AI and Forecast Disclosure

The live Opportunity Score

The live score is a hybrid ranking system. It combines curated habitat and access information, public seasonal data, tides, weather and marine conditions, daylight and moon context, and practical fishability adjustments. The result is converted to a 0–100 percentile among the current location and time-window candidates.

It is not a catch probability, a statement that halibut are present, or a promise that the water is safe or practical to fish. Public inputs can be wrong or unavailable. Trip reports may inform product review, but they do not validate or train the current score.

A separately versioned water-quality advisory overlay can remove an exactly mapped site from recommendations when the official agency reports an active water-contact status. A no-posting result never increases the Opportunity Score. Missing, stale, unmonitored, unavailable, and unmapped status stays unknown, and neither the overlay nor the score establishes water-contact or seafood safety.

Validation evidence today

As of July 19, 2026, CastingCompass has not activated a prospective validation study. The current eligible prospective or confirmatory validation sample therefore contains 0 attempts: 0 eligible target encounters and 0 eligible target non-encounters. Existing trip reports, including catches and skunks, remain product feedback or descriptive context rather than model-evaluation evidence.

The live heuristic has completed 0 preregistered baseline comparisons and 0 probability-calibration runs. No ranking-accuracy, calibrated catch-probability, or model-promotion claim is supported.

The current published negative research result is separate: a geographically held-out seafloor-character probe reached macro F1 0.3914 and was reliably worse than simpler classical structure summaries, so it was not promoted. That experiment measured terrain representation, not California-halibut catch ranking, and it is not evidence about the live Opportunity Score.

Deep-learning research status

CastingCompass has a bathymetry pretraining and model-research pipeline intended to learn useful underwater terrain representations. Unless the product expressly labels a deployed model version as trained and validated, that research output is not the live Habitat Score. The live product will not claim a deep-learning accuracy improvement until geographically blocked evaluation shows a reliable ranking gain over simpler baselines.

Trip-note and gear review

When a signed-in angler completes or edits a trip report, Xiaomi MiMo may review a limited trip payload. It normalizes recognizable rod, reel, lure, and rig names; flags inconsistent or unsafe content; and may prepare a pseudonymous discussion draft. A queued review is authorized only after a final deletion-record check. Deletion committed before that point cancels the request; a provider request already authorized or sent before deletion cannot be recalled, although its response cannot restore the deleted trip or publish a post. The model cannot publish or approve the draft. A human moderator must approve it before it can appear publicly.

Automated review can misunderstand a note or gear product. It does not determine legal guilt, eligibility for a benefit, employment, credit, health care, housing, or another high-impact decision. Users can request correction or removal by contacting bzeng0000@gmail.com.

Data sent for review

The review can receive the curated site and type, trip time, method, effort and catch counts, gear entries, observed fishability, forecast/model context, and up to 1,000 characters of notes. CastingCompass deliberately omits the user’s email, internal account ID, uploaded photo, and structured browser-location or coordinate fields. Free-text trip fields are sent as entered and could contain details the user typed.

How to interpret explanations

Component scores and explanation text describe the inputs that influenced the ranking. They are not proof of causation. Conditions such as tide and weather can be correlated with each other, and a high score can still produce a skunk. Trip reports do not change the current score or enter model evaluation automatically; any future use requires the separate validation protocol.

Regulatory approach

There is no single universal “FTC AI disclosure” form. CastingCompass follows the core consumer-protection principle that claims about automated systems must be truthful, supported, and not misleading. This page, the score explanation, source freshness labels, and the Terms are designed to make the present limitations conspicuous.

California generative-AI training-data and synthetic-media provenance laws may become relevant if CastingCompass later develops or publicly releases covered generative systems. The current opportunity ranker and private note-review workflow are not presented as a general-purpose generative-AI service. This assessment will be revisited when features change.