The Token Report

Every report opens with three main sections — a plain-language summary (what's happening, why it matters, what to watch), a conviction score, and a verified timeline of real events. Then nine deep-analysis sections follow:

Each section returns a one-line summary, a one-word status (Strong / Stable / Moderate / Weak / Risky / N/A), and a set of labeled metrics with notes. Smart-money and social-sentiment are richer — they carry wallet movements and social mentions respectively.

The conviction score isn't a single gut number. The analysis returns an overall score (0–100), a risk tier, a confidence value, a one-sentence verdict, and a breakdown of category scores. The overall is recomputed as the mean of its categories so the headline can never disagree with the parts that produced it.

Category scores are honest-checked: if a category's explanation contradicts its number (e.g. a high score with words like "stagnant", "dead", "illiquid"), it's clamped. The result is a conviction you can actually trace to its evidence.

Risk tiers, from safest to direst:

A token can score well on momentum and still be a trap. A dedicated safety pass runs ahead of the deep analysis — checking contract permissions (mint/freeze/upgrade authority), honeypot risk, tax mechanisms, and transfer restrictions — and surfaces as a safety banner on the report.

A high-conviction token with an obvious red flag is never published as a clean buy; the conviction prompt is explicitly instructed to never reward tokens with obvious red flags.

Every report includes an explicit bull case and bear case — the strongest argument for the upside and the strongest argument against, side by side. No report publishes a one-sided thesis. The bear case is treated as a first-class part of the analysis, not a disclaimer.

The raw overall conviction is an opinion. On the report, it's paired with a calibrated confidence — the realized probability the token-analysis engine has earned at that conviction level, from its fitted model. This is the same calibration layer described earlier, applied to the token-analysis engine specifically.

Because token analysis is the newest engine, its calibrated probability may show a "warming up" state while it inherits a donor engine's model as a cold-start prior, then refines to its own outcomes as predictions resolve. The published probability always carries a confidence interval so you can see how tightly the model knows that number.

The report is anchored to authoritative live metrics from DexScreener — price, volume, market cap, liquidity, holders, age — treated as ground truth that web sources cannot contradict. The prompt forbids inventing or rounding numbers, and the timeline is restricted to real, verifiable events within the token's actual lifespan (no future-dated milestones).

When data is genuinely unavailable, the report says so — "N/A" or an explained gap — rather than fabricating a clean-looking answer. Smart-money and social sections never invent wallet addresses, transactions, or social handles.