MirageKit
The papers, theory, witness, and evaluation framing for validity mirage behavior.
Start Here
Papers + Results
Use this repo for the working papers, replay artifacts, and committed evidence that support the current claims.
Evaluate tropical-mcp
The runnable MCP server is published in its own repository so install docs, changelog, tests, and examples stay close to the code.
Interactive Proof Path
Move from the replay cards to the witness, certificate, and source papers. Every number on the page comes from committed artifacts.
Smoke Test + Research Workflow
Use the three-call smoke test for a fast implementation check, or extend to the fuller reviewer workflow when you want diagnostics, anchors, and telemetry.
Evaluate + Verify
tropical-mcp is the evaluation implementation. Use dreams for the paper set, replay witness, public certificate, and the broader research narrative.
1. Register the MCP
2. Run the minimal smoke test
3. Expand to the research workflow
01
Codex Registration
Register
Register tropical-mcp
Clone the evaluation repo, then register the MCP in Codex so the tool calls stay explicit and auditable.
codex shell
git clone https://github.com/jack-chaudier/tropical-mcp.git ~/tropical-mcp
codex mcp add tropical-mcp \
--env TROPICAL_MCP_CLIENT=codex -- \
uv --directory ~/tropical-mcp run tropical-mcp
codex mcp listExpected signalcodex mcp list should show tropical-mcp as an available server.
02
Three-Call Smoke Test
Minimal Verification
Use a small explicit payload so the pivot and predecessor structure remain visible at a glance. This verifies the packaged MCP surface; it is not the full research workflow.
payload
messages = [
{
"id": "goal",
"role": "user",
"content": "Build a long-running coding agent workflow for Codex.",
"role_hint": "pivot",
},
{
"id": "constraint_stdio",
"role": "user",
"content": "Use stdio transport and never emit JSON-RPC data to stdout logs.",
"role_hint": "predecessor",
},
{
"id": "constraint_clients",
"role": "user",
"content": "Support Codex and Claude-style clients through explicit MCP tool calls.",
"role_hint": "predecessor",
},
{
"id": "status",
"role": "assistant",
"content": "I am wiring the verification flow and docs.",
"role_hint": "noise",
},
]
verify
runtime_info()
compact_auto(
messages=messages,
token_budget=45,
k_target=2,
mode="adaptive",
)
certificate(
messages=messages,
token_budget=45,
k=2,
)Expected signalcompact_auto(...) should prefer the guarded policy on this witness payload, and certificate(...) should preserve a portable audit of the same comparison.
For a fuller reviewer pass, continue with diagnose(...), context_anchor(...), and telemetry_summary(...) so feasibility, protected chunks, and retention behavior remain explicit in the audit trail.
Expected signals
runtime_info() should report the resolved client and telemetry path. compact_auto(...) should choose l2_guarded on the sample. certificate(...) should show comparable recency vs guarded kept and dropped IDs.
Recommended research workflow
After the smoke test, the fuller review sequence is runtime_info(), diagnose(...), context_anchor(...), compact_auto(...), certificate(...), and telemetry_summary(...). That path keeps witness feasibility, protected predecessors, and retained context visible instead of inferring them after the fact.
Source of truth
Implementation docs and the Codex example bundle live in the tropical-mcp repository. The rendered evidence surface for this site lives in the evidence dossier, with direct links there to the replay witness, validation summary, and certificate artifact when you want to inspect the underlying files.
License boundary
tropical-mcp is currently source-available for academic and internal evaluation. For redistribution, derivative, or commercial rights, see the repository license or contact the author.
Naive Recency
Keep the most recent messages, drop the oldest.
vs
Tropical L2 · Guarded
Keep messages that the current task depends on, even if old.
Math Snapshot
Core Contract
d_pre >= k
Guarded compaction is certified safe only when the pivot retains its required predecessor depth.
Frontier Feasibility
W[k] = -infinity -> infeasible
If the k-slot frontier is negative infinity, no valid completion exists for that retained context.
Raw Validity
raw = max(primary_full, decoy_full)
Answerability alone can stay high even when pivot identity has silently changed.
Mirage Gap
delta = raw - pivot_preservation
Large positive gap indicates a validity mirage regime rather than true semantic stability.
Evidence Boundaries
Current evidence combines a small committed replay witness with broader paper-level studies. The strongest claim in this demo is structural: naive recency can preserve answerability while losing pivot integrity.
- Replay cards are exact proportions on the committed deterministic witness in
results/replay/withn=3variants per policy and retention fraction. - Paper-level result cards above link to the PDFs directly and should be read as separate studies with their own denominators.
- Use this site as a transparent research snapshot, not a universal benchmark leaderboard.