LLM ANALYSIS / FILTERING

Model scoring is the last optional filter

ScreenerBot runs deterministic and sourced checks first. Only survivors reach the configured model, which returns a typed pass or reject decision with a 0–100 confidence score and reasoning.

Pipeline position

Enable the real configuration owners

config.toml

[llm]
enabled = true
default_provider = "groq"

[llm.providers.groq]
enabled = true
api_key = "<api-key>"
model = "<model-id>"
rate_limit_per_minute = 30

[llm_analysis]
filtering_enabled = true
min_confidence = 70
fallback_pass = false
use_cache = true
auto_blacklist_enabled = false
auto_blacklist_min_confidence = 90
cache_ttl_seconds = 300

Provider transport

[llm] owns the master switch and the single provider every model-backed feature calls.

Analysis behavior

[llm_analysis] owns filtering, confidence, failure behavior, automatic blacklisting, and the analysis cache.

No fallback chain

There is no per-feature provider list. Configure several clients if useful, but only default_provider is called until you change it.

Decision rules

Pass

A typed pass at or above min_confidence lets the token continue to strategy and entry evaluation. It does not place a trade by itself.

Reject

A typed reject at or above the threshold records an LLM-analysis rejection with its reason, confidence, and provider.

Low confidence or failure

There is no advisory/enforcement mode and no UNCERTAIN pipeline state. Low confidence, malformed output, provider failure, and unknown decisions follow fallback_pass: pass when true, reject when false.

Automatic blacklist

When enabled, a scam result at or above auto_blacklist_min_confidence is persisted through the token blacklist owner. Confidence values are integer percentages from 0 to 100.

Operational guidance

Start fail-closed in paper trading. Keep fallback_pass = false until provider availability and model output are understood; this prevents a provider outage from silently widening admission.
Treat model scoring as one gate. RugCheck and deterministic on-chain checks remain independent evidence; the model is not a replacement for either.
Review decision history. The Assistant page records the model, provider, latency, token usage, confidence, and reason used for each evaluation.
Switch providers deliberately. If the selected provider is unavailable, change llm.default_provider; ScreenerBot does not silently try another provider.

Configuration reference

See every current provider and analysis field.

Troubleshooting

Diagnose selection, rate-limit, output, and cache failures.