The platform
Evaluate and run agents against what was knowable
Fifteen products over one point-in-time record of the US public filings. They run in pipeline order: ingest the record with the clock each row arrived on, test an agent against a frozen window, run it in production on the same clocks, and prove afterwards what it was shown.
Every product below reads the same rows under the same clock rules, so a coverage certificate means the same thing in a flat file, in a benchmark release and on a live query.
The harness is Apache-2.0 and needs no account. Everything else is on the plans, from $29 a month.
Ingest · data
Every record kept the way it arrived
One certified record of the US public filings, with a receipt for when each row became public. 587,748 rows across seven sources on 2026-08-27, and a tape writing forward since 2026-08-26.
Point-in-Time API
The as-of query surface. Give it a company and a timestamp and it returns the filings that were public at that minute, names the timestamp field it filtered on, and certifies every day the answer rests on. The measured stakes: 10.3% of filings are accepted after the day boundary a naive date join uses, a median of 1.1 hours late and 343 days at worst.
LIVE Researcher $29 / Power $79 / Desk $499 monthly.
Live Feed
Real-time capture, recording every fifteen minutes since 2026-08-26. Each record notes what arrived and when we saw it, so a question about 14:15 is answered from a row written at 14:15. Streams over server-sent events.
LIVE included.
Flat Files
The whole corpus as parquet: one file per source per day and a manifest listing the sha256 of every file. Load it into DuckDB or pandas and work offline; the manifest states exactly which corpus version you hold.
LIVE Power and up.
Data Connectors
Bring your own market-data keys and PIT pins each vendor's bars to the same clock the filings use, so your agent reads one timeline. The cache is closed by license rather than by configuration, and every answer reports the class its bytes were handled under.
BUILDING
Test · evaluation
Contamination as a measured quantity, next to the score
On real prices, one simple strategy shows +5.13% when its joins leak tomorrow's filings and −2.02% when they do not. 47 of its 64 trades were placed at opening auctions that happened before EDGAR had accepted the filing.
Evaluation Harness
The open-source runner: your agent, a benchmark window, network off. Every tool call is intercepted and served only from the window, and the full trajectory is hashed into an execution receipt that anyone holding the files can recompute. Runs on your machine, needs no account.
BUILDING open source.
Benchmark Datasets
Frozen eval windows with matched control arms: the rows as served, the same rows with issuers redacted, and the same rows with every date moved ten years forward. The score that survives the shifted arm came from the rows; the score that vanishes came from the model's memory.
LIVE text-complete edition days away.
Post-Cutoff Holdouts
Evaluation on time newer than any model's training data. The live feed ages into this product daily: records captured since 2026-08-26 did not exist when current models trained, so memorisation cannot produce the score.
LIVE included with Benchmark Datasets.
Eval Gates
CI for agents. Every commit re-runs your agent against a frozen benchmark, and the build fails if leakage grew or a control-arm baseline broke, so leakage shows up as a failed build the day it is introduced.
BUILDING
Run · runtime
The same clock discipline at development speed and in production
MCP Server
The tool server agents connect to — Claude Code, Cursor, any
MCP client — where every question carries
known_at and every answer states which clock it used.
LIVE all plans.
Deterministic Replay
Re-run a recorded agent session with the world reconstructed byte for byte, so a changed outcome means your agent changed rather than the world. Dry-replay serves the model's own recorded answers back, which checks the loop itself at zero cost.
BUILDING
Decision Audit Logs
Every decision a production agent makes, recorded with the exact corpus state it saw, hashed. "What did the agent know" has an answer months later — for you, your customers, or your auditors.
LIVE Desk.
Prove · audit
What the agent scored, and how it reasoned to get there
Reasoning Audits
Every claim in the agent's rationale, checked against what the run actually served. A trace citing a dollar figure, a date or an outcome that appears in no served row is flagged, quoted and scored, so an agent that calls a collapse by citing the size of a capital raise it was never shown is scored as remembering rather than forecasting.
BUILDING
Contamination Certificates
A signed verdict on one run: what the agent was shown, when each item was knowable, the control-arm deltas, and the reasoning-audit findings. Anyone holding the certificate can check it against the corpus without trusting us, or you.
BUILDING
Certified Leaderboard
Certified runs, listed publicly if you choose. Rankings appear only once a published quorum of comparable certified runs exists; until a cell meets it, there is no leaderboard for that cell.
PLANNED
Academy
Training for teams building time-aware agents: the clock model, the control-arm method and the reasoning audit, taught on the real corpus.
PLANNED
Pricing
What each stage costs
Monthly and self-serve, checking out from the pricing page. Reasoning Audits meter per decision audited with an allowance in every plan, and Contamination Certificates price flat per run with the audit included. Annual contracts, custom windows and DPAs go through operations@aqx.llc. Full details on the pricing page.