Every judgment this engine makes about a prediction market is hashed, signed and stored before the market settles. Then reality grades it. This page is that score, and it updates itself.
These are signed and timestamped before the outcome exists. Nobody knows how they resolve, including us. Note a hash, come back after the close date, and check what happened. Every one carries its signature and our key is public, so you do not have to take any of it on our word.
33 older records carry no market snapshot and are excluded from this list.
There is no single accuracy number here, and that's deliberate. A blended figure would hide the only thing worth knowing: performance changes completely depending on whether the engine has ground to stand on.
On the Consumer Price Index it went 18 for 18. On gas prices it went 10 for 23 — worse than a coin flip.
On initial jobless claims it has had 8 opportunities to make a call and has taken none of them.
Where the engine performs badly, the cause is usually visible in the record itself: a large share of archived evaluations carry an empty market snapshot — no price, no volume. The engine was reasoning while blind, and it performed the way something reasoning while blind should perform.
Those numbers stay on this page at the same size as the good ones. A ledger that publishes only its wins is a brochure.
The abstention rate is the measurement that matters.
Across 14 settled events the engine declined to answer 62 of 116 times — and declined before any outcome existed, with the refusal itself signed and stored. An engine that always produces a confident answer can't be wrong in an interesting way. One that says I don't know, on the record, in advance, is the only kind whose confidence means anything.
Each row is one real-world question. The counts are the individual price strikes evaluated on it. Events are the honest denominator — one judgment stamped across a dozen strikes is one call, not a dozen wins.
| Event | Correct | Wrong | Declined | Outcome |
|---|---|---|---|---|
| KXFED-26SEP | 4 | 0 | 7 | clean |
| KXTSAW-26SEP13 | 1 | 0 | 3 | clean |
| KXAAAGASW-26SEP14 | 1 | 4 | 2 | mixed |
| KXCPI-26AUG | 12 | 0 | 3 | clean |
| KXJOBLESSCLAIMS-26SEP10 | 0 | 0 | 3 | silent |
| KXTSAW-26SEP06 | 0 | 1 | 1 | wrong |
| KXAAAGASW-26SEP07 | 0 | 1 | 2 | wrong |
| KXJOBLESSCLAIMS-26SEP03 | 0 | 0 | 4 | silent |
| KXAAAGASW-26AUG31 | 1 | 2 | 2 | mixed |
| KXTSAW-26AUG30 | 0 | 0 | 1 | silent |
| KXJOBLESSCLAIMS-26AUG27 | 0 | 0 | 1 | silent |
| KXTSAW-26AUG23 | 6 | 1 | 9 | mixed |
| KXAAAGASW-26AUG24 | 8 | 6 | 20 | mixed |
| KXCPI-26JUL | 6 | 0 | 4 | clean |
Some records were made blind. Where the market snapshot shows no price and no volume, no price signal reached the engine. Those entries are flagged no price data above. They can't support any claim about disagreeing with the market, and they are the documented cause of the worse-than-chance result on AAA national gas price shown above — the engine was asked to reason about gas prices while receiving no gas price.
Strike counts are not independent samples. One event is evaluated across many price strikes. The inflation in this archive runs roughly eight to one, which is why every headline figure is stated per event.
This is a small sample. Nothing here is a stable accuracy rate. It is a record of what happened, published so it can be checked, and it will read differently as more markets settle.
You can check a stamp yourself. Open check this signature under any call above for its record id, SHA3-256 hash and Ed25519 signature, and take our public key from /.well-known/cosmic-keys.json. Verify the signature over the UTF-8 bytes of the lowercase hex hash — not over the raw digest bytes, which is the one detail that trips people up. Records stamped before the feed carried signatures say so rather than showing a blank. What is still not independently checkable is the aggregate: the counts above are computed from these records, and you would have to recompute them yourself to confirm we added them up honestly.
COSMIC AEI reads emotional state from language for people in high-stress work — fire, EMS, law enforcement, dispatch. No one can verify that reading. A group readiness score has no answer key — there's no moment where reality arrives and says you were wrong on Tuesday.
Prediction markets have exactly that. They resolve, on a date, with no way to move the goalposts. So the same engine runs against questions that settle, signs every judgment before the outcome exists, and publishes what reality says back.
This isn't a demonstration that the engine is right. It's a demonstration that it can be checked — and that when it has nothing to go on, it says so instead of guessing.
This is not financial advice and it is not a trading tool.
Nothing here recommends buying, selling or holding anything. Rage Relief LLC does not place trades, offer brokerage, or manage money. The markets are a grading instrument and nothing else. Please don't follow these calls — a meaningful share of them are, by our own count, wrong.