high-velocity-advertising

# High-Velocity Advertising (HVA) **A framework for creative discovery at machine speed.** HVA is a system for generating, deploying, reading, and retiring creative hypotheses faster than competitors **without destroying signal quality**, by treating the platform's learning phase as a real-time market maker for creative discovery. It belongs to the same family as **PIBER** (strategic narrative) and **IDCF** (product strategy): a named, mnemonic, teachable discipline. PIBER governs the story. IDCF governs the product. **HVA governs the spend.** This skill is the spec. The discrete rules below become testable conditions in the commands (`/hva`, `/hva-forge`, `/hva-lint`, `/hva-review`, `/hva-vault`) and the `hva-desk` agent, and the deterministic checks in `scripts/hva-score.py`. > **In this plugin, HVA is an engine, not a replacement.** It runs parallel to the slow-craft `/gtm` lifecycle. One company can run both. Match the engine to the conversion, not to the excitement (see `rules/guardrails.md` → Fit Boundary). --- ## Keystone Insight The constraint was never creative production or the patience the learning phase demands. It is **statistical resolution per dollar.** The edge belongs to whoever reads a true leading signal on clean data faster than the crowd, because volume collapses variance only around a *real* edge and amplifies it around a *false* one. You are not the latency arbitrageur trying to outrun the exchange. You are the **market maker**: you quote continuously, you give the algorithm the creative liquidity it needs, and you harvest the spread of discovery. You provide the options; the algorithm provides the reach; you read its first moves before the rest of the table looks up. The structure stays stable enough for it to keep learning. **Volume compounds a real edge and amplifies a fake one.** That asymmetry is why CLEAR exists. --- ## CLEAR — the decision rule Five letters, the heart of HVA the way P-I-B-E-R is the heart of a case board. Each is a testable condition; see `rules/the-desk.md` for thresholds and `scripts/hva-score.py` for the implementation. | Letter | Rule | One line | |---|---|---| | **C**lean | Trade only on server-side truth (CAPI, not platform-reported numbers alone). | No clean feed, no decision. | | **L**ead | Judge on leading indicators (frequency behavior, the three relevance rankings, cost per micro-conversion), never first-window ROAS. | Read the algorithm's first moves, not the scoreboard. | | **E**conomics | Kill at roughly **3× target CPA** spent with zero qualifying event. Thresholds are spend-relative to your CAC, never a borrowed impression count. | Bounded cost of being wrong. | | **A**symmetry | Cut fast on a single bad leading signal; scale slow, only after the true-conversion clock confirms. | Two clocks, deliberately mismatched. | | **R**eplication | A lone winner among many simultaneous tests is probably a multiple-comparisons ghost. Confirm with a correction (Benjamini-Hochberg) or a clean repeat before betting budget. | Don't crown a coin flip. | --- ## The Stack — five layers Read bottom-up. Each layer exists only to serve the keystone; remove one and the layer above it is trading on a lie. The **keystone** is bedrock, **Layer 0 is the gate**, **Layer 4 is the moat**. ```mermaid flowchart BT K["KEYSTONE INSIGHT<br/>Statistical resolution per dollar is the constraint,<br/>not production or patience.<br/>Volume compounds a real edge, amplifies a fake one.<br/>You are the market maker for creative discovery."] L0["LAYER 0: THE INSTRUMENT<br/>Clean signal. Server-side / CAPI truth.<br/>The gate. No clean feed, no framework."] L1["LAYER 1: THE FOUNDRY<br/>Creative supply. AI-generated diverse concepts.<br/>The fatigue-defeating edge."] L2["LAYER 2: THE STRUCTURE<br/>Deployment. Consolidated, funded ad sets.<br/>Five or fewer creatives each. Test inside stable ad sets."] L3["LAYER 3: THE DESK<br/>Decision logic. The Read Ladder plus CLEAR.<br/>Two clocks: cut fast, scale slow."] L4["LAYER 4: THE VAULT<br/>Compounding. Winners library, creative families,<br/>first-party data. The moat."] K -.->|"underwrites everything"| L0 L0 ==> L1 L1 ==> L2 L2 ==> L3 L3 ==> L4 classDef gate stroke:#16a34a,stroke-width:3px; classDef moat stroke:#d97706,stroke-width:3px; classDef key stroke:#64748b,stroke-width:2px,stroke-dasharray: 5 4; class L0 gate; class L4 moat; class K key; ``` | Layer | Name | What it is | In this plugin | |---|---|---|---| | **0** | **The Instrument** | Clean signal — server-side / CAPI truth feeding everything. The gate, not a feature. The platform undercounts real conversions by default, so a dirty feed makes you fast and wrong. | `skills/meta-ads/rules/pixel-setup.md`, EMQ 8+ rules in `skills/campaign-optimization/SKILL.md`, CAPI fields in `agents/data-analyst.md` | | **1** | **The Foundry** | Creative supply. AI generation producing genuinely distinct concepts (angle, hook, promise, proof, format), not variations of one idea. The durable edge: an inexhaustible supply that outruns fatigue. | `/hva-forge` → `agents/creative-director.md` + `skills/neuro-testing/`. See `rules/the-foundry.md` | | **2** | **The Structure** | Deployment. Consolidated, adequately funded ad sets; five or fewer creatives each; Advantage+ / CBO allocating. Fund each concept enough to reach decision volume in hours. | `skills/meta-ads/rules/campaign-structure.md` + `agents/campaign-operator.md`. **HVA shape:** one concept per ad, ≤5 ads in one funded ad set — so `insights get --ad-id` reads a single creative | | **3** | **The Desk** | Decision logic. Runs the Read Ladder and CLEAR. Converts the algorithm's early behavior into cut-or-scale calls at machine speed. | `agents/hva-desk.md` + `scripts/hva-score.py`. See `rules/the-desk.md` | | **4** | **The Vault** | Compounding. Winners library, systematic fatigue rotation, creative families bred from confirmed winners, and the flywheel of proprietary judgment plus first-party data. **This is the moat.** | `/hva-vault` → `.gtm/hva/vault/`. See `rules/the-vault.md` | The signals commoditize; the system and the data do not. Push the defensibility into the Vault by design. --- ## The Loop Forge, Deploy, Read, Decide (cut or compound), Refeed. Continuous. Every cycle's learnings re-enter the Foundry as sharper generation priors, so the supply gets smarter, not only faster. ```mermaid flowchart LR F["FORGE<br/>generate diverse<br/>creative hypotheses"] D["DEPLOY<br/>consolidated, funded ad sets;<br/>stable structure"] R["READ<br/>leading indicators,<br/>by volume not calendar"] DEC{"THE DESK: CLEAR<br/>cut or compound?"} X["KILL<br/>retire, or tag<br/>under-delivered"] V["THE VAULT<br/>winners library,<br/>creative families"] F ==> D D ==> R R ==> DEC DEC ==>|"compound: confirmed by the slow clock"| V DEC ==>|"cut: bad leading signal, fast clock"| X V -.->|"learnings refeed as sharper priors"| F X -.->|"retest best under-delivered in a cleaner batch"| F classDef desk stroke:#2563eb,stroke-width:3px; classDef vault stroke:#d97706,stroke-width:3px; class DEC desk; class V vault; ``` The Desk (blue) is the only decision point, and it runs CLEAR on every creative. Two exits, two clocks: a **fast** cut on any bad leading signal, and a **slow** promote that waits for the true-conversion clock before sending a winner to the Vault. Both exits loop back to the Foundry. **Three callers, one Desk.** The loop is driven by three interchangeable read mechanisms — the on-demand `/hva-review` command, the hourly `routines/hva-read-loop.md` cloud routine, and the `scripts/hva-watch.sh` local watch loop. All three invoke the *same* Desk → the *same* `scripts/hva-score.py` → the *same* autonomy gate, so safety behaves identically regardless of cadence. --- ## Autonomy modes The Desk's authority is configured per-account in `.gtm/config.json` → `hva.autonomy`. Default is the safest. See `agents/hva-desk.md` for enforcement. | Mode | Pauses (cut) | Budget increase (scale) | Use when | |---|---|---|---| | `recommend` *(default)* | Recommends only — human approves | Recommends only — human approves | Building trust; matches the plugin's PAUSED-by-default culture | | `cut-auto` | **Auto-pauses** losers (saves money, fully reversible) | Recommends only — human approves | You trust the cut logic but not unattended spend | | `full-auto` | Auto-pauses losers | Auto-scales winners, bounded by `scale_bid_cap_multiple` (0.7×) and `max_daily_budget_increase_pct` (20%/day) | Mature account, audited loop, real machine speed | Pausing only ever *saves* money and is reversible, so it is the safe direction to automate first. Raising budget *spends* money, so it stays gated longer. This mirrors CLEAR's **A**symmetry directly. --- ## The Fit Boundary HVA is a tool with a domain, not a universal law. - **Run HVA where it wins:** fast, cheap, well-tracked conversions. DTC, apps, info products, low-friction lead gen, fast purchase cycles, strong creative variation, and budget large enough to hit volume thresholds in hours. - **Do not run HVA where it loses:** high-ticket B2B, long sales cycles, tiny audiences, weak tracking, low-budget accounts, regulated categories, delayed or offline conversion. For the second list, run the **opposite motion** — the slow-craft `/gtm` lifecycle: a small number of high-craft concepts, reads measured in weeks not hours, lead quality scored by sales rather than lead cost scored by the platform, and the feedback loop running through closed-won data in the CRM. The HVA Phase-0 fit check **warns and routes** to `/gtm` when the account doesn't fit; the cloud routine can hard-block. Scoring rubric: `rules/guardrails.md`. --- ## Naming discipline The framework is **High-Velocity Advertising (HVA).** Do **not** call it High-Frequency Advertising: in the platform, *frequency* already names how often one person sees an ad, and the collision will confuse the operators you teach. --- ## Rule files | File | Contents | |---|---| | `rules/the-desk.md` | CLEAR checks in full, the Read Ladder (1–2k / 2–5k / 5k+ impressions), the Diagnostic Table, the two-clocks asymmetry. **The contract `scripts/hva-score.py` implements and is tested against.** | | `rules/the-foundry.md` | Diverse-concept taxonomy and variety-over-duplication; how `/hva-forge` reuses `creative-director` + neuro pre-testing. | | `rules/the-vault.md` | Winners-library schema, creative families, cross-campaign priors, fatigue rotation, the multiple-comparisons philosophy. | | `rules/guardrails.md` | The failure modes HVA exists to kill → the `/hva-lint` checklist; the Fit Boundary scoring rubric. | | `rules/evidence-base.md` | The receipts: learning phase, signal loss, creative count, divergent delivery, multiple comparisons. | ## Close Most operators still think in calendar time. They wait, they watch, they let the machine work over days. HVA is the refusal of that wait — not by outrunning the algorithm, but by feeding it more honest options than anyone else can, and reading its first moves before the rest of the table looks up. **Speed is not the edge. A real edge, executed at speed, is the edge.** The market maker does not predict the next tick. He is simply already there when it prints.