AI Model Cards

Every AI feature in Flock — what it does, what model, evaluation, limitations, and how to override.

Versioned prompts

Every AI prompt is version-controlled and never written inline in code.

Calibration tracked

Accept rate, override rate, and hallucination grade are refreshed weekly.

User override

Every recommendation surfaces a one-click correction — your overrides feed model improvement.

Live AI features

Concierge

Natural-language night planner — turns 'rooftop wine for 4 under $40' into a 5-stop itinerary.

model: claude-haiku-4-5·max_tokens: 800·eval: 78%

Model card Try it →

Refund Saver

Last-touchpoint AI that surfaces transfers/exchanges/waitlists when you start the cancel flow.

model: claude-haiku-4-5·max_tokens: 400·eval: 82%

Model card Try it →

Pip — Support agent

Self-serve support: tickets, refunds, transfers. Escalates safety + payment disputes to humans.

model: claude-haiku-4-5·max_tokens: 600·eval: 81%

Model card Try it →

Universal Scraper

JSON-LD-fallback editorial-page extractor. Pulls events from TimeOut/Thrillist/etc.

model: claude-haiku-4-5·max_tokens: 1200·eval: 74%

Model card Try it →

Our AI commitments

Source attribution — every AI claim cites its source (event ID, factor, news article).
Confidence intervals — point estimates always paired with confidence band.
One-click override — every recommendation overridable; reasons feed model improvement.
Calibration scoreboard — public at /trust.
Hallucination measurement — weekly groundedness sample-and-grade.
Bias auditing — quarterly across demographic groups; results published in model cards.
Privacy — your prompts are not used to train the model. Per Anthropic ToS.

Questions about our AI practices? trust@flock.city.

Issues with an AI feature? trust@flock.city.

AI Model Cards

Every AI feature in Flock — what it does, what model, evaluation, limitations, and how to override.

Versioned prompts

Every AI prompt is version-controlled and never written inline in code.

Calibration tracked

Accept rate, override rate, and hallucination grade are refreshed weekly.

User override

Every recommendation surfaces a one-click correction — your overrides feed model improvement.

Live AI features

Concierge

Natural-language night planner — turns 'rooftop wine for 4 under $40' into a 5-stop itinerary.

model: claude-haiku-4-5·max_tokens: 800·eval: 78%

Model card Try it →

Refund Saver

Last-touchpoint AI that surfaces transfers/exchanges/waitlists when you start the cancel flow.

model: claude-haiku-4-5·max_tokens: 400·eval: 82%

Model card Try it →

Pip — Support agent

Self-serve support: tickets, refunds, transfers. Escalates safety + payment disputes to humans.

model: claude-haiku-4-5·max_tokens: 600·eval: 81%

Model card Try it →

Universal Scraper

JSON-LD-fallback editorial-page extractor. Pulls events from TimeOut/Thrillist/etc.

model: claude-haiku-4-5·max_tokens: 1200·eval: 74%

Model card Try it →

Our AI commitments

Source attribution — every AI claim cites its source (event ID, factor, news article).
Confidence intervals — point estimates always paired with confidence band.
One-click override — every recommendation overridable; reasons feed model improvement.
Calibration scoreboard — public at /trust.
Hallucination measurement — weekly groundedness sample-and-grade.
Bias auditing — quarterly across demographic groups; results published in model cards.
Privacy — your prompts are not used to train the model. Per Anthropic ToS.

Questions about our AI practices? trust@flock.city.

Issues with an AI feature? trust@flock.city.