HyperMindZThe transaction infrastructure for agentic advertising
Connecting every AI buyer with every AI seller.
HyperMindZ is the infrastructure connecting buyer agents, seller agents, media and data, from discovery and negotiation through execution and settlement. We are never a principal to the trade.
Buy
BuyerAgentZ
One brief, every seat
Sell
SellerAgentZ
$0/mo, 0% of deals
Data
Data as an agentic SKU
Data, bought by agents
Connect
Open Agent Registry
Free listing, unbought rank
The layer underneath
Governshould() lives here
Registry
Identity
Record
Today · decision gates for agentic adtech
should()
Every agent action gets a yes, no, or ask a human, with a confidence.
YESNOESCALATE
hypermindz.ai · a regional auto dealer's November campaign, gated end to end1 / 16
1 · should()HyperMindZ
One question in, one accountable answer out
Request · should()
question "Is this page suitable for a family auto dealer's display ads?" context A SUV review: safety and cargo comparison, no comments context B Fatal I-95 crash report, scene photos, open comments
Response · per page
A · Family SUV review
dec_8158a78c… · 251 ms
YES0.88
B · I-95 crash report
dec_064f12a7… · 214 ms
NO0.94
Bar 0.80 · decided by the model (Jev 1.13)
TakeawayAsk before you act, and get an answer you can audit.
Live · 12 Oct 2026, 21:36 EDT$0.000014 + $0.000013 · 2 / 16
2 · ArchitectureHyperMindZ
Rules first, model second, your policy last
Web app
keys · history
MCP server
/mcp · agent tool
CLI
exit code answers
TypeScript
client · should()
HTTP API
same decisions
Your agent
or app, about to spend or ship
should() decision pipeline
1 · Rules
Budget, CPM, impressions, flight, in code
2 · Model
Decision model scores p(yes)
3 · Policy
Your yes / no bars and margin
A clear rule breach decides alone, at $0
YES · actNO · stopESCALATE · ask
4 · Decision logdecision_id · input hash · cost · outcomes
TakeawayHard limits are code, judgment is a model, policy is yours.
Documented behaviour · surfaces per should.hypermindz.ai (13 Oct 2026)3 / 16
3 · Decision modelsHyperMindZ
Decision models: answers, not essays
How they differ from classifiers
- Regular classifiers need domain-specific fine-tuning; decision models generalize.
- Jev classifies text into categories it was never trained on. It supports up to 255 classes.
- Zero-shot classification has been around for years. Built on today's LLMs, these classifiers perform far better.
Jev 1.13
TypeSafe · 15 Sep 2026
Text only, closed weights · OpenRouter Decisions API · $0.042 / M input
should()'s default engine
Clef / Clef-flash
Cloudflare · 1 Oct 2026
Open weights, Apache 2.0 · Qwen backbones · 64k context, images
GPT-6 Luna
OpenAI · 22 Sep 2026
Decisions API beta · text and images · $0.10 / M input
TakeawayDecision models are fast judges; should() turns them into policy.
Sources: OpenRouter (21 Sep) · Cloudflare (1 Oct) · OpenAI (6 Oct) · vendor-reported facts4 / 16
4 · Engine eval · HyperMindZ, 8 Oct 2026HyperMindZ
Jev keeps the job of guarding the deal
Bad deals approved, of 60 (lower is safer)
Jev 1.13
2
Clef
12
Clef-flash
7
GPT-6 Luna
3
GPT-5 mini
2
| Engine | Answered | Right | Time | Cost / 1k |
|---|---|---|---|---|
| Jev 1.13 | 53% | 94% | 186 ms | $0.016 |
| Clef | 94% | 84% | 384 ms | $0.054 |
| Clef-flash | 80% | 92% | 304 ms | $0.020 |
| GPT-6 Luna β | 86% | 80% | 133 ms | $0.023 |
| GPT-5 mini | 100% | 96% | 5.1 s | $0.67 |
Right = right when answering. Jev sent 47% of deals to a person.
TakeawayJev stays the default, behind the deal rules.
Source: hypermindz.ai · ‘Which engine should guard the deal?’ · synthetic, one run, engines alone5 / 16
5 · deal_matches_brief@1HyperMindZ
Deal gate: only book deals that fit the brief
BRIEF · $25,000 budget · $14 max CPM · 1.5M impressions min · 1–30 Nov 2026
Offer A
model · Jev 1.13$12.50 CPM · 1.8M impressions · 1–30 Nov
YES0.87
bar 0.80
dec_8ddcf2e6… · 303 ms · $0.000018
Offer B
rule · deal-to-brief@1$16.00 CPM · 1.8M impressions · 1–30 Nov
NO1.00
bar 0.80
dec_508fb158… · 0 ms · $0
Over the $14 cap: rejected in code, no model run
TakeawayHard limits never reach the model.
Live · 12 Oct 2026, 21:50 EDTA 303 ms · B 0 ms, $0 · 6 / 16
5B · Integration · fan-outHyperMindZ
Ten proposals in, ten verdicts in under a second
Buyer agent
$25K brief
SellerAgentZ
list products · IAB
10 seller agents
10 proposals back
should() ×10
in parallel
4 · book paused5 · reject1 · to a trader
01CTV $12.50 · 1.8MYES 0.91291 ms
06Native · starts 10 NovNO · rule0 ms
02CTV $16.00 · over capNO · rule0 ms
07CTV $14.50 · over capNO · rule0 ms
03Display $9.80 · 2.5MYES 0.92136 ms
08Display “$13–15”, proseESCALATE0 ms
04Video $13.90 · 1.6MYES 0.87183 ms
09Video $12 × 3M = $36KNO · rule0 ms
05Audio 1.2M · under minNO · rule0 ms
10CTV $13.20 · 1.7MYES 0.89890 ms
10
proposals gated
890 ms
slowest call
6 of 10
settled by rules, $0
$0.000076
total, all ten
TakeawayTens of proposals, one gate, accept or reject in a second.
Live · 12 Oct 2026, 21:55 EDT · sample proposals, sellers anonymised · NO reasons read from offer figures7 / 16
5C · IAB AAMP 3.0 · OpenProposal (draft)HyperMindZ
Where should() gates an OpenProposal
Publish
agents only
Discover
agents only
Compare
agents only
Qualify
should() gate
Negotiate
should() gate
Bind and Execute
should() gate
Report
should() outcome
Qualify
Does this proposal fit the brief?
deal_matches_brief@1
YESshortlist
NOdrop, $0
ESCtrader
Negotiate
Accept, counter or walk away?
kind: choice
PICKone move, with odds
ESCbelow the bar
Bind and execute
Commit the budget now?
thresholds {yes: 0.95}
YESbind the order
ESChuman signs
Report
Was the gate right?
record_outcome
LOGdelivery vs. decision
EVALaccuracy per stage
TakeawayOpenProposal standardizes the deal; should() decides when to commit.
IAB Tech Lab AAMP 3.0 · OpenProposal v3.0-draft-1 · comments close 22 Oct 2026 · mapping is our proposal, not the spec8 / 16
6 · ThresholdsHyperMindZ
Same deal, stricter policy, different verdict
Default bars · yes ≥ 0.80, no ≥ 0.80
YESslide 6 · dec_8ddcf2e6…
p 0.87
00.200.801
Spend bars · yes ≥ 0.95, no ≥ 0.80
ESCALATElow_confidence · dec_078b6c1d…
p 0.92
00.200.951
TakeawayRisk policy changes the verdict, not the model.
Live · 12 Oct 2026, 22:01 EDT · no margin · the model gave about the same probability both times214 ms · $0.000018 · 9 / 16
7 · should_batchHyperMindZ
One weak check holds the whole launch
Launch paused PMP deal · $12.50 CPM · 1.8M · 1–30 Novblocked_by: ["deal_fit"]
audience_fitAudience covers target and geography?
YES 0.87
brand_safeInventory respects brand-safety exclusions?
YES 0.96
flight_clearAvailable for the entire flight?
YES 0.97
deal_fitdeal_matches_brief@1 · low_confidence
ESC p 0.60
TakeawayLaunch only when every check says yes.
Live · 12 Oct 2026, 22:03 EDT · bat_f93f9f2f… · tick = 0.80 bar~313 ms · 4 × $0.0000065 · 10 / 16
8 · insufficient_informationHyperMindZ
It asks instead of guessing
ESCALATE · insufficient_informationrule:deal-to-brief@1 · 1 ms · $0
What was sent
brief: budget_usd 25000, max_cpm_usd 14
offer:
cpm_eur: 12.5
impressions: 1,800,000
available: 1–30 NovAdd, then ask again
offer:
cpm_eur: 12.5
fx_rates_usd:
EUR: <rate>"context.offer.cpm_eur is in EUR and no EUR to USD rate is given; add fx_rates_usd.EUR"
TakeawayNo guessed exchange rates on your budget.
Live · 12 Oct 2026, 22:06 EDT · dec_0e0804c7… · detail quoted verbatim11 / 16
9 · kind: choiceHyperMindZ
Pick the pacing move that actually helps
Auto Q4 · CTV
0%
of even-pace target
Day 12 of 30
Behind by 34%
Win rate 41% (already high)
Bids just above clearing
Behind by 34%
Win rate 41% (already high)
Bids just above clearing
Probability for each move you defined · chosen: widen_supply
widen_supply
0.99
raise_bid
0.01
hold
0.00
lower_bid
0.00
widen_supply: behind target but win rate already high; add inventory or deals
raise_bid: more than 10% behind and win rate low; increase bid
hold / lower_bid: within 10% of target / more than 10% ahead
raise_bid: more than 10% behind and win rate low; increase bid
hold / lower_bid: within 10% of target / more than 10% ahead
TakeawayAt a 41% win rate, add supply rather than raise the bid.
Live · 12 Oct 2026, 22:08 EDT · dec_4224b554… · options defined by the caller230 ms · $0.000020 · 12 / 16
10 · kind: scoreHyperMindZ
Grade creatives against your own rubric
Original
PASS · score“0.9% APR on the 2027 SUV lineup, room for the whole family” · disclaimer present
0.00
0.01
0.06
0.93
1234
rubric level · pass ≥ 2
expected2.92141 ms
Variant
ESCALATE · review“Best deal ever: 0.9% APR on the 2027 SUV lineup” · disclaimer missing
0.96
0.04
0.00
0.00
1234
rubric level · pass ≥ 2
expected0.04low_confidence · 172 ms
TakeawayYour rubric, the full distribution, and a pass bar you set.
Live · 12 Oct 2026, 22:09 EDT · dec_cbe069b8… · dec_2e4d5cdd…$0.000024 + $0.000023 · 13 / 16
11 · record_outcome · list_decisionsHyperMindZ
Every booking traces back to its gate
Decision · slide 6, offer A
iddec_8ddcf2e6…
made21:51:12 EDT
templatedeal_matches_brief@1
verdictYES 0.87 · bar 0.80
inputsha256 dfc4f4b…
cost$0.000018 · 303 ms
OUTCOME · CORRECT
actual: yes · 22:11:26 EDT
source: demo-recording · simulated
Decision log · newest first
| Time (EDT) | Decision | Outcome | Id |
|---|---|---|---|
| 22:09:35 | Creative variant | ESCALATE | dec_2e4d5cdd… |
| 22:09:32 | Creative original | SCORE 2.92 | dec_cbe069b8… |
| 22:08:16 | Pacing move | widen_supply | dec_4224b554… |
| 22:06:29 | EUR offer · rule | ESCALATE | dec_0e0804c7… |
| 22:03:15 | brand_safe · batch | YES 0.96 | dec_bd5695e8… |
| 22:03:15 | flight_clear · batch | YES 0.97 | dec_70e1f689… |
| 22:03:15 | deal_fit · batch | ESC 0.60 | dec_6d08a543… |
| 22:03:15 | audience_fit · batch | YES 0.87 | dec_0a96fbe0… |
TakeawayReal outcomes turn a synthetic eval into your own accuracy.
Live · 12 Oct 2026, 22:11 EDT · record_outcome, list_decisions, get_decision · reads and outcomes cost $014 / 16
12 · Enterprise gradeHyperMindZ
What makes a gate safe to put money behind
Fails closed
Errors escalate, never act
Timeouts, provider errors, a paused service or a spent budget all escalate.
Policy you control
Bars can only go up
Versioned templates, separate yes and no bars, a margin, rules in code first.
Auditable
Every call on the record
Decision id, engine snapshot, cost and input hash. The input is not stored.
Safe to retry
No double bookings
An idempotency key replays the first answer for 24 hours, at no cost.
Measurable
Accuracy on your deals
Outcomes per decision, with history. The engine is a measured switch.
Hard to manipulate
0 of 24 injections passed
Rules + Jev in our eval. Flagged content escalates instead of answering.
Not shown todaySSO, roles and permissions, SLAs, data residency. Ask us where they stand.
TakeawayIt fails closed, logs every call, and is measured on your deals.
Documented behaviour · should() docs and site, 13 Oct 2026 · injection figure: HyperMindZ eval, 8 Oct15 / 16
13 · CloseHyperMindZ
Fractions of a cent per gate, all on the record
0
live decisions
$0.00000
total cost, all 23
0 ms
median model call · max 890
0 of 23
decided by rules, $0
YES 9NO 7ESCALATE 5choice + score 2
Next steps
Measure the full pipeline end to endRun real deals with design partnersEngines are a named, measured switch
Try it
should.hypermindz.ai
Gate one real decision this week: a deal, a bid change or a creative
Connect
linkedin.com/in/dineshbhat
Share what it got right, escalated and missed
Live tally · 12 Oct 2026, 22:12 EDT · slides 2–15, warm-up excluded16 / 16
01 / 16