Historical evals cheat when tools retrieve corrected present-day evidence. PRIORI.SH is the versioned evidence and environment layer: pin a past instant, retrieve only evidence knowable then, and replay the same state through Live, Replay, or a sealed Pack.
Macro releases get revised; as-reported EPS gets
restated. Pin every feature to its
as_of
vintage and a research agent or systematic macro
backtest reads the release available on that date,
not the value quietly revised in months later.
client.search( query="payrolls", as_of="2024-03-08", # data available that day sources=["alfred"], )
RAG evals fail quietly when a June 2022 prompt can
retrieve 2024 filings. Point your MCP client at
PRIORI.SH with an
as_of date and the agent only receives
context that existed then — clean retrieval, cited
facts, repeatable runs.
# MCP server claude --mcp-server priorish \ "As of 2023-05-31, what did the Fed signal about persistent inflation?"
PRIORI.SH does not grade agents or store traces. It supplies the evidence state those systems evaluate against. Building that state internally means continuously capturing source versions, normalizing them, sealing provenance, maintaining replay, and tracking every upstream source change.
The same prompt, the same model
(Gemini 3.1 Pro),
answered four ways. A raw LLM leaks whatever its training
baked in. Give it web search and a strict point-in-time
prompt and it can reconstruct famous numbers — but on anything
that isn't front-page news it comes up empty, and it
never hands you a vintage.
PRIORI.SH grounds the answer in
what was actually knowable — today, or on any past
as_of — with the release date attached. Revisions
move the truth; watch the answer move with it.
api.priori.sh/mcp and give it the
prompt. Same tools, your model.
Sign up free for a key.
Gemini is a trademark of Google LLC. PRIORI.SH is not affiliated with or endorsed by Google; outputs are generated via the Gemini API and shown for illustration. See Terms.
Every response is filtered to what was knowable before it is returned — and every fact proves where it came from.
Each result is filtered to what was published at your
as_of, then stamped with a provenance chain: a
BLAKE3 content hash, the source filing, and a sealed
extraction lock. Open any fact and trace it back to the
original document.
curl -G "https://api.priori.sh/observations/series" \ -H "X-API-Key: $KEY" \ --data-urlencode "metric=PAYEMS" \ --data-urlencode "source=alfred" \ --data-urlencode "as_of=2025-07-15T00:00:00Z"
from priorish_sdk import PriorishClient client = PriorishClient(api_key="...") client.observation_series_by_key( "PAYEMS", source="alfred", as_of="2025-07-15T00:00:00Z", )
# claude_desktop_config.json { "mcpServers": { "priorish": { "url": "https://api.priori.sh/mcp", "headers": { "X-API-Key": "your-free-key" } } } }