Build the base graph from public data and industry knowledge.

IW already carries a deep map of how a CPG company works — SKU pyramids, channels, trade promo, distributor economics, festivals, monsoon. NorthCo binds to that map. We then layer in everything that's already public about NorthCo — annual report, investor deck, org chart, product catalog, twelve months of press — so the graph reshapes into NorthCo specifically. The output is a base graph: structurally complete, no internal data yet.

What happens

1. Pick the industry, sub-vertical, geography. CDO selects CPG → Foods + Personal Care → India. Takes 15 minutes.
2. The CPG brain activates. 8,400 standard things every CPG has — SKU hierarchy, channel types, promo mechanics, festival calendar, weather effects on demand — all become available as templates.
3. IW reads NorthCo's public footprint. Annual report (320 pages), Q3 investor deck, public org chart, product catalog scrape, 12 months of press releases.
4. The CPG template reshapes into NorthCo. Generic "Brand → Pack → SKU" becomes NorthCo's actual brands. Generic "geographic zones" becomes NorthCo's actual North/South/East/West. Competitors get named: Haldiram, PepsiCo, ITC Foods, Britannia.
5. 30-min CDO review. Confirms anything the public sources got wrong (typically 2–4 corrections).

Who does what

IW does the loading and extraction (4 days wall-clock, runs in background). The CDO spends 1 hour total — 15 min to pick the industry, 30 min to review the V0, 15 min to confirm the four corrections.

Public sources used

SMEs and consultants thrash out the graph with NorthCo's stakeholders.

The base graph from Week 1 is roughly right but generic. Now SMEs and consultants thrash it out with NorthCo's actual stakeholders — Anuj (demand), Vinod (supply chain), Ritu (category) — fixing what's wrong, adding what's missing, and confirming what's specific to how NorthCo operates. IW asks only the questions that move the graph; the SMEs answer in plain conversation. Three stakeholders × two sessions × ~80 min each. No code, no schema design.

How IW asks only useful questions

1. It scores every part of the V0 for uncertainty. Where is the binding weak? Where do public sources contradict? Where does the CPG template offer multiple options?
2. It generates only the highest-value questions. Each question is scored by how much the graph would change based on the answer. Top 12 get asked.
3. It routes each question to the right person. Demand questions go to the demand lead. Supply questions to supply chain. Categories to category managers.
4. It schedules around their day jobs. 80-min sessions, two per SME, spread across the week.

Real session — Anuj K., Demand Lead

What got added

• Two new SKU hierarchy levels — Variant and Pack-Configuration
• Six named distribution centres — Mumbai, Pune, Delhi, Chennai, Bengaluru, Kolkata
• Seven state-level geographies under the four zones
• Two more named competitors — Parle, Mondelez
• Owner mappings — who in NorthCo owns each part of the graph

Connect data to every node — IW writes its own connectors and resolves conflicts. (Hydration.)

The refined graph from Week 2 has structure but no real numbers in it yet. This week IW hydrates every node — plugs into all 38 of NorthCo's systems (SAP, Oracle, Snowflake, Excel, Tableau, PowerBI, MCP servers), writes the connector code itself, and stops when two systems disagree on the same number. The SME picks which source is canonical. Then the ontology compiles: 184 derived metrics, full lineage on every one, no recomputation at query time.

Sources connected (all 38)

28 more sources abbreviated · full list at iw://northco/sources/

Worked example: how one node — cogs_per_sku — gets populated

To make the data ingestion concrete, here is exactly how IW populates a single node from three disagreeing systems.

Same pattern, 22 more times

The same flow runs for the other 22 conflicts that exceeded tolerance — supplier lead time (SAP MM vs Oracle), DC capacity (WMS vs planning sheet), promo uplift (Tableau Nielsen-backed vs trade marketing's sheet), customer OTIF (PowerBI golden record vs SAP SD), distributor margin (trade terms doc vs Snowflake), and 18 more. Mean time-to-resolution: 18 minutes per conflict.

Then the ontology compiles

Once all base nodes are populated, IW computes the 184 derived metrics — forecast, fill rate, OTIF, days-on-hand, contribution margin, breakage rate, and so on. Each derived value stores its formula, the source nodes it came from, and the version of each source binding. Forward and reverse lineage is materialised — no recomputation at query time.

Morrie elicits tacit knowledge from practitioners — the rules they use that no system records.

The hydrated graph shows Mumbai breakage at 18% versus a 4% network average. The data doesn't say why. Morrie — IW's elicitation agent — asks Sneha, the practitioner who owns that node. She tells him: monsoon humidity warps the cardboard divider, the new ASRS aisles cause forklift hits on turn-3. That becomes a typed rule on the graph: +14% safety buffer for Mumbai monsoon glass. Twenty-three such sessions across NorthCo's practitioners. The graph absorbs the operating wisdom that lives in people, not systems. Then IW commits live.

How Morrie picks who to ask, what to ask

1. Detect anomalies. For every metric, flag values that deviate from peer-group, supergraph priors, or historical baseline beyond a configured threshold.
2. Generate one question per anomaly. Plain language. Includes the data context — so the practitioner knows what triggered the question.
3. Route to the right practitioner. Based on owner mappings (from Week 2), org chart proximity, and prior session history. Morrie balances workload — never asks one person more than 4 questions in a week.
4. Ask for a 15-minute window. No mandatory meeting. Async (typed) or live (voice transcribed). Practitioner chooses.
5. Adapt as it goes. Morrie's next question depends on the previous answer. Graph state visible alongside the chat — practitioner can correct it directly.
6. Commit as a typed rule. Each insight becomes a TacitRule with a trigger predicate, an action, an owner, and explicit links to the nodes it affects.

Real session — Sneha K., Mumbai planner · 14 min

Anomaly: breakage_rate(dc=mumbai, sku∈glass_bottles) = 18.4% vs network mean 4.1% (z = 4.7)

Some of the 25 typed rules captured

• tacit/418 · Mumbai monsoon glass +14% safety buffer (Sneha)
• tacit/422 · Festival demand 8-week pre-build pattern (Anuj)
• tacit/491 · Demand planner override after long-monsoon weeks (Vinod)
• tacit/507 · Personal Care humidity-sensitive SKU substitution (Ritu)
• tacit/623 · Cricket schedule beverage spike, IPL window (marketing)
• tacit/641 · South India filter-coffee seasonality (regional sales)
• tacit/688 · Modern trade Q4 negotiation pattern (key accounts)
• tacit/712 · Distributor channel-stuffing detection (sales ops)
• tacit/734 · Diwali credit terms exception window (finance)
• tacit/751 · ASRS Mumbai turn-3 forklift hit pattern (Sneha)
• 15 more · full list at iw://northco/tacit/

Then IW goes live

End of week 4: validation passes confirm structural integrity, lineage completeness, and tacit-data linkage. IW is released as iw://northco/v1.0.0. Read access to 23 named users across demand, supply, and category. Write access only via SME-vetted action types. CDO + Demand Lead sign-off in 52 minutes.

IW writes its own orchestrator, traversal, and planner agents. Your team builds custom agents on top.

With the live IW in hand, IW generates its own standing set of orchestrators (route a task into sub-tasks), traversals (walk a sub-graph and return a structured result), and planners (propose action sequences under a goal). 14 OTP agents in total — Demand-Forecast, SKU-Health, Stockout-Risk, Trade-Promo, and 10 more — derived from the graph itself and honouring the tacit rules from Week 4. Engineers approve 11 the same week; 3 stay in review. Then the agent builder opens to designated business users. Anuj ships his first custom agent in 18 minutes, no engineer required.

The three agent kinds

• Orchestrator — receives a task or query, decomposes into sub-tasks, routes results. Example: DemandForecastOrchestrator.
• Traversal — walks a defined sub-graph and returns a structured result. No external action. Example: SkuHealthTraversal — walks SKU → COGS, margin, velocity, breakage.
• Planner — given a goal and constraints, proposes an action sequence. Calls action types under the user's permissions, never silently. Example: TradePromoPlanner.

The 14 OTP agents IW generated

4 more abbreviated · full registry at iw://northco/agents/

Then the agent builder opens

Designated business users — initially 9 across demand, supply, category, and trade marketing — get the agent builder. Custom agents declare a trigger, the OTP agents they inherit, the actions they may invoke, and the owner. Inherited OTP agents carry their tool contracts and tacit-rule honoring forward — the new agent does not re-encode any of that.

Anuj's first custom agent — 18 minutes

Live to the demand desk by lunch. The monsoon buffer (tacit/418) is honoured automatically because StockoutRiskOrchestrator already honours it — no re-encoding.