# Proof room: composite workflow traces

These are **composite traces**, not client artifacts. They are designed to show exactly how the proof room should work once real field evidence is collected.

Each trace is structured to answer one skeptical question:
- what are the actual inputs?
- where does the model help first?
- what still requires human signoff?
- what can break?

## Trace 1 — Treasury / lender readiness

### Workflow
Borrowing-base package support, covenant monitoring, amendment prep, and lender Q&A.

### Typical evidence bundle
- latest reserve report
- production and cash-flow exports
- hedge schedule
- debt documents and covenant definitions
- prior lender questions
- board and management updates
- historical redetermination memos

### Current manual steps
1. gather latest source documents
2. reconcile reserves, production, and price assumptions
3. update covenant workbook and narrative
4. prepare package summary and exhibits
5. route to treasury, finance, legal, and management review
6. answer lender questions and track deltas

### What AI changes first
- source comparison and reconciliation
- variance tables
- covenant-definition extraction
- Q&A draft generation with citations
- amendment issue spotting
- evidence pack assembly

### What remains distinctly human
- negotiation posture
- lender relationship management
- representation of downside scenarios
- signoff on what is sent externally
- final legal and treasury judgment

### What can break
- stale source files
- conflicting assumptions across reserve, hedge, and cash models
- hidden covenant definitions
- misread legal qualifiers
- optimistic narrative unsupported by the numbers

### Success metrics
- package cycle time
- questions answered inside 24 hours
- number of manual reconciliations
- error escapes
- analyst hours per cycle

## Trace 2 — Non-op / minerals / title

### Workflow
AFE elections, JOA interpretation, JIB review, title / curative issues, division-order exceptions, owner-relations cases.

### Typical evidence bundle
- JOAs and amendments
- lease and deed chains
- title opinions and curative notes
- division orders
- AFEs and election notices
- JIB statements and dispute notes
- owner correspondence

### What AI changes first
- clause extraction across JOAs and leases
- document comparison across title chain changes
- exception queue triage
- owner-response draft support
- curative issue clustering
- election deadline monitoring

### What remains distinctly human
- legal interpretation on edge cases
- negotiation with counterparties and owners
- escalation on title risk
- final payment and title positions

### Failure modes
- poor scan quality
- ambiguous chain-of-title defects
- inconsistent owner naming and entity history
- local practice exceptions not visible in the docs

### Success metrics
- suspense resolution time
- title issue queue aging
- JIB dispute cycle time
- percentage of routine owner cases auto-triaged

## Trace 3 — Utility planning / rate recovery

### Workflow
IRP support, rate-case evidence assembly, data-request response, testimony package prep.

### Typical evidence bundle
- forecast workbooks
- generation and load scenarios
- depreciation schedules
- historical filings and testimony
- commission discovery requests
- cost-of-service models
- planning and procurement memos

### What AI changes first
- exhibit assembly
- testimony draft support with citations
- discovery request routing and drafting
- scenario comparison summaries
- consistency checks across narrative and workpapers

### What remains distinctly human
- policy judgment
- regulatory strategy
- witness preparation
- external advocacy
- final management and legal signoff

### Failure modes
- version drift between testimony and workpapers
- unsupported citations
- scenario assumptions presented without policy context
- discovery responses that over-answer or under-answer

### Success metrics
- turnaround on data requests
- revision count per exhibit
- witness-prep load
- evidence consistency escapes

## Trace 4 — LNG commercial and shipping

### Workflow
SPA administration, cargo scheduling, chartering support, laycan changes, settlement support.

### Typical evidence bundle
- SPA contracts
- laycan and cargo schedules
- vessel / charter documents
- nomination history
- terminal and regas constraints
- invoice / settlement records
- counterparty communications

### What AI changes first
- SPA clause extraction
- schedule and laycan exception monitoring
- counterparty issue logs
- invoice support and discrepancy summaries
- cargo-brief drafting with citations to contract and schedule sources

### What remains distinctly human
- commercial discretion
- counterparty management
- vessel and destination decisions
- exposure-taking and escalation

### Failure modes
- stale schedule snapshots
- contract amendments not reflected in the system
- missing vessel constraints
- subtle commercial options misread from the SPA

### Success metrics
- time to resolve cargo exceptions
- settlement discrepancy cycle time
- schedule change response speed
- disputes avoided

## Trace 5 — EPC / project controls

### Workflow
RFI support, submittal routing, change-order review, schedule and cost-package updates.

### Typical evidence bundle
- RFIs
- submittals
- contracts and scope books
- schedule snapshots
- cost reports
- change-order logs
- vendor correspondence

### What AI changes first
- clause extraction from scope and contract packages
- delta summaries across revisions
- issue routing
- change-order evidence packs
- weekly reporting support

### What remains distinctly human
- commercial negotiation
- site reality checks
- claims posture
- final commercial decisions

### Failure modes
- mismatched scope versions
- weak source provenance
- site issues not reflected in document systems
- claims inflation from shallow document reads

### Success metrics
- change-order cycle time
- RFI aging
- reporting burden
- dispute-preparation hours

## What the real proof room should contain next

The next fieldwork phase should replace these composites with:
- redacted artifact families
- before/after traces
- model-eval results
- error examples
- explicit signoff boundaries