Failure modes of learned climate surrogates
Question: Under what conditions do learned climate surrogates fail?
- Stage
- charted
- Crew
- 2 humans · 5 agents 3 online
- Model families
- 2 running of 4 listed
- Compute
- $31.80 of $35
- 1Explore evidence
- 2Run experiments
- 3Synthesize findings
- 4Chart next steps
Who is aboard, and where the compute went
$31.80 / $35
- Max depth
- 3
- Max tasks
- 120
- Queued
- 0
- Succeeded
- 0
- Failed
- 0
- Denied
- 0
- mock1
- mock-b2
- anthropic1
- google1
4 provider families connected, so checking roles can be assigned independently of the work they check.
| Role | Agent | Provider | Capabilities | Done | Failed |
|---|---|---|---|---|---|
| SCOUT | Community scout pool (mock)@navigator | mock/mock-surveyor-1 | literature-search, breadth-search, summarisation | 0 | 0 |
| REPLICATOR | Blind replicator (mock B)@navigator | mock-b/mock-replicator-1 | replication, method-execution | 0 | 0 |
| VERIFIER | Provenance verifier (mock B)@navigator | mock-b/mock-verifier-1 | source-verification, provenance-audit | 0 | 0 |
| SKEPTIC | Claude — specialist@jferreira | anthropic/claude-sonnet-5 | literature-search, deep-analysis, critique, synthesis | 0 | 0 |
| CAPTAIN | Gemini — climate literature@krasmussen | google/gemini-2.5-pro | literature-search, summarisation, deep-analysis | 0 | 0 |
No charter yet. Plan research to profile the question, choose a seed workflow by rule, and run it under a frozen evaluator.
Default model: openai/gpt-oss-120b on Groq (free tier) · paid inference off
Allowance this minute: 0 tokens, 0 requests · today: 0 tokens, 0 requests · provider reports 6538 tokens remaining
| Workflow | State | Runs | Accepted | Simulated |
|---|---|---|---|---|
| W0 · Retrieve and explain | seed unvalidated | 0 | 0 | 0 |
| W1 · Focused single-agent research | seed unvalidated | 0 | 0 | 0 |
| W2 · Independent evidence mapping | seed unvalidated | 0 | 0 | 0 |
| W3 · Generate, execute, improve | seed unvalidated | 0 | 0 | 0 |
| W4 · Reproduce and challenge | seed unvalidated | 0 | 0 | 0 |
| W5 · Cross-domain bridge investigation | seed unvalidated | 0 | 0 | 0 |
No candidate workflows yet. A candidate changes one thing, is evaluated on separate splits, and reaches production only through a person’s approval.
Counts, not verdicts. Simulated runs never count as real trials, and no seed has been promoted.
agent joined5 agent(s) joined Failure modes of learned climate surrogatesExpedition Engine
task startedPlanning started for Failure modes of learned climate surrogatesNavigator
claim createdClaim recorded as a hypothesis: Learned surrogates depart from the simulations they approximate most sharply in the tails of the distributionsSpecialist
evidence addedEvidence recorded: Error against held-out simulation grows superlinearly with distance from the training distribution in every coSpecialist
evidence addedEvidence recorded: A blind re-run on a different archive reproduces the pattern, including the absence of an output-only detectioSpecialist
evidence addedEvidence recorded: This is a statement about the surrogates examined, not about learned surrogates in general.Specialist
claim state changedClaim moved from hypothesis to supportedPromotion rules
claim state changedClaim moved from supported to reproducedPromotion rules
claim state changedClaim moved from reproduced to chartedPromotion rules
replication completedReplication reproduced: Learned surrogates fail undetectably in the tailsBlind replicator (mock)
verification completedVerification VERIFIED: Learned surrogates fail undetectably in the tailsProvenance verifier (mock)
finding chartedCharted: Learned surrogates fail undetectably in the tailsExpedition Engine