Method

Never-claim

We never treat a species list as ecological condition, and we never use “recovered” as a label without persistence evidence.

Presence is not persistence. Habitat is not occupancy. Non-detection is not absence. A long species list is not ecological condition. Satellite greening is not faunal recovery.

Change with effort

Relative detection frequency is computed from a fixed survey recipe (20-minute 2-ha, four visits, ~8 hours per site per year). Like-for-like seasonal windows only. A decline in records is not interpreted as biological decline unless effort is stable or modelled. Occupancy language is withheld unless detection probability is addressed.

Confidence

Public labels: High, Moderate, Low, Insufficient evidence. Internally a 0–1 score considers source quality, identification confidence, observation effort, temporal consistency, independent corroboration, method suitability, geographic precision, and model uncertainty. Confidence is shown separately from trajectory so a fast scan cannot conflate “declining” with “we are sure”.

Intelligence chain

Entity / System → Observation → Evidence / Source → Provenance → Model / Method → IntelligenceObject → Validation → Confidence → Permissions → Decision → Outcome → Learning

Agents

Logical specialist roles, event-driven or scheduled — not continuously running. Last seed run:

• observation — Ingested demonstration surveys, acoustic index, camera event, eDNA detection, two Swift Parrot records.
• taxonomy — Mapped Themeda australis → T. triandra and Olea europaea → subsp. cuspidata without rewriting originals.
• effort — Survey hours ~8/site/year across 2018–2026. Decline is not explained by reduced effort.
• change — Woodland-dependent relative detections down ~31%. Sittella non-detection flagged as signal.
• evidence — Acoustic index corroborates direction at Castlereagh. Citizen records do not independently prove assemblage decline.
• ecology — Miner dominance + edge + olive as a coherent ecological hypothesis, not a closed cause.
• critic — Rainfall variability and restoration lag remain live alternatives.
• provenance — Demonstration source labelled MOCK. Live GBIF/iNat adapters are separate and do not overwrite this object.
• validator — Confidence 0.71 moderate. Direction stronger than attribution.
• sentinel_liaison — Published relationship to LandSentinel fragmentation and restoration-cover objects. Queued for Meta Sentinel.