Enterprise AI evidencePublic-safe · evidence current 2026-10-07

Verified AI Capability

Four personally built systems mapped across workflow, source boundaries, ownership, testing, failure modes, monitoring, and adoption. Unknown outcomes stay labeled instead of becoming portfolio mythology.

Verification lens

One evidence model across very different systems

Built + shippedSources + actionsReal-intent testingFailure diagnosisKnown ownerMonitoringAdoption outcomes

Four systems, one evidence standard

What the work proves — and what it still does not

Each system keeps its strongest proof next to the evidence that is still missing, so confidence and uncertainty stay visible at the same time.

System 01

CSI New Analyst Toolbox

Versioned, in-workflow analyst capability with explicit authority boundaries and a monthly update path.

Shipped + maintained

Strongest proof

Versioned product ownership, in-workflow support, authority boundaries, and a maintained release path.

Still unmeasured

Active users, repeat usage, time saved, quality movement, and runtime telemetry are not yet measured.

User / workflow: New and developing CSI analysts investigate and label messages without leaving the case page. A floating Chrome extension combines pace tracking, threat-intel lookups, queue and SOP guidance, case tracking, and on-device translation.
Sources / actions: Reads supported page context; surfaces SPF, DKIM, DMARC, domain, URL, DNS, and registry evidence; provides queue-aware guidance, safe links, local state, and CSV/JSON export. Approved internal guidance controls; team discussion only corroborates scoped decisions.
Scope limits: No autonomous labeling. Official SOPs, queue rules, escalation paths, and analyst judgment remain authoritative. Snapshot evidence is labeled non-live; unresolved conflicts escalate; exported files may contain work data.
Owner: Marquetta Moore — product concept, workflow research, architecture, implementation, documentation, testing, and 30+ releases. Human guide owners retain policy authority.
Test method: Regression coverage across manifest and permissions, labeling rules, migration, fallbacks, translation, secure CSV, accessibility and keyboard use, narrow screens, and packaging using offline fixtures.
Failure modes: Blocked DNS or RDAP, stale or conflicting guidance, missed page regions, import type mismatch, hidden navigation state, and managed-browser policy. Live authenticated selectors and enterprise-network variation remain partially unverified.
Monitoring: Versioned changelog plus feedback form. A daily rule monitor feeds a human-reviewed monthly preview queue, month-end release checks, and first-of-month changelog rollover. Runtime telemetry is not yet measured.
Adoption: Audience and workflow are documented. Active users, repeat usage, time saved, QA disagreement, and behavior change are not yet measured.

Internal guide build v3.34 · 30+ releases

Open case study

System 02

Launchpad

Reusable instruction system with explicit routing, portable state, security controls, and repeatable outputs.

Active · v2.1 / 2.1.0

Strongest proof

Reusable instruction architecture, explicit routing, portable state, and security-conscious artifact generation.

Still unmeasured

Formal intent benchmarking, semantic consistency, usage telemetry, and user outcomes are not yet measured.

User / workflow: Career changers or applicants provide a target role plus a resume, LinkedIn profile, or work samples. Launchpad scores readiness, maps skill, proof, language, and positioning gaps, assigns one task, and updates a command center.
Sources / actions: Uses user-provided evidence and role requirements; routes across assessment, progress, materials and content, and momentum flows. Generates self-contained HTML storing goals, streaks, score history, completed work, and content pipeline.
Scope limits: Career-readiness support, not a hiring prediction or guarantee. Scores must follow evidence; missing or conflicting inputs are disclosed. No database; the user retains the portable file. Links are limited to HTTP/S and stored content is escaped.
Owner: Marquetta Moore — Cowork skill v2.1 and ChatGPT plugin v2.1.0; aligned flow logic, references, dashboard, packaging, documentation, and changelog.
Test method: Security and scope audit verified HTML escaping, URL and status allowlists, closing-script protection, Flow 2 versus Flow 4 routing, v1 migration, state separation, and half-score support. Formal intent benchmarking is not yet measured.
Failure modes: Wrong flow for ambiguous requests, missing role or profile, conflicting evidence, stale or absent dashboard, unsafe content insertion, cross-platform drift, and inflated scoring. Guards exist; semantic consistency across runs is not yet measured.
Monitoring: Every tracked flow updates visible state; score history and the content pipeline preserve change reasons. A versioned changelog exists. Usage telemetry, error logging, and post-launch review cadence are not yet measured.
Adoption: Installable in two formats. Unique users, return rate, task completion, score movement, and job-search outcomes are not yet measured.

Cowork v2.1 · ChatGPT plugin v2.1.0

Open case study

System 03

Aegis

Formal launch-gate evidence: intent test, scoring rubric, failure taxonomy, measured rates, and a no-go boundary.

Evaluated prototype

Strongest proof

A 25-intent evaluation, scoring rubric, failure taxonomy, measured rates, and an explicit launch boundary.

Still unmeasured

Live semantic quality, provider reliability, recurring users, and business outcomes are not yet measured.

User / workflow: A creative or marketing reviewer supplies a public competitor URL or source text. Aegis extracts, analyzes the offer, audience, hooks, and claims, generates five concepts, and runs a separate risk review with rewrite options.
Sources / actions: Public URL or manual text via Firecrawl or a saved sample; Gemini stages for strategy, concepts, and Shield review; Zod contracts, stage-specific fallbacks, preserved partial results, and visible human review.
Scope limits: Deployed prototype and portfolio demo only. Shield does not guarantee platform approval, replace current policy, or provide legal advice. Source-linked traceability is incomplete. Production or pilot launch is a no-go.
Owner: Marquetta Moore — independent five-day build covering product, UX, architecture, code, deployment, documentation, and evaluation.
Test method: 25-intent deterministic source-level benchmark: 52% pass, 12% partial, 36% fail, 58.6% weighted readiness, 0% unhandled errors, and 100% known-fault capture. Live semantic benchmark has not yet run.
Failure modes: Manual-text-only routing, malformed or unsupported input, mislabeled fallback provenance, prompt-injection exposure, and Shield severity, summary, and rewrite contradictions. Three critical failures remain open.
Monitoring: Re-run the pack after P0 fixes; a 48-row live semantic worksheet is ready for three-run testing. Production telemetry, provider reliability, latency, and campaign outcome monitoring are not yet measured.
Adoption: Public deployed prototype. Recurring users, approved campaigns, production use, and business outcomes are not yet measured.

Public app + repository · 25-intent evaluation · 2026-10-07

Open case study

System 04

Glean Labeling-Rule Monitor

Enterprise source-governance workflow that separates authoritative change from discussion before downstream release.

Baseline · needs review

Strongest proof

Enterprise source governance, authority hierarchy, privacy controls, human approval, and a downstream release path.

Still unmeasured

Repeat-run accuracy, review participation, false-positive rate, and downstream release impact are not yet measured.

User / workflow: CSI and TI guide owners and analysts need viable labeling-rule changes detected, reviewed, and queued for monthly Toolbox updates without changing analyst behavior prematurely.
Sources / actions: Searches approved knowledge pages, the issue tracker, approved team channels, and the training archive; compares a 48-hour window; writes a structured report and routes viable candidates to the monthly preview queue.
Scope limits: AI-generated draft; human review required. Published guidance and verified current instructions control. Peer discussion and MotW are context or training only. Customer, case, message, identifier, and attachment details are omitted.
Owner: Marquetta Moore built and maintains the monitor. Designated subject-matter experts and published-guide owners decide policy. Permanent absorbing owner and review SLA are not yet formalized.
Test method: First baseline compared controlling sources, in-window changes, training signals, owner questions, and privacy treatment; a reviewer checklist checks authority hierarchy and safe generalization. Repeat-run accuracy is not yet measured.
Failure modes: False novelty on the first baseline, peer discussion promoted as policy, stale source coverage, permissions gaps, missed rule changes, source conflict, and sensitive-data leakage.
Monitoring: Daily monitor page with a 48-hour reporting window; monthly preview queue; human approval; month-end release checks; first-of-month changelog rollover. Run reliability and alert SLA are not yet measured.
Adoption: First baseline completed; no analyst action was required. Reviewer participation, approved candidates, missed-change rate, false-positive rate, and downstream release impact are not yet measured.

First baseline · 48-hour reporting window

Public-safe summary only

Portfolio-wide read

The strongest evidence is in the system design

The clearest proof is product ownership and maintenance in the CSI Toolbox, instruction and state design in Launchpad, formal evaluation and failure diagnosis in Aegis, and enterprise source governance in the Glean monitor.

The largest remaining gap is measured adoption and behavior change for real user cohorts.

Evidence is more useful when the gaps stay visible.

Internal source links, production data, customer details, and private screenshots are intentionally omitted. Documentation views are context, not adoption metrics.