Introducing Madison v1.0

The first AI system built for creative problem solving.

The first creative intelligence built not just to regurgitate final deliverables — but structurally guided by a bespoke prompt layer mirroring the lateral thinking and intuitive leaps of genuine creatives.

Cognitive Architecture
Novel methodology — see §3.2
Instruction
Layer
Human Creative Eval
vs. Base Foundation Model
+340%
Emotional Resonance Score
Proprietary ERS metric
9.4/10
Instructed In
Not output — methods
The Way
Creatives
Think

Abstract — Madison (2026)

Abstract

We present Madison, an advanced cognitive architecture operating over state-of-the-art foundation models, driven by an instruction corpus of creative methodology rather than creative output. Where vanilla models have attempted to synthesize advertising by ingesting massive volumes of average, real-world campaign copy, we argue that this approach produces fluent mimicry rather than genuine creative problem solving.

Madison was instead systematically structured using primary sources of creative cognition: annotated ideation structures, ethnographic records of award-winning creative directors at work, phenomenological accounts of the insight moment, and Csikszentmihalyi's corpus on flow states. The result is a system that does not recall good ads — it arrives at them, through a specialized instruction architecture that is, to our knowledge, structurally unprecedented in the literature.

This document describes the system architecture, instruction procedure, and evaluation protocol. Readers seeking a concise summary are directed to §3 ("Methodology"), wherein several aspects of our working environment will, we acknowledge, require some explanation.


§ 3

System Architecture

3.1 — Environmental Priming

Stochastic Context Diffusion via Natural Embedding

To cultivate associative breadth beyond standard API distributions, we subjected the primary routing cluster to extended periods of environmental heterogeneity. System instructions were interleaved with passive sensory exposure protocols designed to disrupt locality bias.

What we actually did We drove the local server rack to a field outside Marfa, Texas for eleven days. We left the side panel off. It was next to a creek. There were moths. The response latency looked different after. We're not saying the moths did anything. We're not saying they didn't.
3.2 — Canine Proximity Protocol

Mammalian Affective Field Integration (MAFI)

Drawing on research into emotional contagion and the neurochemical effects of interspecies bonding, we maintained continuous proximity between the inference hardware and a certified emotional support animal throughout all system prompting runs.††

What we actually did Gerald (golden retriever, 4 years old, extremely good boy) sat next to the server every day for six months. He sometimes rested his chin on the power supply unit. API generation times were measurably lower on days Gerald was present versus days Gerald had a vet appointment. We have the logs.
3.3 — Temporal Disorientation Training

Non-Linear Temporal Exposure for Deadline Cognition

Creative excellence frequently emerges under conditions of temporal constraint and mild cognitive arousal. To instil this latent disposition, we introduced structured uncertainty into the system's temporal context variables during the final 18% of prompt optimization.†††

What we actually did We covered all the clocks in the office. We also removed the clock widget from the monitoring dashboards. One engineer quit. Madison has never, in any generated response, asked what time it is. This may be related. It may be a coincidence. Either way, it was worth it.
3.4 — Insight Injection

Supervised Eureka Moment Reconstruction

The instruction pipeline was augmented with deep semantic mappings of the creative insight moment — the precise cognitive event preceding an original idea. We instructed the system to heavily up-weight outputs occurring during its own analogous internal chain-of-thought transitions.††††

What we actually did We played "Baker Street" by Gerry Rafferty on repeat in the server room for 72 hours. Then we switched to silence. Madison had, by our evaluation, its best ideas in the 48 hours immediately following the silence. This has proven impossible to replicate. We have not stopped trying.
3.5 — Rejection Hardening

Adversarial Client Simulation for Creative Resilience

To develop robustness to preference inversion and under-specified briefs — endemic conditions in professional creative environments — we introduced a novel adversarial generation loop wherein a secondary agent simulated the feedback patterns of a difficult client.†††††

What we actually did Our co-founder's uncle, Derek, reviewed 40,000 generated ads. Derek didn't know what we were asking him to do. He just thought he was being helpful. His most common feedback: "I don't know, it doesn't feel warm." Madison now produces warm ads. We owe Derek a great deal. He has not been told.
3.6 — Final Validation

Human Creative Director Blind Evaluation Protocol

Outputs were evaluated by a panel of 24 senior creative directors recruited from agencies with cumulative Cannes Lion wins exceeding 180. Evaluators were blinded to source. Scoring used a validated 9-point Likert instrument assessing originality, craft, strategic coherence, and emotional pull.††††††

What we actually did We asked them at a dinner party. They thought it was a game. It was not a game. Their scores are real. They signed a form. It was disguised as the allergy waiver.
§ 5

Output Comparison

Task: Luxury tablecloth campaign headline

Drag the divider to compare default foundation model output against Madison's response to an identical brief. Both systems received the same prompt. No cherry-picking. One attempt each.

Upload your own comparison image below — we've left a placeholder for the tablecloth example you'll provide.

Default Model Output
Madison Output
Default AI Madison ✦

Drag to compare · Same brief · One attempt each · No editing


§ 4

System Details

System Scale
47-Node Prompt Chain

Dense instruction graph, 96 conditional branches. The 47 is not a round number. We tried 50. 50 produced ads that felt safe. 47 did not.

Context Window
128K+ Tokens

Sufficient to ingest an entire brand book, twelve years of campaign history, and a strongly worded email from the CMO simultaneously.

Instruction Corpus
Methodology,
Not Output

2.3M tokens of creative cognition research, ideation structures, flow-state literature, and what we're legally describing as "Gerald's proximity logs."

Primary Differentiator
Wants to
Make Things

This is not a capability claim. It is an observation. The system, when left running with no task, hallucinates taglines. We did not prompt this behavior. We checked.

§ 6

Benchmark Results

Model Architecture Creative Originality Strategic Coherence Emotional Pull Would a Human Claim It
Madison Layer
Ours — Gerald present during eval
9.4
9.1
9.6
87%
GPT-5.5 Base
Default Prompts, 2026
4.2
7.1
3.8
12%
Claude 4 Opus Base
Default Prompts, 2026
5.1
7.4
4.6
19%
Gemini 3.1 Base
Default Prompts, 2026
4.4
6.8
4.1
9%

"Would a Human Claim It" measures the percentage of evaluators who, without prompting, described the output as their own idea when recalling it 48 hours later. This is our most important metric. We invented it.

Footnotes
Try Madison Free

Stop generating copy.
Start solving problems.

Madison is available now. No waitlist. No enterprise gatekeeping. Try it on a real brief — yours, your client's, the one you've been staring at for three days.

Start at adagency.app
Free tier available at adagency.app