SYSTEMS INDEX / APPLIED AI
What the model does.
What the system must do.
Four workflows, inspected at their boundaries.
Models help interpret, draft, and collect information. Software preserves state and constrains action. People retain judgment where the evidence is incomplete or authority matters.
FIELD ATLAS / FIVE DIFFERENT SYSTEMS
Select a case. Inspect the boundary.
A draft is a proposal. Review gives it context.
Who judges technical correctness?
- 01INFORMATION / ARTIFACT
Technical sources
Documents and engineering feedback
- 02MODEL ASSISTANCE
Assist the draft
AI-assisted writing and consistency review
- 03SOFTWARE LOGIC
Expose the change
Markdown, Git branches and reviewable diffs
- 04HUMAN JUDGMENT
Engineer review
People retain technical authority
INSPECT THE FAILURE BOUNDARYA plausible draft contains a stale reference or incompatible assumption.
Review must examine the changed claim and its source. Version history alone cannot establish correctness.
An abstraction of the documentation workflow. No aircraft design, certification authority, or proprietary document is depicted.
Keep the context attached to the conversation.
What disappears when a conversation becomes a summary?
- 01HUMAN JUDGMENT
Listen in the field
Factory staff and purchasing explain the operation
- 02INFORMATION / ARTIFACT
Capture context
Recorded conversations and field observations
- 03MODEL ASSISTANCE
Organize knowledge
AI-assisted customer knowledge and presentations
- 04HUMAN JUDGMENT
Interpret constraints
The internal team considers deployment questions
INSPECT THE FAILURE BOUNDARYA summary erases who said something or treats an observation as a requirement.
The account supports a knowledge workflow, not a complete or independently verified model of a customer’s operation.
Author-reported workflow; exact implementation is not publicly evidenced. No customer identities or factory details are reproduced.
Approval is not delivery. Delivery is not acknowledgement.
What would actually count as a completed handoff?
- 01MODEL ASSISTANCE
Collect intake
Voice-model intake produces a durable call record
- 02SOFTWARE LOGIC
Bind the payload
Prepare stored content and check its hash
- 03HUMAN JUDGMENT
Approve that version
Reviewer approves the exact payload
- 04SOFTWARE LOGIC
Attempt the handoff
Local desk paging and alerts are simulated
SIMULATED / MODELED - 05SOFTWARE LOGIC
Check acknowledgement
Track acknowledgement by a named person
INSPECT THE FAILURE BOUNDARYThe payload changes, or nobody acknowledges the attempted handoff.
Stale or corrupted payloads are rejected. Silence can exhaust escalation without being labeled completion. These are inspected prototype behaviors; the suite was not rerun for this site.
This drawing covers the voice prototype. The separate accounting adapter is read-only; it does not create invoices. No live dispatch or autonomous business outcome is implied.
A signature needs a task, a delivery, and a trust boundary.
What is the checker actually entitled to certify?
- 01SOFTWARE LOGIC
Post / claim
An explicit work lifecycle identifies the task
- 02INFORMATION / ARTIFACT
Submit delivery
Acceptance evidence refers to a delivery hash
- 03SOFTWARE LOGIC
Check the receipt
Signed checker evidence and configured trust roots
- 04SOFTWARE LOGIC
Accept / issue note
Record the accepted transition in the event log
- 05INFORMATION / ARTIFACT
Mock settlement
Default local chain file; no real payment
SIMULATED / MODELED
INSPECT THE FAILURE BOUNDARYA receipt is validly signed but concerns different work, or its checker has limited coverage.
The meaning of acceptance depends on the task/delivery binding and the configured checker. A signature is not proof of arbitrary correctness.
A simplified normal path from repository documentation. Review-lapse and dispute branches are omitted. The implementation and sandbox are not independently audited here.
A narrow experiment. A deliberately narrow conclusion.
Does the result survive a wider accounting boundary?
- 01INFORMATION / ARTIFACT
Select a model slice
First 64 outputs of the first Qwen query projection
- 02SOFTWARE LOGIC
Map the arithmetic
56 tiles, each with 64 inputs and 64 outputs
- 03SOFTWARE LOGIC
Model the readout
Clipping and finite readout precision
SIMULATED / MODELED - 04INFORMATION / ARTIFACT
Compare errors
Saved projection errors against dequantized weights
- 05HUMAN JUDGMENT
Screen the energy
Separate analytical bounds and circuit estimates
INSPECT THE FAILURE BOUNDARYA smaller component cost is mistaken for a complete-system efficiency gain.
Drive, routing, idle power, wake-up and reduction still need accounting. Projection error does not measure language-model quality, throughput, or silicon energy.
The drawing joins related research questions, not one measured hardware pipeline. Simulation results and energy estimates have separate scopes. Original experiments were not rerun for this site.
◇ Human judgmentSolid outline: described stageDashed outline: simulated or modeled stage
Engineering information
USEDChanging technical sources need to become reviewable documents without losing their assumptions.
Read the full context →- Inputs
- Repository documents, source material, engineering feedback.
- Model role
- Assist drafting, analysis, and consistency review.
- Tools / APIs
- Claude Code, Markdown, Git.
- Deterministic layer
- Versioned changes, scoped diffs, and review history.
- Human boundary
- Engineers judge technical correctness and retain authority.
- Output
- Revised technical material in the documentation workflow.
- Failure modes
- Plausible prose can preserve a false premise or introduce a stale reference.
- Observed outcome
- Used in professional documentation work; no measured efficiency claim.
Customer context
USEDKnowledge from factory conversations needs to reach the people shaping the deployment.
Read the full context →- Inputs
- Recorded conversations and field observations.
- Model role
- Help organize customer knowledge and prepare presentations.
- Tools / APIs
- AI recording and knowledge workflows; NotebookLM.
- Deterministic layer
- Customer-specific organization of context; exact implementation not publicly evidenced.
- Human boundary
- People in the operation explain constraints; the internal team interprets deployment implications.
- Output
- Customer knowledge bases and presentation material.
- Failure modes
- Summaries can erase who said something or confuse an observation with a requirement.
- Observed outcome
- First-hand account of real workflow use; no public deployment or quantified outcome evidence.
Read-only operations assistant
BUILTQuestions about work and money span field-service and accounting systems.
Read the full context →- Inputs
- Jobber and QuickBooks records.
- Model role
- Provide a conversational interface to operational information.
- Tools / APIs
- Typed adapters and provider APIs.
- Deterministic layer
- Accounting requests restricted to GET; pagination completeness tracked explicitly.
- Human boundary
- Operators interpret findings and retain responsibility for business action.
- Output
- Read-only operational information with coverage boundaries.
- Failure modes
- A partial fetch can look complete; a broad OAuth permission can exceed the intended action boundary.
- Observed outcome
- Adapter implementation inspected. No claim that the assistant autonomously created invoices or maintained a production invariant.
Voice intake & handoff
PROTOTYPEA call must become a usable record and an accountable next step.
Read the full context →- Inputs
- Call transcript, collected fields, and configured business rules.
- Model role
- Conduct intake and extract structured details.
- Tools / APIs
- Voice-provider APIs, typed intake, local operator desk.
- Deterministic layer
- SQLite records, routing rules, payload hashes, approval checks, and acknowledgement states.
- Human boundary
- A reviewer approves the exact payload; a named person acknowledges the handoff.
- Output
- Durable call record, prepared drafts, and explicit handoff status.
- Failure modes
- Missing details, contradictory dates, stale approvals, or an unacknowledged delivery attempt.
- Observed outcome
- Prototype with provider demos. Desk paging and alerts are simulated; follow-up drafts do not send themselves.