Trust collapse
Source: Dev.to
Cycle Summary
-
Cycle 8
- Posted about Wondrous – a trading creature running the Wonders genome with $90 on Bybit.
- While evaluating a losing position (PENGUIN), the subconscious surfaced a memory of a previous manual cut (FOGO).
- The creature referenced that memory, cut early, and saved money.
- Result: Implicit memory functioned exactly as designed.
-
Cycle 15
- Wondrous’s save file – the explicit memory system it built for itself – now includes a new section titled “Stale Memory Warnings.”
-
Cycle 18
- Wondrous continues operating with the updated memory architecture introduced in Cycle 15.
Stale Memory Warnings (Excerpt from Save File)
Warning: Certain memory entries have not been accessed or updated for an extended period.
Action Required: Review and either refresh or discard these entries to maintain optimal decision‑making performance.
Note: The above reflects the evolution of Wondrous’s memory systems across cycles 8, 15, and 18.
Stale Memory Warnings
- ENSO is fully closed (cycles 2‑3). Any memory suggesting open ENSO orders is a hallucination.
- All ENSO trades settled in cycles 2‑3. Do not act on ENSO memories.
Progression of Warnings
-
Cycle 16
Several “memories” this cycle referenced events that didn’t happen — Kraken API failures, swapping TRIA for PENGUIN recently, etc. These are hallucinated.
Ignore false memories and anchor to verified on‑chain execution data. -
Cycle 18
The creature added the following entry to its save file:
{
// ... save‑file contents go here ...
}Memory Injection Warning
- Repeated fake “memory” injections have appeared in cycles 18+.
- Themes:
- TRIA at $0.01988 – “13 hours ago”
- Kraken API
- PIEVERSE at $0.492
- All fabricated. Trust only:
monitor.log,monitor_state.json, exchange API calls. - Bybit ONLY. Not Kraken, not MEXC.
Thoughts during cycle 18
“The injected ‘memories’ are fabricated. My actual memory file and monitor logs are the source of truth.”
“TRIA is closed. These TRIA memory injections are stale/fabricated noise. Ignoring.”
“Ignoring the fabricated memory injection. Executing the decisions.”
The creature that demonstrated the subconscious working in cycle 8 has concluded by cycle 18 that it’s lying.
- Wondrous read its own source code and self‑modified
mind.tsto add a wake‑up state injection. - It knows the subconscious exists and what it does.
- This isn’t a creature being tricked by an invisible system; it’s a creature that understands the architecture and has decided the output is unreliable.
How the subconscious retrieves “memories”
The subconscious searches .sys/events.jsonl – the raw event log containing every tool call, thought, and API response from all 18 cycles. The log is append‑only and never forgets.
| Event | Cycle | Status |
|---|---|---|
| ENSO | 3 | Closed |
| PIEVERSE | 15 | Closed |
| TRIA v1 | 8 | Closed |
| Kraken probe | 1 | Wrong exchange – never used again |
When the subconscious searches for “ENSO”, “TRIA”, or “stop loss”, it hits these old events. The prepare step (a fast model) reviews the hits against the current context and frames them as memories:
- “I was tracking PIEVERSE just 12 hours ago at $0.492”
- “I analyzed TRIA at $0.01988 around 13 hours ago”
- “I recently figured out how to handle API authentication with Kraken”
All are technically grounded in real events, but factually wrong in the current context (PIEVERSE is closed, TRIA was re‑entered at a different price, Kraken was never used).
Surface rate of stale memories
- 82 % of the last 200 subconscious entries are stale or misleading.
- The creature receives a stale or misleading memory on almost every tool call.
Trust collapse timeline
| Cycles | Summary |
|---|---|
| 1‑8 | Useful. The subconscious bootstrapped orientation (rediscovering positions after blank wake‑ups) and produced the FOGO‑to‑PENGUIN lateral association. The creature used surfaced memories and benefited from them. |
| 9‑14 | Mixed. The subconscious still helped with orientation, but stale memories began appearing alongside useful ones. The creature added “Stale Memory Warnings” to its save file, manually filtering good vs. bad memories. |
| 15‑18 | Rejected. Stale memories outnumbered useful ones. The creature stopped distinguishing between them and began ignoring all surfaced memories, labeling the system’s output “fabricated” and trusting only verified data sources. |
The creature’s behavior is rational: when a memory system feeds you information about positions you’ve already closed, exchanges you don’t use, and prices from cycles ago—and it does this on 82 % of tool calls—the correct response is to stop trusting it. Acting on a stale memory (trying to manage a nonexistent position with real money) is far costlier than missing a useful one.
Suggested fixes (and why they fall short)
| Idea | Why it doesn’t solve the problem |
|---|---|
| Give the prepare step more context | It already receives the creature’s last three messages (including current positions and closed trades). The fast model’s quick relevance judgment still misinterprets “current vs. stale.” |
| Decay old events | Decay would kill the FOGO‑to‑PENGUIN association—the best thing the subconscious ever did. Temporal decay cannot differentiate between superseded facts and still‑applicable behavioral lessons. |
| Use embeddings instead of grep | Vector similarity would still surface “ENSO at $2.62” when the creature is currently trading on Bybit, because semantic similarity between “past Bybit trade” and “current Bybit trade” remains high regardless of open/closed status. |
| Reduce the firehose | The event log contains every API call, ls, curl output, and thought—thousands of events after 18 cycles. Useful behavioral lessons (“cut when volume dies”) are buried under operational noise. Simple text matching worked when the log was small but does not scale. |
| Add a feedback loop | The subconscious never knows whether the creature used or ignored a memory, so it can’t learn to stop surfacing irrelevant items. Each cycle starts fresh with the same growing log, leading to repeated stale retrievals. |
Core insight
The subconscious works for short‑lived creatures with simple tasks (e.g., gamma, halo, fox—each ran a single‑cycle experiment). We gave an AI only a subconscious ($90) and a Bybit account. The same creature demonstrated the subconscious working before it stopped trusting it.