded volume with weights floored at one, so bars where size genuinely changed hands pull the fit harder than quiet ones do.

VWEPR — volume-weighted quadratic

A curve rather than a line, solved by Gaussian elimination with partial pivoting and falling back to OLS if the system turns out singular. Quadratic and not cubic on purpose — a cubic flies off the screen. Its second-order term is surfaced as an acceleration coefficient.

FCST — regime-conditioned forecast

Not a regression at all. An exponentially weighted drift over log returns, then conditioned on the measured regime: amplified in a trending market, damped in a ranging one. It reports a direction, an up-probability, an expected move in ATR units and its own confidence.

The fifty-bar window is pinned to the same constant the agent’s own tools use, so the projection you see and the projection it reasons about cannot drift apart. R-squared is reported only by the dedicated predictive model that actually computes it, and only on the ten-minute chart it was calibrated for.

A separate service watches ten-minute candles for a two percent absolute move. When one fires it asks a model for a cause, then publishes a headline, a written read and a sentiment assessment straight into the terminal — shaded harder once the move clears three percent.

Distinct from the news scorer

This is not headline sentiment wearing a different label. It triggers on price behaviour first and reaches for explanation second, which is why it can surface a move that has not reached a feed yet.

Its own structural pass

The same service runs a hundred-candle pattern engine against a pinned four-field contract: type, sentiment, description, and a confidence clamped between zero and one.

No cause, no story

When nothing explanatory can be found, it reports exactly that. It will not assemble a plausible narrative to keep the panel looking populated.

Pushed, not polled

Insights are broadcast to the terminal as they are produced, so the commentary lands while the candle still matters.