IMPACTS / SINGULARIS
SINGULARIS™ · The Decision Engine

Every opportunity, reduced to one rational decision.

Beneath the interface is a disciplined act of inference: read a market‑member state, estimate the expected value of acting on it, and commit capital only when the math clears — impression by impression, in 10–12 minute cycles.

The architecture → How we handle signals →
[ The unit of decision ]

It begins with a state, not a segment.

For every market member, household, or context, SINGULARIS maintains a live probabilistic state vector — Si,t — re‑estimated continuously as new signal arrives. It is not a profile. It is a calibrated, consent‑based estimate of who this is, what they intend, what they are worth, and whether they can be reached and measured — right now.

{{ grp.g }}
{{ it }}
[ Confidence regimes ]

Confidence is not hidden. It is priced.

Every estimate carries an identity‑confidence regime. Lower confidence does not mean lower value — it means a disciplined, lower bid and a heavier burden of proof. Nothing ineligible is ever actioned.

{{ r.v }}
{{ r.k }}
{{ r.d }}
[ The governing equation ]

The expected value of a single impression.

EV(impression) =
P(valid)REAL · NON‑FRAUD
×
P(viewable)SEEN OR COMPLETED
×
P(attention)| CREATIVE · CONTEXT · STATE
×
P(conversion)| ATTENTION · PATH · OFFER
×
ENRCEXPECTED NET REVENUE
×
P(lift)INCREMENTAL · NOT HARVESTED
media_costTHE CLEARING PRICE
fatigue_costFUTURE RECEPTIVITY SPENT
risk_penaltyBRAND · SUITABILITY · COMPLIANCE

Plain reading: pay only for impressions likely to be seen, to earn real attention, to convert, to add net value, and to be incremental — minus what they cost, what they spend in future goodwill, and what they risk.

max_CPM  ≈  1000 × EVper‑impression × confidence_adjustment
The most we will ever rationally pay to reach this state, now.
Inside ENRC — and the term most systems omit
ENRC = Σt δt·E[ margin ]   ( acquisition + service + discount + cannibalization )

Expected net revenue is a discounted stream, not a single sale — and it subtracts the cost most platforms quietly ignore: cannibalization, the value that would have happened without paid media. If a customer would have bought anyway, the ad did not create value — it harvested demand. Refusing to pay for harvested demand is what separates capital allocation from spending, and it is why the engine needs ROAS360 incrementality downstream to keep itself honest.

[ MatchRate™ ]

A single, calibrated score for every opportunity.

MatchRate™ collapses the decision math into one number from 0.00 to 99.99 — a calibrated percentile, not a raw probability. A gradient‑boosted core with Bayesian hierarchical priors keeps new or sparse states honest; the score moves the instant reality does.

It is fluid by design. A score decays in the hours after a conversion, so spend is not wasted on the already‑won; it climbs on an abandoned cart or a return visit. Members and contexts flow between value tranches continuously.

Tranche transitions — a state machine, not a list
T3 → T1checks availability · rain raises weekend urgency
T2 → T1completes 75% of video · later branded search
T1 → T4purchases — immediate acquisition value collapses
T1 → ⊘fatigue · opt‑out · recent conversion → suppressed
Scoring function
MatchRate = percentile(
  w₁·P(conv) + w₂·ENRC
  + w₃·P(attention) + w₄·affinity
  + w₅·timing
  − w₆·fatigue
  − w₇·post‑conversion_decay )
{{ t.t }} {{ t.k }} {{ t.d }}
[ From value to bid ]

We do not buy media. We allocate acquisition capital.

The expected value sets the ceiling; a constrained optimization sets the bid. Every dollar competes against every other dollar, across platforms, in real time — pacing, frequency, and confidence all binding the result downward, never up.

Bid ceiling
bid_ceiling = min(
{{ b.t }} {{ b.d }}
)
[ One loop ]

Decide. Verify. Prove.
Then decide better.

Every decision the engine makes is checked against what actually happened — attention measured by ATTENVERA, return proven by ROAS360 — and folded back into the next estimate. The model that priced this impression is sharper for the next.

See it in the field → The case studies → Privacy & governance →