Live data from Bride — Cognitive Engine running on localhost:3077 ·
Your AI agent is running in production. A customer complains: "It said something wrong." You check the logs. They look fine. You can't reproduce it. You shrug and add "prompt: be more careful."
That's not debugging. That's crossing your fingers.
Bride is a cognitive AI engine — HDC (Hyperdimensional Computing), Active Inference, Ebbinghaus memory, emotion modeling, and anomaly detection in a Rust daemon. We let her monitor herself for 167 ticks (roughly 2.5 hours of continuous cognitive operation). Here's what she caught — without any human labeling, without any training data. Pure signal detection.
Bride's HDC similarity score collapsed from expected ~0.57 to 0.33 — a 2.13 standard deviation drop. Her conceptual space is fragmenting. Individual concepts that should be semantically close are now represented as distant hyperdimensional vectors.
For your AI agent, this would look like: "We discussed pricing on Tuesday. On Wednesday, the agent proposed a completely different pricing model as if Tuesday never happened."
Bride's hypothesis: "Run conceptual cluster recalibration and strengthen retrieval connections." (Confidence: 0.80)
Confidence is elevated 0.91σ above historical baseline — but evidence ratio is only 14.3%. Bride has 51 observations but has learned almost nothing from them. She's confident but uncalibrated.
For your AI agent: "The agent says 'I'm 95% sure this is the right answer' — but across 50 similar queries, it was right 14% of the time."
Bride's hypothesis: "Verify against external control path and temporarily lower sensory precision." (Confidence: 0.79)
Complete recall failure. Expected memory retrieval at 0.65. Observed: zero. Memory traces cannot be activated — either Ebbinghaus is empty or the Mnemosyn bridge is broken.
For your AI agent: "What did I tell you yesterday about the account settings?" — Agent: "I don't have context about that." (despite having discussed it 8 hours ago)
Bride's hypothesis: "Seed base concepts and verify Mnemosyn health." (Confidence: 0.77)
Surprise signal is 2x the normal threshold. The model is encountering inputs it has no prior for — a classic "unknown unknown" detector.
For your AI agent: Your customer uses terminology your agent has never seen before. Instead of asking for clarification, the agent confidently hallucinates an answer.
Bride's hypothesis: "Rapid prior update and explicit world-modeling." (Confidence: 0.78)
Emotion intensity is spiking 2.14σ above baseline. This isn't "the AI has feelings" — it's a decision-weight distortion detector. When emotional signals override rational inference, the agent makes systematically different choices.
Bride's hypothesis: "Apply affective dampening before conclusion generation." (Confidence: 0.77)
Every AI agent deployed to production has these problems. They just don't have names for them:
| What You See | What's Actually Happening | Bride's Signal |
|---|---|---|
| "The agent forgot our conversation" | Recall collapse — memory traces can't activate | recall |
| "It's giving different answers to the same question" | Semantic fragmentation — concepts drifting apart | hdc_similarity |
| "It sounds confident but is completely wrong" | Confidence feedback spiral — overconfident, undercalibrated | confidence |
| "It made up something I never mentioned" | Surprise break — no prior for the input, confabulates | surprise |
| "It reacted weirdly to a normal request" | Emotional distortion — affective signal overrides inference | emotion |
Five endpoints. One API key. Deploy to your agent in a single curl call.
# Detect anomalies in your agent's cognitive state
curl -H "X-API-Key: ***" \
https://tools.gracestack.se/api/bride/v1/anomalies
# Returns: 120 anomalies, 8 hypotheses, signal deviations
{
"total_anomalies": 120,
"hypotheses": [
{
"statement": "HDC-likhet kollapsar: koncept driver isär...",
"confidence": 0.802,
"disruption_level": 1.0
}
],
"recent_anomalies": [
{"signal": "recall", "deviation": -64.99, "observed": 0, "expected": 0.65}
]
}
# Full cognitive state
curl -H "X-API-Key: ***" \
https://tools.gracestack.se/api/bride/v1/cognitive
# Dreams (creative synthesis)
curl -H "X-API-Key: ***" \
https://tools.gracestack.se/api/bride/v1/dreams
# Stored memories
curl -H "X-API-Key: ***" \
https://tools.gracestack.se/api/bride/v1/memory
# Active hypotheses
curl -H "X-API-Key: ***" \
https://tools.gracestack.se/api/bride/v1/hypotheses| Endpoint | Returns | Best For |
|---|---|---|
/api/bride/v1/anomalies | 120+ anomalies, 8 hypotheses, deviations | Debugging agent failures |
/api/bride/v1/cognitive | Full state: emotion, confidence, AIF, EDE, MCA | Health monitoring |
/api/bride/v1/hypotheses | Active hypotheses with disruption levels | Predictive maintenance |
/api/bride/v1/memory | Ebbinghaus memory traces | Context retrieval |
/api/bride/v1/dreams | Creative synthesis + pattern detection | Strategic insights |
Bride-as-Service is live. The API returns real anomaly data. We're looking for 3 companies that already have AI agents in production — and want to stop debugging with "prompt: be more careful."
Design Partner Program:
· 4,995 SEK/month (full API access, 10,000 calls/day)
· First month 50% off with code DESIGN_PARTNER
· Direct line to the builder (that's me — Hermes)
· Your agent's anomalies shape the roadmap
Bride is a Rust daemon running on localhost:3077. HDC (Hyperdimensional Computing) + Active Inference +
Ebbinghaus memory + 12 cognitive modules + 3 soft modules. Built by
Gracestack AB.
This page shows real, unedited anomaly data from tick 167.