Methods · policy set v0.2.1 · live

The rules HuntFish owns, in the open.

Published methods. Versioned, tested, replaced when measured evidence supports a better method. Not validated science.

Source hierarchy v0.1.0

When sources disagree, the higher rank wins. Corroboration: Multiple sources repeating one announcement are not independent corroboration. Abstention: When reliable current evidence cannot be obtained for a rule-dependent question, say so and give the official next step. A generic disclaimer is not a substitute.

Rules and access

  1. official authority — e.g. state wildlife agency rule pages; federal land manager pages; official emergency notices. Property-specific restrictions and emergency notices are reconciled with general rules.
  2. official dataset — e.g. USGS PAD-US public access field. Public ownership is not permission for every activity.
  3. user private observation. Most relevant for activity on the user's own ground; establishes neither law nor access.
  4. credible reference — e.g. research; expert references
  5. community report. Labeled as reported; never overrides an applicable current official rule.

Habitat and technique

  1. credible reference
  2. user private observation
  3. community report — Well-labeled local observations and discussion.

Routing v0.2.1

Use the cheapest configured route that meets the task's quality, evidence and privacy requirements. Retrieve current evidence when a claim depends on current conditions or rules. Never call several models for a basic question. A fast answer may not skip essential source verification.

Classification: stage 1 is deterministic keyword rules, run on our server before any model is called; stage 2 is claude-haiku-4-5 structured classification, max 600 in / 60 out tokens, only when stage 1 is ambiguous. Default task class: chat_simple.

What runs today: stage 1 only (stage 2 is planned), official pages from Source Watch, and bounded web search on research routes. The NWS forecast, USGS water and PAD-US land tools listed below are planned and not called yet — they are priced at $0, so they never change a quote.

RouteWhenBounded planEvidence
classifyTask classification only.claude-haiku-4-5: ≤600 in / ≤60 outnot required
chat_simple
escalates to chat_grounded
General technique/gear/behavior questions with no legal, safety or current-conditions dependency.claude-sonnet-5: ≤5,000 in / ≤1,600 outnot required
chat_grounded
escalates to chat_research
Questions that depend on rules, access or current conditions. Must retrieve evidence or abstain.claude-sonnet-5: ≤8,000 in / ≤1,500 out; tools: source_watch_lookup ×4, nws_forecast ×2, usgs_water ×2, padus_lookup ×2required
chat_researchMulti-source research inside a conversation. Its bounded plan costs at most 18 credits. A research question is held to your per-Scout maximum (25 credits by default) instead of the 10-credit per-question maximum; if it would need more, HuntFish asks you to narrow it or raise that maximum.claude-sonnet-5: ≤10,000 in / ≤2,000 out; tools: anthropic_web_search ×2, source_watch_lookup ×6, padus_lookup ×4required
photo_identifyIdentification from a metadata-stripped, task-appropriate crop. Consent required before transfer.claude-sonnet-5: ≤4,000 in / ≤1,200 outnot required
scout_researchOne bounded, checkpointed Scout research step under the Scout's reservation. Each step costs at most 18 credits and is held to your per-Scout maximum (25 credits by default).claude-sonnet-5: ≤10,000 in / ≤2,000 out; tools: anthropic_web_search ×2, source_watch_lookup ×10, padus_lookup ×10, nws_forecast ×4required
sandbar_scoutSandbar Scout: one bounded pass that reads free public map candidates (OpenStreetMap, NOAA charts, USGS names), up to 5 public-domain aerial photos and at most 2 web searches for local reports, then picks the best sandbars with a calculated time window. Held to your per-Scout maximum (25 credits by default); nothing is charged when the free map pass and the around-you photo give nothing to check.claude-sonnet-5: ≤18,000 in / ≤2,500 out; tools: anthropic_web_search ×2, osm_overpass ×1, noaa_enc ×3, usgs_gnis ×2, naip_imagery ×5, noaa_tides ×1, nws_forecast ×2required

Freshness v0.1.0

Evidence labels: observed, forecast, calculated, reported, inferred, regulatory. Never show a universal 'verified today' label when only part of an answer was refreshed today; each evidence card carries its own retrieved_at. Evidence older than its recheck interval is shown with its age and reduces evidence strength; it is never silently dropped or silently trusted. Access and regulation evidence attached to a Trip Card is rechecked within 24h before the trip date.

Evidence kindRechecked at least every
regulation_page7d
emergency_notice24h
public_access_map30d
acquisition_announcement7d
nws_forecast1h
nws_observation20m
usgs_water1h
noaa_tides6h
padus180d

Privacy defaults v0.1.0

Plain-English version on the Trust page. The machine-readable defaults the app and connector enforce:

{
  "version": "0.1.0",
  "permissions": {
    "save_privately": {
      "default": "explained_in_feature",
      "meaning": "Keep the requested account data for the user's use."
    },
    "external_ai_processing": {
      "default": "ask_first",
      "meaning": "Send the necessary content to named approved providers.",
      "providers_at_launch": [
        "anthropic"
      ]
    },
    "share_with_community": {
      "default": "off"
    },
    "contribute_to_training": {
      "default": "off"
    }
  },
  "location": {
    "signup_requires_precise_gps": false,
    "default_precision": "approximate",
    "approximate_grid_km": 5
  },
  "photos": {
    "strip_metadata_on_derivatives": true,
    "originals_immutable": true,
    "send_crop_not_original_when_sufficient": true
  },
  "logging": {
    "never_log": [
      "prompts",
      "precise coordinates",
      "tokens",
      "signed photo URLs",
      "payment data"
    ],
    "session_replay_on_private_surfaces": false
  },
  "connectors": {
    "default_scopes": [
      "field.read",
      "credits.read"
    ],
    "precise_coordinates_shared": false,
    "original_photos_shared": false,
    "grant_default_expiry_days": 30,
    "default_spending_allowance_micro": 0
  },
  "community_aggregation": {
    "live_private_camera_heatmap": false
  }
}

Pricing v2026-09-23.1

Retail usage = 3.0x eligible metered provider cost. Rounded to micro-USD once per job. Rounding: half up at job level. Provider list prices as published on 2026-09-23; re-verify at implementation. Effective negotiated rates replace list rates when they apply.

ProviderModel / toolUnitList rate
anthropicclaude-haiku-4-5input token$1.00 / M
anthropicclaude-haiku-4-5output token$5.00 / M
anthropicclaude-haiku-4-5cache write token$1.25 / M
anthropicclaude-haiku-4-5cache read token$0.10 / M
anthropicclaude-sonnet-5input token$2.00 / M
anthropicclaude-sonnet-5output token$10.00 / M
anthropicclaude-sonnet-5cache write token$2.50 / M
anthropicclaude-sonnet-5cache read token$0.20 / M
anthropicclaude-opus-5-5input token$4.00 / M
anthropicclaude-opus-5-5output token$20.00 / M
anthropicclaude-opus-5-5cache write token$5.00 / M
anthropicclaude-opus-5-5cache read token$0.20 / M
toolanthropic_web_searchcall$0.01 each
toolnws_forecastcall$0 each
toolusgs_watercall$0 each
toolpadus_lookupcall$0 each
toolsource_watch_lookupcall$0 each
toolnoaa_tidescall$0 each
toolosm_overpasscall$0 each
toolnoaa_enccall$0 each
toolusgs_gniscall$0 each
toolnaip_imagerycall$0 each

Retail usage = eligible provider cost × 3. Credits are display only: 1 credit = 10,000 micro-USD = one cent. Customer pricing.