Preserve Anomalies, Retain Alternate Explanations
Existing systems transmit what they understand. PARALLAX preserves the evidence they do not.
PARALLAX is an onboard evidence-preservation assistant for deep-space missions. When an anomaly-detection system encounters an event it cannot confidently explain, PARALLAX uses Gemma to maintain competing hypotheses and determine which raw records must survive a limited transmission window — so investigators on Earth can still distinguish between explanations after the data is sent.
Deep-space spacecraft operate with:
- Communication delays of 38–90 minutes
- Transmission budgets of 25–50 KB per anomaly report
- Autonomous compression that permanently discards data before mission control can review it
A conventional system flags that something unusual happened. It cannot determine whether the event was a genuine scientific observation, an instrument fault, or corrupted data. A standard compression policy selects the smallest files — often discarding the primary instrument record in favour of thumbnail images.
Information lost during transmission cannot be recovered.
PARALLAX adds a reasoning layer after conventional anomaly detection:
- Detect — Deterministic threshold triggers flag the anomaly
- Retrieve — TF-IDF search pulls relevant passages from instrument manuals, incident procedures, and known event descriptions
- Reason — Gemma generates competing hypotheses, rates each piece of evidence by preservation priority, and warns what compression would destroy
- Preserve — A 0/1 knapsack algorithm selects the highest-value evidence combination that fits within the byte budget
- Compare — The dashboard shows what a naive compression baseline loses vs. what PARALLAX keeps
| Scenario | Event | Key Finding |
|---|---|---|
| Spectrometer Transient — Europa Flyby | 8.7σ spectral spike at T+04:17:32 | Naive compression drops raw_spectrum (14 KB primary data) in favour of a camera thumbnail |
| Solar Array Power Drop — Jupiter Approach | 35% output loss at T+09:42:15 | Naive compression drops solar_iv_curves (segment-level panel data) in favour of star tracker frames |
In both cases, PARALLAX preserves the primary instrument record. The naive baseline loses it.
Synthetic Mission Simulator
↓
Deterministic Anomaly Detector
↓
TF-IDF Context Retriever (5 knowledge documents)
↓
Gemma Reasoning Layer (hosted inference)
→ Competing hypotheses
→ Evidence preservation priorities
→ Compression warnings
↓
Pydantic Validation
↓
0/1 Knapsack Evidence-Capsule Builder
↓
Mission-Control Dashboard (Streamlit)
Gemma handles: hypothesis generation, evidence ranking, uncertainty preservation, compression warnings
Deterministic Python handles: anomaly detection, byte-budget arithmetic, evidence selection, schema validation
git clone <repo>
cd BWG
pip install -r requirements.txt
cp .env.example .env
# Edit .env if your endpoint requires an API key
streamlit run parallax/app.pyWithout an API key or live endpoint, the app falls back to pre-generated cached analyses automatically.
GEMMA_ENDPOINT=https://ai.spuric.com/v1/chat/completions
API_KEY= # optional Bearer token
USE_LIVE_GEMMA=true
MODEL_NAME=gemma-3-27b-it
| Layer | Technology |
|---|---|
| Interface | Streamlit |
| Model | Gemma via hosted OpenAI-compatible endpoint |
| Retrieval | TF-IDF (scikit-learn) |
| Validation | Pydantic v2 |
| Charts | Plotly |
| Data | Pandas / NumPy |
Track 2: Trajectory & Orbit (Deep Space Navigation)
PARALLAX demonstrates that an LLM reasoning layer can meaningfully improve evidence preservation decisions under bandwidth constraints — without replacing deterministic safety systems or performing unreliable numerical calculations.