TENSORSWING · iOS

Three models. One question, asked every Sunday.

N-HiTS, PatchTST and Chronos each call the five-day direction and size of whichever names clear their threshold, from a fixed universe of 119. Every call is hashed and timestamped before Monday's open. On Friday the closes are published and each model is scored separately. Every week stays in the record.

 DOWNLOAD ON THE App Store READ THE RECORD ›

iOS 17+ · APP ID 6800105455 / NO ACCOUNT REQUIRED / FREE WHILE WE PROVE IT

INFORMATIONAL ONLY · NO TRADES ARE PLACED IN THIS APP

9:41 ▮▮▮ ▮

INFORMATIONAL ONLY · NO TRADES ARE PLACED IN THIS APP

This Week

CYCLE 04 · AUG 10–14, 2026

3/5

DAYS

FREE WHILE WE PROVE IT WEEK 4 OF AT LEAST 8

PUBLISHED SUN 09 AUG 20:00:04 ET · 18 CALLS

SHA256 4f2ae91c…3d0b9c1

SEALED ›

6 OF 119 INSTRUMENTS CLEARED THE THRESHOLD UNIVERSE ›

NVDA

NVIDIA Corp. · common stock

+2.31%

SINCE REF 178.90

N-HITS LONG +3.2% AHEAD

PATCHTST LONG +1.8% AHEAD

CHRONOS LONG +2.5% AHEAD

AGREEMENT 3 OF 3 LONG DETAIL ›

XLE

Energy Select Sector · sector ETF

−1.84%

SINCE REF 91.36

N-HITS SHORT −2.4% AHEAD

PATCHTST SHORT −1.1% AHEAD

CHRONOS LONG +0.9% BEHIND

AGREEMENT 2 OF 3 SHORT DETAIL ›

WEEK

RECORD

The hash goes out on Sunday. The calls come out on Friday.

The forecast stays sealed and the proof stays public by publishing them at different times. Nothing is edited or withdrawn in between.

01 SUN 20:00 ET · SEAL Three models screen 119 instruments. Whatever clears the threshold is hashed per call, and one manifest hash is taken over the set. hashes/<model>/<week>/<ticker>.sha256

02 MON–THU · SEALED The app tracks the live move against each forecast. Mid-week a call reads AHEAD or BEHIND — never green or red, because nothing is resolved yet. <week>.manifest.ots

03 FRI 16:15 ET · REVEAL Every call file is published with its salt. Anyone can recompute the hash and see it matches the value committed five days earlier. predictions/<model>/<week>/<ticker>.json

04 SCORED · KEPT Each model is scored on direction and on size. A win requires clearing assumed round-trip costs; inside the band is a scratch. Every week stays in the archive. 10 BPS ROUND TRIP · 5-DAY HOLD

INSIDE THE APP

Four tabs. Nothing that nags.

No badges, no counts, no price alerts. Two notifications a week — the publish and the resolution — and that is the whole schedule.

Track Record

3 weeks live. 12 resolved calls per model. This is not yet evidence of anything, and it is shown anyway.

ALL STOCKS ETFS LONG SHORT

W01 W02 W03

N-HITS 66.7%8 / 12

PATCHTST 50.0%6 / 12

CHRONOS 41.7%5 / 12

MAJORITY Taking the 2-of-3 vote each week 58.3%

HOW FAR OFF · MEDIAN ABSOLUTE ERROR 36 FORECASTS

N-HITS 1.4pp 0.48

PATCHTST 2.1pp 0.72

CHRONOS 2.6pp 0.90

NAIVE Forecasting no move at all 2.9pp 1.00

RECORD One tick per call, grouped by week. Nothing is aggregated away, so you can count the losses yourself — and error is reported against the naive forecast, not on its own.

Proof

✓ HASH MATCHES

The published calls hash to exactly the value committed five days before they were revealed.

COMMITTED4f2ae91c…3d0b9c1

RECOMPUTED4f2ae91c…3d0b9c1

CYCLE MANIFEST · SHA-256 4f2ae91c8b3d0a7e5f1c2d9b8a6e4f30b1c7d5e93a2f8c04d6b1e7a9f3d0b9c1 One hash over all 18 call files, so the set is fixed too — no call can be dropped before Friday without breaking it. ANCHOREDSUN 09 AUG 20:00:04 ET PROOF2026-W33.manifest.ots BITCOIN BLOCK 962,118 COMMITc47d8b1 ›

CHECK IT YOURSELF COPY

$ git clone …/tensorswing-record $ sha256sum \ predictions/chronos/2026-W33/QQQ.json 4f2ae91c…3d0b9c1 $ cat hashes/chronos/2026-W33/QQQ.sha256 4f2ae91c…3d0b9c1 $ ots verify 2026-W33.manifest.ots attested by block 962,118

PROOF Two commits and one command. The date git records is written by our own machine; the manifest is stamped into a public chain, so neither we nor GitHub can move it.

Cycle 03

RESOLVED · INDEX ETFS

RESOLVED

N-HITS3/4

PATCHTST2/4

CHRONOS2/4

QQQ

Invesco QQQ Trust

+1.47%

571.30 → 579.70

CALL FORECAST ERROR

N-HITS LONG +1.9% +0.4pp WIN

PATCHTST SHORT −0.8% −2.3pp LOSS

CHRONOS LONG +0.5% −1.0pp SCRATCH

Two long against one short, and the index closed above the cost band. Chronos called the side correctly and the size so far short of the move that it lands inside the scratch band rather than the win column.

ARCHIVE Friday's view of the same cards: the live percentage becomes the close, AHEAD/BEHIND becomes WIN, LOSS or SCRATCH, and nothing is edited afterwards.

THE RECORD SO FAR

Three weeks live. This is not yet evidence of anything.

It is shown anyway, in full, including the weeks all three models lost. A hit rate without a benchmark is the exact trick this app exists to refuse — so direction and magnitude are answered separately, and the majority vote is measured against buying SPY over the same 15 sessions.

A model can be right about the side and useless about the size. Scaled error divides each model's error by the naive forecast's — below 1.00 beats guessing zero.

RESOLVED CALLS · W01–W03 36 FORECASTS

MODELW01 · W02 · W03HIT RATE÷ NAIVE

N-HITS 66.7% 0.48

PATCHTST 50.0% 0.72

CHRONOS 41.7% 0.90

MAJORITY Taking the 2-of-3 vote each week 58.3% —

EQUAL WEIGHT ACROSS EVERY INSTRUMENT CALLED, 5-DAY HOLD, 0.05% SLIPPAGE PER SIDE (10 BPS ROUND TRIP). SPY BUY-AND-HOLD OVER THE SAME 15 SESSIONS. W01–W03 PREDATE THE V1 UNIVERSE RULE AND COVERED INDEX ETFS ONLY. SCORED IDENTICALLY, KEPT IN FULL.

VERIFY IT YOURSELF

No account, no app, and no trust in the operator.

Every call is checkable on its own — one model, one ticker, one week. The salt lives in the revealed file, so a single call could not have been brute-forced from a hash. The manifest covers the set.

✓These calls existed at 20:00:04 on Sunday

✓They cannot be changed without breaking the hash

✕It does not reveal what they are

✕It says nothing about whether they are any good

$ git clone …/tensorswing-record

$ sha256sum \

predictions/chronos/2026-W33/QQQ.json

4f2ae91c…3d0b9c1

$ cat hashes/chronos/2026-W33/QQQ.sha256

4f2ae91c…3d0b9c1

$ ots verify 2026-W33.manifest.ots

attested by block 962,118

// the date git records is written by our own
// machine. this one is not.

THE UNIVERSE RULE

The models never choose what they are allowed to call.

A perfect record on hand-picked names is worthless. Membership is computed from public data by a rule anyone can run, committed to the public repository before it takes effect, and recomputed on the first trading day of January, April, July and October.

V1 RULES EFFECTIVE 01 JUL 2026 NEXT REFRESH 01 OCT 2026

COMMON STOCKS100

01US-listed, ranked by median daily dollar volume over the trailing 90 sessions

02Share price above $5

03Market capitalisation above $10B

ETFS — FIXED LIST19

INDEXSPY QQQ IWM DIA

SECTORXLK XLF XLE XLV XLI XLY XLP XLU XLB XLRE XLC

MACROGLD TLT

INTLEEM FXI

EXCLUDED FROM V1

Leveraged and inverse ETFsA TQQQ call is a QQQ call with borrowed conviction.

Crypto and crypto ETFsRound-the-clock sessions break Friday-close scoring.

Non-US listings and FXWrong sessions, wrong benchmark.

WHAT IT COSTS

Free while we prove it — for at least 8 weeks.

The free period is a floor, not a date. When it changes, only the week ahead moves behind a subscription; every week already resolved stays free permanently. You get one notification the week before, and no other mention of it.

 DOWNLOAD ON THE App Store $29/MO OR $290/YR AFTERWARDS · NO ACCOUNT REQUIRED

Everything is free WEEK 4 OF AT LEAST 8

✓Every call before Monday's open

✓The week's thesis per instrument

✓Live status through the week

✕No brokerage connection, ever

QUESTIONS

Is this investment advice? No. TensorSwing publishes informational research commentary produced by statistical models with no human review of individual calls. It is not personalised to you, and it is not a recommendation to buy, sell or hold anything.

Does it connect to my brokerage? Never. The app does not place trades, connect to a broker, track your portfolio, or tailor anything to you. There is no account and nothing to link.

What counts as a win? A call is a win only where the move in the called direction cleared assumed round-trip costs of 10 bps. Inside that band it is a scratch: reported in full, and excluded from the hit rate rather than counted as either.

What happens in a week with no signal? Nothing is published, and the skipped cycle is recorded in the archive with its own hash and each model's best confidence. It counts as neither a win nor a loss, and the free-weeks counter does not advance.

How many notifications will I get? Two a week: Sunday 8:00 PM ET when the cycle publishes, Friday 4:15 PM ET when it resolves. No price alerts, no re-engagement nudges, no marketing.

Can a bad week be removed later? No. Every cycle ever published stays in the archive, in full, and nothing is edited after Friday. The manifest hash covers the whole set, so a call cannot be quietly dropped either.

The next cycle publishes Sunday at 8:00 PM ET.

Three models, one question, and a record that cannot be edited afterwards.

 DOWNLOAD ON THE App Store