Reproducibility
The same frozen input produces the same graph.
AIKG / EVIDENCE SYSTEM01 / 15
01 / Advertising Intelligence Knowledge Graph
From archived data to evidence-backed intelligence.
A deterministic system for reconstructing advertising history—and knowing when its answers deserve to be trusted.
AIKG
02 / Product thesis
Can historical advertising data become more useful than a searchable archive?
Award sites, campaign pages, agency credits, production records and archived pages contain fragments. Customer value lives in the relationships between them.
A catalog answers
- Where does this name appear?
- Does this campaign exist?
- Who received credit?
Intelligence must answer
- Which relationships persist?
- Which offices produced the work?
- What evidence supports the conclusion?
03 / Information architecture
The value is in the connections.
Select a node to see how AIKG turns an isolated record into a traceable claim.
SELECTED / CAMPAIGN
A campaign is a junction, not a document.
It connects brands, agencies, people, roles, markets and media—while retaining the evidence that justifies each relationship.
campaign relationshipassertionevidence bindingarchived source
04 / Deterministic reconstruction
The graph is a rebuildable projection—not the source of truth.
Select any stage. Every transformation remains inspectable and replayable from frozen source evidence.
01 / CAPTUREFrozen HTML and source snapshots remain independent of downstream interpretations.
05 / System properties
Determinism is a governance choice.
Auditability
Every suspicious answer can be traced back to source evidence.
Debuggability
Changes can be isolated to acquisition, parsing, identity or interpretation.
Governed AI
LLMs assist investigation without silently defining the graph.
answer→relationship→assertion→evidence→source
06 / 0.7.11 generation
Scale arrived after the evidence contract.
The early demo proved six bounded intelligence questions on 420 pages. The same architecture then expanded without changing the trust model.
380,753UNIQUE NODES
1,197,473UNIQUE RELATIONSHIPS
2,387,207EVIDENCE RECORDS
68,675SOURCED PAGES
240,104 individuals62,068 campaigns47,203 agencies30,842 brands500K+ CONTRIBUTED_TO edges
07 / Customer intelligence
The graph began surfacing commercially useful patterns.
Repeated collaboration, network overlap and account history become queryable rather than anecdotal.
DDB→Volkswagen
DDB→McDonald’s
Leo Burnett→McDonald’s
TBWA→Nissan
Publicis→Renault
Saatchi & Saatchi→Toyota
Cheil→Samsung
Wieden+Kennedy→Nike
But these signals also exposed the system’s next, more difficult problem.
08 / Discovery
The moment that changed the project
Every structural check passed. Yet the highest-ranked “individuals” included a school and a city — Miami Ad School and Hamburg. The graph could perfectly prove where a fact came from while being wrong about what it meant. That shifted the project from graph construction to entity truth.
0.7.11.3.8 · zero endpoint failures · zero excluded-page leakage — the validation worked; it exposed the semantic layer as the next frontier.
Provenance
is not
interpretation.
Miami Ad SchoolHamburgAlter EgoMPCCo3
09 / Presentation-layer risk
A valid graph can still produce a misleading ranking.
Raw “top individuals”
25
Structurally valid records returned by the graph.→ENTITY-TYPE
QUALITY FILTER
QUALITY FILTER
Presentation-ready
16
Nine entries were held back from the presentation layer — a heuristic flag, not a verdict. Obvious misclassifications (Miami Ad School, Hamburg, Toronto) were separated from cases that still need evidence (Alter Ego, MPC, Co3, The Mill, Getty Images).
10 / Bounded investigation
The frontier shifted from graph scale to entity truth.
Type correctness
Is this actually a person, agency, brand, geography or another semantic type?
Identity correctness
Same entity, presentation variant, network and office, parent and subsidiary—or unresolved ambiguity?
Interpretation quality
Could a technically valid result still give a customer the wrong conclusion?
11 / Information architecture, not cleanup
Entity resolution is often relationship modeling.
BBDO network
BBDO New York
BBDO
AMV BBDO
Preserve organizational distinctions while making network relationships explicit. Ambiguity remains visible instead of being erased.
12 / Four layers of trust
Quality became a layered system.
01Structural validationDoes the graph reproduce correctly?
02Semantic diagnosticsDo entity types and relationships look plausible?
03Evidence calibrationDoes original context support the conclusion?
04Human adjudicationCan a reviewer determine what the record represents?
These layers are kept separate instead of being collapsed into a single, misleading confidence score.
13 / Evidence before metrics
A 93-record corpus—stratified by what the evidence can support.
22direct and unambiguous
27direct, source-match ambiguous
25derived, suitable for type review
19evidence-insufficient
No precision number was invented.
Only evidence-sufficient, genuinely human-reviewed records can contribute to the first precision measurement.
49 direct HTML contexts49 original credit lines42 credit-group headings27 ambiguous multiple matches
14 / Trustworthy AI intelligence
Five principles that travel beyond advertising.
- 01Preserve original evidenceRetain enough source structure to investigate every fact.
- 02Separate observation from inferenceA source credit and an inferred relationship are not equivalent.
- 03Treat ambiguity as data“Unresolved” can be the most accurate state.
- 04Validate interpretationsReal customer questions expose what infrastructure tests miss.
- 05Use AI as investigatorAssist judgment without becoming an invisible authority.
PRODUCT STRATEGY · INFORMATION ARCHITECTURE · DATA ARCHITECTURE · QUALITY SYSTEMSJAY SETHI / AIKG
15 / Outcome
A reproducible evidence system — and an operating model for trustworthy AI.
AIKG now includes a deterministic rebuild pipeline — 380K entities · 1.2M relationships · 2.4M evidence records · 68K source pages, rebuildable from frozen sources; versioned interpretations kept separate from the evidence layer; a four-layer quality system spanning structural, semantic, evidentiary and human review; and a 93-record human-calibration corpus.
“The hardest problem is not generating an answer. It is preserving enough structure, evidence and uncertainty to know whether the answer deserves to be trusted.”
380K entities · 1.2M relationships · 2.4M evidence records · 68K source pagesJAY SETHI — PRODUCT STRATEGY · INFORMATION ARCHITECTURE · SYSTEMS DESIGN
01 / 15