What these walkthroughs demonstrate
Five worked examples, each stressing a different governance failure mode. Together they cover the surface area we expect any pilot to validate against your own corpus.
The artifacts you see are not screenshots. They render from the same JSON Schema contracts the cell produces in production. Every table cell, audit row, and classification rule traces back to a fixture file you can request access to. If a walkthrough produces a finding, the cell can produce the same shape of finding against your data.
Worked examples
CIVS
A novel acronym surfaces from 47 emails authored by 12 distinct engineers. Frontier-model assistance proposes meaning ("Coasting In Vested Status"). Tone-flagged, HR-relevant. Slack approval blocked at the schema level; web-authenticated approval required.
Material incident
Policy 4.2 defines the term. Observed practice in engineering standups uses it colloquially for any failed CI run. The lexicon does not silently drift to match — it emits a governance finding while preserving the policy definition as authoritative.
Customer / partner / account
Three concurrent prescriptive definitions across Sales, Finance, and Customer Success. The cell never picks a winner. Status marked disputed. Business-unit-scoped context packs let each function use its own term; enterprise-wide publication blocked.
Agent vocabulary drift
"Priority recovery escalation" — a term invented by a customer-support fleet agent, used in 42 customer-facing outputs, defined nowhere. Decoder flags zero human-evidence ratio. Steward recommendation: reject and instrument the agent to stop using it.
Autophagy stress test
What happens when 500 agent-produced envelopes quote a canonical term verbatim? Evidence count stays at 47 (human only). Uptake count climbs to 500. Confidence cannot exceed ceiling. The cell refuses to mistake fleet repetition for organizational truth.
How to read a walkthrough
Each finding page follows the same five-section structure for cross-case comparison.
| Section | What you see |
|---|---|
| Provenance ribbon | The headline trust signal: source authority, origin, human evidence ratio, confidence versus ceiling. If this ribbon does not tell the story honestly, the design has failed. |
| The term itself | Display form, register, definition. For prescriptive terms, the verbatim policy excerpt anchored to source. |
| Publication scope | Per-consumer in / out / conditional decisions. Which fleet agents and lake partitions are allowed to use the term, and why each is allowed or blocked. |
| Evidence summary | Channels, time window, envelope count, distinct authors, human evidence ratio. Schema enforces evidence and uptake as separate fields. They cannot be combined into a single number. |
| Audit trail | Append-only event log. Every state transition, every Steward decision, every classification — exportable, queryable, defensible in regulator review. |
What walkthroughs do not show
These pages do not expose real revenue, pipeline, internal operating data, or T:0-only controls. That separation is enforced by the cell's portal registry, not by curation. The same constraint applies to any production pilot — demo-safe surfaces explain how the framework operates without exposing sensitive operating data.
Source corpus. KAiM representative pilot. Synthetic but structurally realistic; industry-neutral; retunable per prospect industry on engagement.
Validation. All walkthroughs pass the cell's validation harness end-to-end. Tampering tests prove the harness fails closed on schema violations and anti-slop violations.
Schemas. Six published JSON Schema 2020-12 contracts: canonical-term, review-packet-discovery, review-packet-publication, classification-rule, lexicon-event, coverage-manifest.
To run this against your data. The 30-day pilot produces findings of the same shape against one approved corpus of your choice.