The furthest‑along arc: the demo landed, the deck is in their hands, and the proof of concept was granted out loud on the Jul 30 call. This file holds what the room is (Unified, inside iHeart), what their machine does by hand, and how the backtest converts a verbal yes into a track record.
Their pain in their own words, timestamped, priced by the research.
iHeart's broadcast money closes by hand: under‑delivery spawns makegoods and audience‑deficiency units, political preempts scheduled spots, and quarter‑end is, in Josh's words, "Tetris combined with the loudest person in the room." On the digital side, Matt's optimizers move pacing, bid, and budget, and every new business rule waits in an offshore code backlog ("never ending… right now we're working on Spotify, the digital out of home"). Nobody moves inventory levers, nobody prices the allocation decision, and the rule that would catch yesterday's mistake takes a backlog slot. Precise referees both: prices each allocation decision, calibrates predicted against realized, turns rules into plain sentences, and writes every decision down completely, signed. Recommendations, not writes. The goal in their words: deliver in full at lower cost.
| # | Fact | Why it sharpens the pitch |
|---|---|---|
| F1 | Broadcast shrinks while digital carries growth: FY2025 revenue $3.865B flat, Multiplatform −4.2%, Digital +14.2%, programmatic ~$200M and up ~50% in Q1 2026 | The money is moving onto exactly the decision surface we price |
| F2 | Cost pressure is structural: $100M in‑year cuts plus a new $50M program, FCF pointed at 10.875% second‑lien debt | "Lower the cost while delivering in full" is their balance sheet talking; it lands in Bressler's office unassisted |
| F3 | They are building our decision surface right now: broadcast became DSP‑buyable Dec 2025 (Viant, Yahoo, Amazon) and AudioGraph launched Jun 2026 with predictive models and outcome attribution | A fresh predicted‑vs‑realized gap, created by their own growth story, with no calibration discipline behind it yet |
| F4 | Quarter‑end is still manual: centralized traffic desks, spot times keyed by hand, and AdBuilder's own terms repay under‑delivery with bonus impressions rather than cash | Under‑delivery‑as‑inventory‑cost is contractual and company‑wide; pricing it is found money |
| F5 | AudioGraph ships a "75% higher KPI vs demo‑based plans" claim, and Backer is credited by name on it and on the Magellan attribution deal | Predicted‑vs‑realized holding up is personally his exposure; a referee protects his own claim |
Not a proposal: a conversion. Both principals said yes on the call; the job now is making the yes real before it cools.
What was agreed, verbatim: Josh at 30:14, "no reason not to plug some campaigns in… recommendations… without rights to any campaign changes. Seems like an easy test." Pilcer: "easy POC." The mechanics they proposed themselves: he and Josh brief Phil and AJ and ad ops, pick closed orders, extract log‑level data to an S3 bucket. The reverse demo was Pilcer's own suggestion. Adam's frame that landed: the hedge‑fund backtest.
The shape we run: a retrospective backtest on their chosen orders: contribution value per publisher and per data partner scored against what actually ran, rule checks expressed as plain sentences, and a holdout slice of campaigns left untouched so recommended reallocations get judged on realized outcomes rather than curve‑fit. No write access, no credentials, no new interface, historical daily logs only (their intra‑day feed is phase two, Beeswax first). Portfolio framing throughout, never single‑campaign: Pilcer at 29:27, "Josh and I don't think about individual campaigns very much. It's the portfolio."
The follow‑up is booked: Friday, August 7, 1:30pm. The three moves that convert the yes: the Phil and AJ walkthrough (the self‑running piece is the deck they hold), the S3 extract with a named owner and date, and the reverse demo scheduled. Friday is where all three get their dates; walking out of it with I02 and I12 answered is the whole meeting.
What Josh asked for is the surface: say the rule, skip the sprint. What sits underneath is a full product on the trace system, and it is worth showing them more than they asked to see.
The one engineering sentence that makes this trustworthy: the language model authors rules; it never enforces them. A sentence compiles into a typed, versioned rule (scope, condition, threshold, action, exceptions) that executes deterministically inside the governed container. Authoring is generative; enforcement is a state machine. That is why a rule can run against every log line at no marginal model cost, why the same inputs always produce the same verdict, and why an auditor can re‑run any decision the rule ever made.
| They asked for | The product underneath |
|---|---|
| Rules in natural language instead of the offshore backlog | The whole authoring loop: sentence, compile, read‑back with examples, dry‑run on their own history, then arm. "Why don't you have it yesterday?" becomes: you have it in the same conversation, already backtested. |
| Exceptions, "blah blah blah" | Exceptions as first‑class decisions. "Except during political windows" or "snooze until Friday" is a typed modifier with an owner and an expiry, written to the record like any other decision. And exceptions get graded: one that keeps renewing is a policy change trying to happen, surfaced monthly with its cost attached. |
| Catching mistakes after they happen | The reverse direction, pre‑draining the backlog. We mine the Salesforce order instructions and the historical violations they already log, and propose the rules their operation already implies. The first fifty rules arrive written; their team edits sentences instead of filing tickets. |
| A violation alert | Every verdict on the trace. Each evaluation is a signed event: which rule version, what inputs it saw, the verdict, any exception applied, and the outcome once known. "Why did this flag" is answered in one click, forever, and a rule's whole history survives its own edits. |
No freeform model judgment ever sits on a money path: the model writes and explains, the compiled rule decides. No rule arms itself; a person confirms after the dry run. And the dry run is not optional ceremony, it is the same backtest discipline as the POC: a rule meets their history before it meets their operation. Said to Pilcer's team in their terms: the offshore pipeline goes into a state‑machine box, and the box keeps their issues table, their workflow, and their control.
The room is Unified, iHeart's in‑housed ad‑tech arm. Their stack: Triton (the majority, also iHeart‑owned), Beeswax, MadHive.
| Who | Seat | What they need to hear |
|---|---|---|
| Josh Backer | Co‑founder & COO of Unified, in‑housed as iHeart's data and ad‑product arm; credited by name on AudioGraph and the Magellan deal | Plain‑language rules replace the backlog ("every new type of rule needs to be established as code… I'm interested in natural language"), contribution value per publisher and data partner, and a referee that lets him defend AudioGraph's numbers upward without hiring a rival optimizer. |
| Matt Pilcer | CTO of Unified since 2024; enterprise‑implementation background; Backer's execution counterpart | Headless, no new interface (his stated requirement), the S3 extract of closed orders, one intra‑day log feed as the only eventual dependency, and his offshore rules pipeline put into a box that works without the backlog. |
| Lisa Coffey | Chief Business Officer; the executive voice on AudioGraph | The advertiser‑facing claim gets a referee behind it. Likely sponsor when this scales past the POC. |
| John Rosso | President, iHeart Technology Solutions; CEO of Triton Digital | Anything near pacing or allocation crosses his stack. Referee‑not‑player, brought in early rather than around; Triton‑majority reality means his door is on the path, not beside it. |
| Rich Bressler | President, COO, CFO | Quarter‑end allocation is a revenue decision; "deliver in full at lower cost" is his sentence before it was ours. |
| Phil · AJ · ad ops | The named next‑step audience from the call | The self‑running walkthrough, then the order‑picking session. They choose the campaigns; the proof belongs to them. |
Do not repeat "full security review done": it was hedged on the call ("One of us did?"); verify the NDA status first. Confirm MadHive logs before claiming them (Pilcer: "I have to double check"; Beeswax and Triton are confirmed). Josh's "five percent more margin by dropping two publishers" stays attributed to Josh as his own hypothetical, never presented as measured. The dashboard shown was demo data, said once. And the re‑derivable‑score, open‑models claim was not demonstrated on the call; it needs backing ready before a technical audience pushes on it.
The interface. Each one buys a piece of the build; most convert a verbal yes into a date, an owner, or a schema. Grounded in the transcript and the research.
| # | Question | What it buys |
|---|---|---|
| I01 | What exactly is signed today (NDA, security), and what does your team need from us for the S3 extract to clear? | The paper path, verified instead of assumed |
| I02 | Who owns the extract on your side, and what date works? | The verbal yes becomes a calendar entry |
| I03 | Which closed orders would Phil and AJ most want scored first, and what would genuinely impress them? | Their proof, their pick; the acceptance audience chooses the material |
| I04 | What does the log schema look like per platform: Beeswax we know, what grain does Triton export, and can you confirm MadHive? | The honest coverage read before any claim |
| I05 | What is the outcome definition per order type: delivered in full against which measure, maturing over what window? | The objective everything prices against |
| I06 | What are the first ten rules your team would write as sentences, today, if writing were free? | The editor's seed list, and the check on our reverse‑mining |
| I07 | Which violations from last quarter cost the most, and which rule would have caught each one? | The backlog, priced; the dry‑run demo picks itself |
| I08 | What does the issues table expose (create, snooze, escalate), so verdicts land in your workflow with no new interface? | Pilcer's headless requirement made concrete |
| I09 | How do makegoods and ADUs show up in the data, and who decides their placement today? | The Tetris door, opened gently for phase two |
| I10 | What would make AudioGraph's results independently defensible to an advertiser who pushes back? | Backer's exposure, offered a referee rather than a rival |
| I11 | When does the Triton conversation happen, and who makes the introduction to Rosso's team? | The technical home, entered early rather than around |
| I12 | At the end of the backtest, what does success look like, in whose words, shown to whom? | The acceptance rule written by its owners; the yes becomes their conclusion |
I02 and I12 are the meeting inside the meeting: a date and an acceptance rule convert everything else from conversation into work.
The balance sheet behind their sentence.
FY2025 revenue $3.865B, flat, with Multiplatform (broadcast) down 4.2% and Digital up 14.2%; podcasting leads the industry at $563.7M, up 25.6%. Q1 2026 brought the first GAAP operating profit in years and programmatic at roughly $200M growing about 50%. The Dec 2024 debt exchange cut over $440M and pushed maturities to 2029–2031 at hard coupons, so free cash flow serves debt and two cost programs run at once ($100M in‑year plus $50M new). Leadership: Pittman and Bressler at the top, Conal Byrne on Digital Audio, and the seat that matters most to this thread: John Rosso, President of iHeart Technology Solutions and CEO of Triton Digital. The 2026 strategic story is broadcast entering the digital TAM: DSP‑buyable radio since December, AudioGraph since June, industry‑wide in 2027, with a "Guaranteed Human" content pledge running above it all. A company selling verifiable trust in its content is a company primed to buy verifiable trust in its buying.
Broadcast by hand, digital by optimizer, and the seam between them unowned.
Broadcast sells on Nielsen‑rated audience guarantees. When delivery misses, makegoods and audience‑deficiency units repay the gap in free spots; political preempts what was scheduled; bonus spots fill unsold avails. Placement is decided by traffic desks (consolidated into central Centers of Excellence), and quarter‑end means hand‑reshuffling repayment spots against remaining avails before the books close. On the digital side, Unified's optimizers move pacing, bid, and budget across Triton (the majority), Beeswax, and MadHive; log data is assembled by hand today and used for reach and frequency, not optimization. Every decision class from the other two arcs has its analog here at the seam: which publisher earns the next slot, what a makegood is worth against which advertiser, which rule pays for its backlog slot, and what the quarter‑end reshuffle actually costs. The machine's own growth plan (DSP pipes plus AudioGraph's predictive audiences) mints new predicted‑vs‑realized gaps every week, and calibrating them is a service to their story, never a challenge to it.
The corpus and the call, with the lines worth repeating verbatim.
| Asset | What it carries | Where |
|---|---|---|
| The checkable deck"What the Listen Was Worth" · 10 slides | The two asks from the call; the seam ("your optimizers pull pacing, bid, and budget levers; we move inventory"); governance in, events out; headless with the plugin fallback; recommendations as expiring options; the three asks (backtest, reverse demo, the Phil‑and‑AJ walkthrough) | the deck, live |
| The Jul 30 transcript32 minutes, complete; copies (0)–(5) are prefixes of the same call | The primary source: Josh's backlog pain (10:58), Pilcer's headless requirement (14:05), the plugin exchange (15:38), the inventory‑levers seam (25:20), Tetris (26:37), the portfolio frame (29:27), and the yes (30:14) | the annex |
| The demo dashboard | What Matt walked live: governance module, CSV constraints in, aggressiveness packs, per‑segment recommendations including "turn this one off," everything logged as events | shown on the call |
Lines to reuse, exactly as said: "they're not pulling inventory levers" · "lower in cost while delivering the campaign in full" · "Tetris combined with the loudest person in the room" · "never ending" (the backlog) · "why don't you have it yesterday?" · and ours that landed: the expiring option, the hedge‑fund backtest, and the offshore pipeline "into a state‑machine box that 95% of the time works."
Their words, plainly, so we speak radio and ad tech both.
| Their word | What it means, plainly |
|---|---|
| makegood · ADU | Free spots that repay under‑delivered audience; the audience‑deficiency unit is the accounting form of the apology |
| avails · bonus spots | Unsold slots, and the freebies that fill them |
| preemption | A higher‑rate or political order bumping a scheduled spot |
| quarter‑end Tetris | Hand‑reshuffling makegoods against remaining avails before the books close |
| pacing · bid · budget levers | What Unified's optimizers already move, and keep |
| inventory levers | Which publishers and slots get the spend; untouched today, and the seam we fill |
| headless | Our logic inside their screens and workflow; no new interface |
| expiring option | A recommendation priced with a use‑by time, instead of a report read after the money moved |
| contribution value | What each publisher or data partner actually added, relative to cost |
| the five‑field record | A decision written down completely: what was known, chosen, predicted, what happened, what it taught; signed |