Loom: Ask the Catalog, Then Check the Answer | CMD+RVL
Written by CMD+RVL.
Real Loom sessions against the public catalog, shown whole: the answer, the source it came from, the transformation behind it, and the date it was true.A model will answer a question about an interest rate with the same confidence whether it traced the number or invented it. That is fine until the number enters a decision.
Loom is chat over our public catalog — finance and procurement datasets, from Fed rates to wholesale vehicle prices to filings. The part worth showing is not that it answers. It is what happens when you ask it to prove the answer.
Below are real sessions, run on 29 July 2026 and reproduced whole. Nothing has been rewritten to read better. Where something was cut or corrected, the transcript says so.
Ask for evidence and it hands you the chain
The first question is the one that matters. Not what is the rate — anything can return a number — but what evidence do you have.
Session, 29 July 2026
· In testing
Asked
What evidence do you have that overnight SOFR is 3.64%?
Answered
Here's the complete provenance chain showing how that rate value was produced:
Data Lineage
- Source: AWS S3 bucket s3://edgar-data-full/curves/sofr (raw JSON files from NY Fed)
- External Table: edgar_db.source.curves_sofr_daily_data_ext (Snowflake external table pointing to S3)
- Staging Table: edgar_db.dbt_staging_curves.stg_curves_sofr_daily_data (transformed & validated)
Transformation Logic
The rate is extracted from JSON using this code:
try_to_decimal(rate_obj:percentRate::string, 18, 6) as percent_rateThe transformation:
- Parses JSON field percentRate from the NY Fed API response
- Converts to a decimal with 6 decimal places
- Filters to only valid rates between -5% and 20%
- Deduplicates by effective_date, keeping the most recently modified file
Verified Data
Query results showing the actual values in the warehouse:
| Date | Rate | Volume (billions) |
|---|---|---|
| 2026-07-24 | 3.640000% | $2,979 |
| 2026-07-23 | 3.640000% | $2,971 |
| 2026-07-22 | 3.620000% | $3,026 |
| 2026-07-21 | 3.610000% | $2,975 |
| 2026-07-20 | 3.570000% | $3,012 |
Confidence level: PARSED_LINEAGE (highest tier) — the value comes directly from the NY Fed's official SOFR API JSON feed with no intermediate calculations, verified against live warehouse data.
Traced to
- markets.newyorkfed.org — SOFR API
- stg_curves_sofr_daily_data.sql — internal transformation, not publicly readable
Asked 2026-07-29
Three things are worth noticing. The answer names the original source, the New York Fed's own API, which anyone can open and check against. It shows the transformation that turned the Fed's JSON into the number returned, including the validation range and the deduplication rule. And it grades its own confidence, because a value read straight from a source feed is a different kind of claim from one computed through several steps.
The transformation itself lives in a private repository. We show what it does rather than linking somewhere a reader cannot follow — a dead link inside a provenance claim is worse than no link, because it invites the one test a reader can actually run.
The same question, backwards
The second session asks for a historical series rather than a single value.
Session, 29 July 2026
· In testing
Asked
What is the 3-month average SOFR at end of Q2, and where does that number come from?
Answered
The 3-month (90-day) average SOFR at the end of Q2 2025 (June 30, 2025) was 4.342100%
For historical context, here are the end-of-Q2 rates for recent years:
- 2026-06-30: 3.63459%
- 2025-06-30: 4.34210%
- 2024-06-28: 5.353310%
- 2023-06-30: 4.99569%
- 2022-06-30: 0.69870%
- 2021-06-30: 0.01578%
- 2020-06-30: 0.04767%
Original source: New York Federal Reserve (NYFED). The NY Fed publishes official SOFR (Secured Overnight Financing Rate) data including various averages. Data is fetched via HTTP API calls that return JSON responses.
[Internal pipeline steps between the Fed API and the queried table.]
[Amended: 2024-06-28 was absent from this capture: a pipeline bug skipped the only quarter-end in the range that fell on a weekend. Fixed, and the value confirmed on a rerun.]
Asked 2026-07-29
This one found a bug. The original run skipped 2024 — the only quarter-end in the range that fell on a weekend, so the correct date is the last Fed business day, 28 June, rather than the 30th. The pipeline handled the weekend case incorrectly and dropped the row. It has been fixed, the value confirmed on a rerun, and the transcript carries an amendment note saying exactly that.
We left the note in rather than quietly editing the record. A catalog that tells you when it was wrong is worth more than one that has never appeared to be.
Why this is the product
Every other piece in our portfolio shows one result pulled out of the catalog: a capital structure, a procurement signal, a scored contract book. This shows the catalog itself, answering, with its working attached.
That is the whole argument. An answer you cannot check is a guess wearing a suit — and the check has to be cheap, or nobody does it.