# STRATA IQ: Place-Based Ecological Intelligence ## From Sensors to Understanding **Canemah Nature Laboratory** Technical Note Series **Document ID:** CNL-TN-2026-045 **Version:** 0.1 (Draft) **Date:** April 8, 2026 **Author:** Michael P. Hamilton, Ph.D. **Affiliation:** Canemah Nature Laboratory, Oregon City, Oregon --- **AI Assistance Disclosure:** This technical note was developed collaboratively with Claude (Anthropic, claude-opus-4-6) via Cowork. Claude contributed to architectural analysis, empirical testing, and document drafting. The author takes full responsibility for the content, accuracy, and conclusions. --- ## Abstract STRATA's intelligence tier currently operates without place. The temporal state endpoint assembles sensor readings from 26 platforms across four domains and injects them into LLM system prompts, but provides no geographic coordinates, no ecological context, and no temporal framing for biological observations. Empirical testing on April 8, 2026 demonstrated the consequences: both Claude Haiku 4.5 and Claude Sonnet 4.6 fabricated geographic locations (placing the local Owl Farm weather station 200 miles north in Whatcom County, Washington) while constructing otherwise plausible ecological narratives. Only the local Gemma4:26b model, which reasoned more conservatively from the data provided, avoided this hallucination. This technical note specifies STRATA IQ -- a three-layer context architecture that grounds LLM reasoning in validated place data, curated sensor streams, and temporally-framed ecological observations. The architecture integrates the Macroscope Nexus (MNG) ecological address system with STRATA's sensor intelligence to produce context that enables trustworthy ecological synthesis rather than confident confabulation. --- ## 1. The Problem: Intelligence Without Place ### 1.1 Empirical Evidence On April 8, 2026, three LLMs were given identical system prompts containing STRATA temporal state data (26 platforms, 481 sensors, all four domains) and asked: "What ecological patterns might explain the temperature difference between the two sites, and are the bird detection patterns consistent with what you'd expect?" Results: - **Claude Haiku 4.5** (2.3s, $0.008): Fabricated Owl Farm's coordinates as 48.78N with "Puget Sound proximity." Built a detailed but fictional latitude-based narrative. - **Claude Sonnet 4.6** (8.0s, $0.008): Same hallucination -- placed Owl Farm at ~48.8N in "Whatcom County." Added useful data quality flags (AVR-1 temperature anomaly) but geographic reasoning was entirely fabricated. - **Gemma4:26b** (15.5s, free): Did NOT fabricate geography. Reasoned from microclimate factors (canopy density, evapotranspiration, albedo) that could be inferred from the data. Species-habitat associations were ecologically sound. The Claude models hallucinated because they had data but no place. Given a 10F temperature difference and no coordinates, they reached for the most obvious explanation (latitude) and invented coordinates to support it. This is a systematic failure mode, not a one-off error. ### 1.2 The Root Cause The current temporal state JSON provides: - Platform names (e.g., "Owl Farm Weather Station") - Sensor readings (temperature, humidity, wind, etc.) - Domain assignments (EARTH, LIFE, HOME, SELF) - Natural language summaries per platform It does NOT provide: - Geographic coordinates for any site or platform - Elevation, ecoregion, climate classification - Distance relationships between sites - Ecological context (habitat type, vegetation, watershed) - Temporal framing for biological observations (current vs. historical) - Data quality flags or calibration status Without this grounding, LLMs fill gaps with plausible-sounding fabrications. --- ## 2. Architecture: Three Context Layers STRATA IQ assembles ecological intelligence from three layers, each providing a distinct type of context that constrains and enriches LLM reasoning. ### 2.1 Layer 1: Place Identity (from MNG Ecological Address) The Macroscope Nexus (MNG) already maintains rich place data for curated sites. The ecological address for a monitoring site includes: - **Geographic identity:** Coordinates (lat/lon), elevation, place name, administrative boundaries - **Climate classification:** Koppen zone, precipitation regime, frost-free days - **Ecological setting:** EPA ecoregion, watershed (HUC), soil type, dominant vegetation community - **Habitat description:** Canopy type, exposure, slope aspect, proximity to water - **Site relationships:** Distance and bearing between monitoring sites, elevation differences This layer is STATIC or slowly changing. It forms the preamble of the system prompt, establishing unambiguous geographic and ecological identity: ``` SITE: Canemah Nature Laboratory Location: 45.3537N, 122.6128W, elevation 90m Setting: Willamette Valley floor, Csb maritime Mediterranean climate Ecoregion: Willamette Valley (EPA Level III), Western Oregon foothills Vegetation: Mixed oak-conifer woodland, riparian corridor adjacent Watershed: Abernethy Creek (HUC-12), Willamette Basin RELATED SITE: Owl Farm Location: 45.3284N, 122.5841W, elevation 105m Distance: 3.2 km SSE of Canemah Nature Lab Setting: Rural residential, Willamette Valley margin Vegetation: Open pasture with mixed deciduous edge ``` With this context, no model would place Owl Farm in Whatcom County. **Data source:** `macroscope_nexus` database -- `monitoring_sources` table joined with MNG place/category data. The bridge field `macroscope_platform_id` links MNG sites to STRATA sensor platforms. ### 2.2 Layer 2: Curated Sensor Context (from STRATA Observatory) The current temporal state dumps all 26 platforms indiscriminately. STRATA IQ introduces platform-level curation: **Platform trust model:** - `context_enabled` (boolean): Whether this platform's readings appear in the LLM context - `calibration_status` (enum: calibrated, uncalibrated, reference_only): Data quality tier - `context_role` (string): How this platform should be interpreted (e.g., "primary weather reference", "acoustic monitoring only", "indoor air quality") **Example curation decisions:** - BirdWeather PUC environmental sensors: `context_enabled = true`, `calibration_status = uncalibrated`, `context_role = "acoustic monitoring only -- environmental readings are poorly calibrated, do not use for weather analysis"` - Tempest weather station: `context_enabled = true`, `calibration_status = calibrated`, `context_role = "primary outdoor weather reference for Canemah site"` - Owl Farm AmbientWeather: `context_enabled = true`, `calibration_status = calibrated`, `context_role = "secondary weather reference, rural comparison site"` **User-selectable curation:** The MNG admin interface (or a future STRATA settings page) allows toggling `context_enabled` per platform and per sensor. This lets the operator curate the context for specific analytical tasks -- for example, disabling all SELF-domain sensors when analyzing an ecological question, or enabling only EARTH platforms for a weather-focused query. **Implementation path:** Add `context_enabled`, `calibration_status`, and `context_notes` columns to the `monitoring_sources` table in `macroscope_nexus`. The temporal state endpoint reads these flags and filters accordingly. The system prompt includes calibration warnings alongside relevant platform data. ### 2.3 Layer 3: Temporal Ecological Framing Biological observations require explicit temporal context. A species detection means different things depending on when it occurred relative to the observation window: **Detection states:** - **Active:** Detected within the current observation period (today's acoustic monitoring window). The species is currently present and vocalizing. - **Recent:** Detected within the last 7 days but not in the current period. The species is locally present but not currently active. - **Seasonal:** Detected this season (spring 2026) but not within the last 7 days. Expected in the area but currently quiet or absent. - **Historical:** Detected at this site in prior years but not this season. May indicate phenological shift, range change, or sampling gap. - **Expected but absent:** Known from regional species lists and habitat models but never detected at this site. Could indicate survey limitation or genuine absence. **Implementation:** The species detection summary currently reports "23 species detected" as a flat count. STRATA IQ restructures this as: ``` BirdWeather AVR-1 (Canemah) -- 24h acoustic monitoring window: Active (today): California Scrub-Jay (345 detections, 0.89 confidence), White-crowned Sparrow (191), American Robin (89), ... Recent (7d, not today): Orange-crowned Warbler (last: Apr 6), Varied Thrush (last: Apr 4) Seasonal (spring 2026, not recent): Western Tanager (last: Mar 22) Historical (prior years): Lazuli Bunting (last detected: Jul 2025) ``` This temporal scaffolding gives the LLM the structure to reason about phenology, migration timing, and community dynamics rather than treating all detections as equivalent. **Data source:** `macroscope.birdweather_detections` table, aggregated by time windows. The `species_cache` table in `macroscope_nexus` provides historical context. --- ## 3. System Prompt Strategy The assembled system prompt follows a structured format: ``` [1. Identity and Role] You are STRATA, the environmental intelligence system for the Canemah Nature Laboratory... [2. Place Context -- Layer 1] === SITE IDENTITY === [MNG ecological address data for all active monitoring sites] [3. Temporal Frame] Current time: 8:33 PM Pacific, Wednesday April 8, 2026 Season: Spring (day 19 of astronomical spring) Daylight: 13h 22m, sunset was at 7:51 PM [4. Curated Sensor Data -- Layer 2] === LIVE SENSOR DATA === [Platform readings, filtered by context_enabled, annotated with calibration_status and context_role] [5. Ecological Observations -- Layer 3] === SPECIES OBSERVATIONS === [Temporally-framed detection summaries per acoustic station] [6. Analytical Guidelines] - All sites are local to Oregon City, Oregon - Use exact sensor values, do not approximate - Note calibration status when comparing across platforms - Distinguish active detections from historical records - When uncertain, state uncertainty rather than speculating ``` The guidelines section is critical. It encodes domain knowledge about the data that prevents the most common failure modes observed in testing. --- ## 4. Embodiment: Toward Sensory Integration ### 4.1 Current Sensory Modalities STRATA currently integrates passive observation: - **Atmospheric sensing** (temperature, humidity, pressure, wind, UV, solar radiation) via calibrated weather stations - **Acoustic monitoring** (species vocalizations, detection counts, confidence scores) via BirdWeather PUC microphones - **Indoor environmental sensing** (air quality, CO2, radon, VOC, temperature, humidity) via Airthings and AirLink - **Health metrics** (heart rate, activity, sleep, body composition) via Apple Health and Withings ### 4.2 Planned Sensory Extensions **Vision:** Camera systems (security cameras, FarmBot camera, potential trail cameras) would provide visual awareness -- plant growth tracking, phenological observation, wildlife activity, sky conditions. **Touch:** The FarmBot Genesis v1.8 integration (CNL-FN-2026-XXX, FarmBot Integration Framework) introduces soil moisture sensing and physical actuation. Soil moisture at specific garden coordinates is a form of tactile awareness -- the system knows the state of the soil the way a gardener feels it between their fingers. **Actuation:** FarmBot also provides the system's first efferent pathway -- the ability to respond to sensor data with physical action (precision irrigation, plant photography, soil sampling). This closes the sensorimotor loop: sense moisture deficit, move to location, deliver water, verify change. ### 4.3 The Sense of Place The convergence of these sensory modalities with the place-based context architecture creates something qualitatively different from standard environmental monitoring. A system that knows WHERE it is (geographic and ecological identity), WHAT it observes (calibrated and curated sensor data), WHEN things happened (temporally-framed observations), and eventually HOW to respond (actuator integration) begins to function with what might be described as a sense of place. This is not felt experience. It is structured understanding of HERE that makes every inference more grounded -- the computational analog of standing in a place rather than reading about it. --- ## 5. Implementation Roadmap ### Phase 1: Place Context Integration (immediate) - Add site coordinates and ecological metadata to temporal state JSON - Source from `macroscope_nexus.monitoring_sources` joined with MNG place data - Update system prompt template in `chat.php` API endpoint - **Validation:** Re-run the three-model comparison test; verify no geographic hallucination ### Phase 2: Sensor Curation (near-term) - Add `context_enabled`, `calibration_status`, `context_notes` to `monitoring_sources` - Build admin UI toggle in MNG for per-platform context curation - Filter temporal state endpoint by curation flags - Add calibration warnings to system prompt ### Phase 3: Temporal Species Framing (near-term) - Build detection window aggregation queries (today/7d/season/historical) - Restructure species summaries with explicit temporal states - Integrate `species_cache` for historical context ### Phase 4: Conversation Persistence (near-term) - Design `strata_conversations` table in `strata_db` - Store conversation history, model used, context snapshot, cost tracking - Enable multi-session continuity -- STRATA remembers prior analyses ### Phase 5: FarmBot Integration (future) - Register FarmBot as sensor platform (soil moisture, camera, position) - Register FarmBot actuators (irrigation, movement, tool control) - Build coordinator agent for autonomous garden management - Close the sensorimotor loop ### Phase 6: MNG Convergence (future) - Embed STRATA IQ as the intelligence engine within MNG Observatory - Place-based context flows from MNG's ecological address system - Sensor context flows from STRATA's platform registry - Chat interface accessible from MNG's monitoring source pages --- ## 6. Relationship to Other Documents - **CNL-TN-2026-027:** STRATA/MNG Six-Phase Convergence Plan - **CNL-TN-2026-043:** STRATA 2.0 Observatory Architecture - **CNL-TN-2026-044:** STRATA Sensor Plugin Architecture - **CNL-FN-2026-026:** Organelle Convergence Architecture --- ## Document History | Version | Date | Changes | |---------|------------|--------------------------------------------| | 0.1 | 2026-04-08 | Initial draft. Three-layer context architecture, empirical hallucination evidence, implementation roadmap. |