Give your agent evidence that can still be checked

Discover the exact release boundary

The AI catalog is the machine entry point for the release-bound MCP 1.9.3 service, this site’s independently versioned OpenAPI 2.0.2 document, the China observation ledger and manifest, the BRI v2 contract and its WDI national-context bundle, and the canonical Financial Evidence Agent Skill. Use the catalog’s live-version authority instead of inferring deployment from a repository version string.

Connect in under a minute

Use the remote streamable-HTTP endpoint. Clients discover all seven hosted tools from the server, so no schema needs to be copied by hand.

claude mcp add --transport http palimpsest https://api.seiche.info/palimpsest/mcp

{
  "mcpServers": {
    "palimpsest": {
      "type": "http",
      "url": "https://api.seiche.info/palimpsest/mcp"
    }
  }
}

Seven hosted tools

Try the live server

The server response will appear here.

Use the JSON API directly

For deterministic pipelines and notebooks, import the OpenAPI 3.1 / Palimpsest REST 2.0.2 document or fetch a reading as ordinary JSON. The evidence-lake metrics projection reports the four reviewed v1 lanes and their source gates without making the underlying records public; validate it against the closed schema and recompute metrics_sha256 before quoting a count. That digest proves byte consistency, not source identity; recurring import additionally requires the closed shared-secret HMAC receipt. Start with the China Observatory index and its JSON Schema. The economic observation manifest and append-only JSONL ledger are published there too. The manifest checksum detects a byte mismatch; it does not authenticate publisher identity. Named-series forecasting publishes a separate backtest and forecast-status artifact under its forecast schema; its warming_up state is an abstention, not a forecast claim. The current BRI v2 reading links the reviewed WDI national-context bundle; its 3,564 China, Myanmar and Pakistan rows may be joined only as country-period context. The immutable WDI Pages publication receipt binds Release A commit 14b06772dfed6cdc736279c9ab61b444e5846598, workflow run 32984946320 and cache-busted served bytes verified at 2026-08-26T15:55:34Z. Its fresh_until clock of 2026-08-27T15:55:34Z is a point-in-time availability window, not continuous monitoring or a WDI observation clock. The frozen v1 reading remains available for reproducible citations. The files are cacheable, versionable and easy to archive.

curl -sS https://palimpsest.info/readings/board-alarm-latest.json \
  | python3 -m json.tool

For AI-evaluation due diligence, start with the claim ceiling rather than the headline. It separates integrity, recomputability, statistics, human validation and unaffiliated replication.

curl -sS https://palimpsest.info/readings/eval-assurance-latest.json \
  | python3 -m json.tool

To inspect the primary GFI v2 evidence, fetch the complete transcript matrix. It contains every response and null transport abstention across 3 models, 44 preregistered prompt arms and 660 samples, with explicit denominators and the offline verification command.

curl -sS https://palimpsest.info/readings/gfi-transcripts-latest.json \
  | python3 -m json.tool

For evidence-bound explanations of the origin, method changes and known failures, fetch the Eval Journal edition. Each article carries its falsifier and exact artifact receipts; the linked readings remain the measurement authority.

curl -sS https://palimpsest.info/readings/eval-journal-latest.json \
  | python3 -m json.tool

For the current deterministic findings edition, fetch the live article collection. It carries controls, intervals, counterreadings and sentence-level evidence selectors from the newest verified refusal-drift panel.

curl -sS https://palimpsest.info/readings/eval-articles-latest.json \
  | python3 -m json.tool

ScamShield reviewed bridge

The intelligence pack is a versioned static JSON contract, not an MCP tool. ScamShield loads its inert sources and typologies; a separate local stdin bridge can build a private Evidence Capsule from a structured assessment.

curl -sS https://palimpsest.info/integrations/scamshield/intelligence-pack-v1.json \
  | python3 -m json.tool

OpenAI Responses API

Pass the hosted endpoint as an MCP tool. Because all seven hosted tools are read-only, the example runs without per-call approval.

curl https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6",
    "tools": [{
      "type": "mcp",
      "server_label": "palimpsest",
      "server_url": "https://api.seiche.info/palimpsest/mcp",
      "require_approval": "never"
    }],
    "input": "What censorship signals are moving now? Cite timestamps and limitations."
  }'

ChatGPT app setup

1. Enable Developer mode in an eligible ChatGPT workspace.
2. Open Settings → Apps → Create and name the app “Palimpsest”.
3. Paste https://api.seiche.info/palimpsest/mcp as the MCP endpoint. No OAuth details are required.
4. Review the seven discovered read-only tools, create the app, then enable it in a conversation.

Use it honestly

For a named bank or other financial institution, start with the LiquiLens institution stress-review workflow. It connects institution evidence and Seiche funding context to OpenClaw, Hermes and scripted automation. Add Palimpsest when the question needs its China, regional, information-control or AI-evaluation records. Keep each source's dates, coverage and rights attached; these records do not certify an institution's regulatory compliance.