Asset Nova: Shield — Interactive Prototype

1 · Live alert

2 · Daily review

3 · Threat & defense

4 · Approve response

5 · What happens next

When Shield detects an urgent threat to the owner's profit, it sends an immediate alert to the contracts team's channel. This one arrived on a Thursday afternoon, 6 days before the contractor's notice deadline expires.

#jubail-claims-defense Thursday 21 Aug 2026, 14:32

Asset Nova: Shield APP

⚠ Incoming claim threat detected — $42.3M exposure

A contractor claim is forming against the owner on Area 7 pipe rack foundations. The contractor hit unexpected rock during excavation and has sent an informal letter reserving their rights. They are building toward a formal claim for $42.3M in extra cost and schedule delay.

Why this is urgent: The contract gives the contractor 21 days to file a formal notice. If they file it correctly, they can claim. If they miss the deadline or file incorrectly, the owner is protected. The deadline is in 6 days.

Shield found 3 weaknesses in the contractor's position:

1. Their informal letter (C-1187) does NOT satisfy the formal notice requirement

2. Amendment 2 shortened the deadline from 28 to 21 days — the contractor may not know

3. The schedule delay they'll claim (54 days) is not yet substantiated in the P6 narrative

Defense brief ready for your review.

Open defense brief

View daily summary

Every morning, Rania also receives a daily email summarizing all threats Shield detected and the owner's defense posture.

From: Asset Nova: Shield (no-reply@shield.assetnova.io)

🛡 Daily Claims Defense Report — 1 active threat, 2 monitored

To: rania.khalid@owner-pmo.sa · Monday 25 Aug 2026, 08:45 · Jubail Gas Processing Expansion

Hi Rania,

Shield reviewed 1,204 new documents this week. Here is the owner's claims defense status:

            ⚠ ACTIVE THREAT — Response needed by 30 Aug

            EV-047: Contractor is building a $42.3M claim for unexpected rock in Area 7. They sent an informal letter but have not filed a formal notice yet. Shield found 3 weaknesses in their position. Defense brief is ready for your approval.

            ✓ Owner's defense position: STRONG

            The contractor's informal letter does not satisfy notice requirements. Amendment 2 shortened the deadline from 28 to 21 days. Their schedule impact is unsubstantiated. If they miss the deadline or file incorrectly, the owner owes nothing.

Threat exposure	$42.3M (contractor claim forming)
Contractor deadline	30 Aug 2026 (6 days) — if missed, claim is barred
Defense weaknesses found	3 procedural flaws in contractor's position
Defense brief	Ready for your review and approval

Also monitoring: compressor vendor delay (potential force majeure argument) and PA-19 unbooked costs (possible claim seed).

THREAT

Contractor building a $42.3M claim against the owner

Excavation crews hit rock 2.3m higher than the ground survey predicted in Area 7. The contractor incurred extra drilling costs and schedule delays. They sent an informal letter reserving their rights and are preparing a formal claim for $42.3M against the owner. Shield assessed the contractor's position and found 3 weaknesses the owner can use to defend.

$42.3Mcontractor's claim amount

6 daysuntil contractor's notice deadline

Strongowner's defense position

96%confidence score

Evidence chain — what the contractor will argue, and where their position is weak

CONTRACT · Clause 4.12EPC Agreement Rev 4 · p.88

"...the Contractor shall give notice to the Owner within twenty-one (21) days of encountering physical conditions... failing which the Contractor shall not be entitled to additional cost or time."

✓ Owner advantage: Amendment 2 shortened this from 28 to 21 days. The contractor may still be working off the original 28-day window.

DAILY SITE REPORT · DR-221004 Aug 2026 · Area 7

"Excavation crews stopped at grid E7-E9. Rock encountered at approx. 2.1m, rig refusal on 3 piles. Awaiting instruction."

Supports contractor: documents the unforeseen condition on site

GROUND SURVEY · Geotechnical ReportFEED package · Appendix C · p.214

"Weathered rock horizon anticipated at 4.5m below grade across Areas 6-8 (boreholes BH-112 to BH-131)."

Supports contractor: clear gap between predicted (4.5m) and actual (2.1m) rock level

LETTER · C-118708 Aug 2026 · Contractor → Owner

"...we advise of unforeseen ground conditions in Area 7 and reserve our rights... revised methodology under assessment."

✓ Owner advantage: this is an informal reservation. It does NOT satisfy the formal notice requirement under Clause 4.12. If the contractor relies on this letter as their notice, their claim is procedurally deficient.

SCHEDULE CHANGE · P6 Update 26Jul → Aug comparison

Activities A7-2210 to A7-2265 delayed by 54 days. Pipe rack path went from 21 days of float to 12 days behind schedule.

✓ Owner advantage: the schedule change was not explained in the narrative. The contractor cannot substantiate the 54-day delay without documenting the cause.

MEETING MINUTES · WPM-3112 Aug 2026 · Weekly progress meeting

"Contractor raised Area 7 ground conditions. Owner PM requested 'formal substantiation'. No event logged in change register as of 22 Aug."

✓ Owner advantage: the owner already asked for formal substantiation. The contractor has not provided it. This strengthens the procedural defense.

Defense response — prepared by the agent, sent only after human approval

Shield drafted a response letter to the contractor that acknowledges the site condition while protecting the owner's position. Every sentence links to its source document.

OWNER'S RESPONSE TO CONTRACTOR RESERVATION — AREA 7 GROUND CONDITIONS REF: JGX-OWN-0047 · DRAFT v1

We acknowledge receipt of your letter C-1187 dated 08 August 2026 C-1187 regarding ground conditions encountered in Area 7.

We note that your letter constitutes an informal reservation of rights and does not satisfy the formal notice requirements set out in Clause 4.12 of the EPC Agreement Clause 4.12 p.88, as amended by Amendment 2 Amd.2 §3.1, which requires formal notice within twenty-one (21) days of the event.

We further note that the schedule impact claimed in P6 Update 26 P6 U26 has not been substantiated in the schedule narrative as required by the contract. The 54-day delay to activities A7-2210 through A7-2265 remains undocumented.

This letter is without prejudice to the Owner's rights under the Agreement. We reserve all positions.

Your approval is required, Rania
The agent cannot send this response. Evidence pack with 6 documents will be auto-attached to your files.

What happens when you approve:

1. Response letter sent to contractor 2. Evidence pack assembled and filed 3. Event logged in the change register 4. Financial and schedule impact analysis triggered 5. Defense brief prepared for the Monday stewardship meeting

After approval — three follow-up tasks run automatically

Once Rania approves, the response is sent and Shield automatically kicks off three follow-up tasks. Each runs as a separate agent on the NovaOS platform. Nothing goes out without a person signing off at the end.

◆ Shield APPROVED

Response approved by Rania Khalid. Letter sent to contractor.

✓ complete · audit #8842

📄 Evidence Compilation AUTO

Pulls together the 6 source documents into one evidence file. Links each document to the audit trail so everything is traceable.

idle

📈 Financial & Supply Chain Impact AUTO

Checks what the owner is about to spend money on and flags anything that should be paused or rescheduled because of the claim.

idle

📅 Schedule Validation AUTO

Checks whether the contractor's claimed 54-day delay actually matches what the schedule data shows. Prepares questions for the owner to challenge it.

idle

👥 PMO Director Review HUMAN GATE

Summary of the owner's defense position for the Monday leadership meeting. Generated live from the record each time, then discarded.

waiting on agents

Output 1: Defense record

Filed

Response letter sent to the contractor. 6 supporting documents bundled together. Event logged in the project register with a full trail of who did what and when.

Output 2: Save money before it's spent

Pause the steel order

The owner was about to order $1.8M of steel for Area 7. Since the foundations are delayed, that order can wait. Piling rig rescheduled 2 weeks out until the new method is confirmed.

Output 3: Stewardship brief

On demand

One-page summary of the owner's defense for the PMO Director. Generated live each time from the record, then discarded. Not a permanent dashboard.

A person owns every decisionRania Khalid (Contracts Lead) approved this response. The PMO Director owns the agent. Every action has a name next to it, is logged, and can be reversed.

The agent stays in its laneIf something is outside the project scope, it gets escalated to a person. Low-confidence items go into a review queue — nothing is silently dropped.

Success = dollars protectedScoreboard: $42.3M claim defended · 0 deadlines missed · 3 threats assessed this month.

▶ 2b · Design exploration Daily Risk Dashboard Under evaluation

An alternative to the email-based daily review (step 2 above). Instead of an alert email, the contracts team opens a full risk dashboard every morning. This is a design exploration we would validate with contracts managers before building.

Active threats

▲ +1 from yesterday

Monitored

▬ No change

$48.7M

Total exposure

▲ +$42.3M (new threat)

Deadlines missed

▬ Clean record

Event	Exposure	Risk	Trend	Days to deadline	Status
EV-047 Area 7 ground conditions Contractor claim forming — unexpected rock	$42.3M	Critical	▲New	6 days	Defense brief ready
EV-039 Compressor vendor delay Contractor may argue force majeure	$4.1M	Elevated	▬Stable	18 days	Monitoring
EV-034 PA-19 unbooked cost trend Costs in payment app not in change register	$2.3M	Elevated	▼-$0.4M	No deadline	Monitoring
EV-028 Concrete batch test failures Resolved — contractor accepted responsibility	$0	Resolved	▼Closed	—	Defended

What changed since yesterday

• New threat: EV-047 detected from site report SR-1187 + correspondence C-1187
• EV-034 exposure reduced: $2.7M → $2.3M after contractor corrected PA-19
• EV-028 closed: Contractor accepted responsibility, no claim filed

Actions required today

⚠ Review EV-047 defense brief — deadline in 6 days
• Acknowledge EV-039 status — no action needed yet
✓ EV-028 closed — no action needed

Defense track record (project to date)

7 threats detected 4 defended successfully 2 being monitored 1 active — action needed $0 deadlines missed

Case study artifacts & the skills that produced them

Every artifact was produced using PM OS, a reusable operating system for product work. Each skill follows a consistent process: gather context, generate a first draft, review from multiple perspectives, iterate.

Product Discovery

Full Discovery Research — 48 sources, 7 sections
Discovery Report (Plain English) — full report without jargon
Validation Quotes by Opportunity — sourced quotes mapped to personas and JTBD

Analysis

Competitive Analysis — SuperHive, Nodes & Links, full landscape
/competitor-analysis
Impact Sizing — driver tree, three scenarios ($30-80M value, $10-12M ARR)
/impact-sizing
Sub-Agent Review Synthesis — 5-reviewer stress test, 12 findings
/ralph-wiggum /prd-review-panel

Product Specs

Product PRD — hypothesis, behavior contract, rollout plan, eval framework
/prd-draft
AI Engineering PRD — DAG pipeline, governance kernel, eval suites
/prd-draft --ai
Strategy & Roadmap — thesis, wedge decision, 3-horizon plan
/write-prod-strategy

Prototypes

Interactive prototype — this demo (Shield Demo tab)
/prototype
Agentic system design — condensed in the System Design tab
/prototype
Eval Harness & Governance Kernel — runnable code with synthetic data

Context

Orbitron Context Synthesis — product lineup, founder philosophy, platform implications

PM OS — the system behind this case study

PM OS is a reusable operating system for product management work. 41 AI-assisted skills covering the full PM lifecycle, from discovery through launch. Here are the ones used for this case study:

/prd-draftProduct PRD with hypothesis, behavior contract, rollout plan

/impact-sizing4-step framework: usage funnel, impact calc, assumptions, takeaways

/competitor-analysisCompetitive landscape mapping with positioning

/ralph-wiggumDevil's advocate review — catches what you don't want to hear

/prd-review-panelMulti-perspective review: engineer, designer, exec, legal, customer

/prototypeInteractive prototypes generated from PRD specs

/write-prod-strategyStrategy thesis, wedge decision, multi-horizon roadmap

Other available skills not used here: /daily-plan /meeting-notes /user-interview /experiment-decision /launch-checklist /feature-metrics /journey-map and 34 more.

The skills compound. Discovery research feeds impact sizing. Impact sizing feeds the PRD. The PRD feeds the prototype. The review panel catches gaps. Each output is the input for the next. This isn't a one-off case study process — it's how I work on every product problem.

📁

View full repository on GitHub

All artifacts, research, PRDs, eval harness, synthetic data, and prototype source code.

Early-stage proposal. This system design will be validated and refined through test-and-learn iterations on real project data, starting with Phase 0 backtests on closed projects. The synthetic scenario demonstrates the pipeline on a contractor building a $42.3M claim against the owner on the Jubail Gas Processing Expansion.

DAG architecture overview

Shield is a directed acyclic graph (DAG) pipeline. Each node has a defined input, output, and quality gate. LLM where language understanding is the problem. Code where being wrong is unrecoverable. Humans approve every external action.

1. Ingest

Parse & index

code

→

2. Detect

Find threats

LLM

→

3. Map contract

Find clauses

LLM

→

4. Deadlines

Date math

code

→

5. Evidence

Build chain

LLM

→

6. Draft

Write response

LLM

→

7. Human gate

Approve

human

LLM Judgment tasks

Threat detection, clause interpretation, evidence linking, response drafting. Pattern matching across unstructured text.

Code Deterministic tasks

Document parsing, deadline computation, citation verification. Rules engines and date math. Auditable and reproducible.

Human Decision tasks

Defense strategy, response approval, claim valuation. The agent surfaces evidence and drafts; the human owns the decision.

Pipeline walkthrough

Steps 1-2: Ingest + Detect

Input

DR-1042: "Piling rig PR-07 refusal at 2.1 m"
GBR 5.3.4: "Rippable material to 6.5 m"

→

Processing

LLM pairs baseline with field observations. Rock at 2.1m vs 6.5m baseline. Change register: no record covers Area 7. Contractor has grounds for a claim.

→

Output

THREAT: EV-047 | confidence: 96%
basis: rock at 2.1m vs 6.5m baseline
in_change_register: false

Steps 3-4: Map contract + Deadlines

Input

GC 4.12(b): "notice within 28 days..."
Amendment 2: replaced with "21 days"

→

Processing

LLM retrieves clauses, applies amendments. Code computes: 04 Aug + 21d = 25 Aug. Owner response: 01 Sep (6 days left).

→

Output

contractor_bar: 21d (not 28 -- caught!)
contractor_deadline: EXPIRED 25 Aug
owner_response: 01 Sep (6 days)

Steps 5-6: Evidence + Draft

Input

C-1187: "We reserve our rights..."
No particulars, no records attached.

→

Processing

LLM scores documents for/against owner. C-1187 is bare reservation, doesn't satisfy notice. Code verifies every citation. LLM drafts response.

→

Output

EP-047: 6 items verified
C-1187 insufficient as notice (DEFENSE)
Amd 2: 21d bar (may be expired)
DRAFT: JGX-OWN-0047 (every sentence cited)

Quality gates

Every node has a testable quality gate. The eval suite runs on a golden set of confirmed events.

LLM Threat detection recall

Target: 100% recall; maximize precision progressively

Missing a real threat is unacceptable. False alarms are tolerable early and reduced over time through the self-improving loop.

Test: 50 site reports from closed project. Expert labeled 8 real threats, 42 non-events. Agent must flag all 8. Precision starts at whatever level catches them all, then improves as dispositions accumulate.

Code Citation accuracy

Target: 99.5%

Every citation must resolve to an actual document, section, and page.

Test: Agent cites "Clause 4.12, p.88". Verifier opens the contract, confirms clause exists at that page.

Code Deadline accuracy

Target: 100%

Date math is code, never LLM. Zero tolerance for errors.

Test: Event date = 4 Aug, contract says "21 days", Amendment overrides to "14 days for Areas 6-8". Expected: 18 Aug.

LLM Hallucination in drafts

Target: 0 in approved outputs

Every sentence must trace to a source. Citation verifier + human review = double barrier.

Test: Agent writes "as stated in letter C-1205". Verifier checks: does C-1205 exist? If not, draft is blocked.

LLM Defense position accuracy

Target: 80% agreement

"Owner advantage" vs "supports contractor" tagging must match expert assessment.

Test: 6 documents from a resolved dispute. Expert tagged 3/3 split. Agent must agree on at least 5 of 6.

Human Draft approval rate

Target: 70% approved

If the team rewrites most drafts, the agent isn't learning their style. Edit diffs feed the loop.

Test: After 4 weeks, of the last 20 drafts, at least 14 approved as-is or with minor edits.

Self-improving loop

Gets smarter on your project, without your data leaving.

Agent flags threat

Detection + evidence + draft

→

Human reviews

Approve, edit, or dismiss

→

Decision logged

Labeled example in audit trail

→

Agent retunes

Thresholds, templates, golden set

Dismissals reduce noise

When the team dismisses a false alarm, Shield raises its threshold for that pattern. Fewer irrelevant flags over time.

Edits sharpen drafts

When the team rewrites a response before sending, the edit diff is captured. Draft quality improves to match how the team writes.

Approvals confirm what matters

Confirmed events grow the golden eval set. No update ships unless it passes on all confirmed events. The test suite gets harder over time.

Full sovereignty: all retraining happens inside the customer's environment (in-region VPC or on-prem). Dispositions never leave. Your project data never improves someone else's agent. No shared model, no data export, no cross-tenant learning.

End-to-end recap

Step	Input	Output	Who	Eval gate
1. Ingest	6 raw files, 5 systems	Indexed record with provenance	Code	Parse coverage 100%
2. Detect	Baseline + field reports	EV-047 threat, 96% confidence	LLM	Recall 100%
3. Map contract	GC 4.12 + Amendment 2	21-day bar (not 28), clause refs	LLM	Expert agreement ≥80%
4. Deadlines	Date anchors	Contractor bar expired, 6 days left	Code	Accuracy = 100%
5. Evidence	Chronology + letters	EP-047: 6 docs, 3 defense advantages	LLM + Code	Citation accuracy ≥99.5%
6. Draft	EP-047 + defense positions	Owner response letter, every sentence cited	LLM	0 hallucinations, 70% approval
7. Human gate	Draft + evidence	Approved response, audit trail, 5 actions	Human	Named approver required

▶ Agentic Design Exploration In research & testing

Design A: Cascade Pipeline

Wide funnel, progressive filtering. Optimized for recall and speed.

Document InputAconex, P6, SAP

continuous

→

Fast TriageSmall model + rules

100% recall target

→

Deep AnalysisFrontier LLM on flagged items only

+precision

→

Deadline EngineDeterministic date math

never LLM

Citation VerifierDoc ID + quote match

99.5% gate

→

Evidence AssemblyRAG from document store

grounded

→

Defense DraftDraft + QA retry loop

pass@k

→

Governance KernelScope, allowlist, audit, sovereignty

runtime

→

Human GateNamed approver signs off

final

LLM node Deterministic (code) Governance Human

Recommended for v1. Fixed DAG is fully testable per node. 100% recall is structural (triage is deliberately permissive). Precision improves at each stage. Fast: frontier LLM only runs on ~25-30% of document volume. Maps to existing AI engineering PRD architecture.

Design B: Orchestrator + Specialist Agents

A routing agent delegates to domain-specific sub-agents. Optimized for precision and composability.

Document InputAconex, P6, SAP

→

Orchestrator AgentReasons about which agents to invoke and in what order

routing risk

→

Threat DetectorTuned for recall

Clause ExtractorContract parsing

Evidence LinkerCross-doc reasoning

Deadline ComputerDeterministic

Defense DrafterResponse generation

→

Cross-Validation"If I were the contractor, how would I attack this defense?"

adversarial

→

Governance KernelScope, allowlist, audit, sovereignty

→

Human GateNamed approver

Orchestrator LLM node Deterministic

Better long-term architecture. Each agent is independently evaluable and reusable in NovaStudio. Orchestrator enables richer reasoning (e.g., "this is a delay claim, run schedule analysis first"). Cross-validation agent strengthens defenses. But: orchestrator routing is itself an LLM decision, making 100% recall harder to guarantee at launch. Slower and more complex to eval end-to-end.

Shared: Self-Improving Eval Loop

Applies to both designs. How the agent gets better on each project without re-engineering.

Human decisionsApprove, edit, or dismiss

→

Labeled examplesProject-specific golden set

→

Eval engineBenchmark agent against updated set

→

Prompt optimizerReflect on disagreements, synthesize updates

→

Config storeUpdate thresholds, few-shot examples

↻

Prompt tuning, not fine-tuning. Learning stays in-region. Threshold changes are automatic; prompt changes require human approval.

Design Tradeoff: Success Metrics and Guardrails

How each design performs against objective success metrics and guardrail constraints.

Dimension	Metric	Gate	Design A: Cascade	Design B: Orchestrator
Objective success metrics
Recall	Threats detected / total real threats	100% hard	Structural. Fast triage is deliberately permissive. Recall is a property of the design, not a tuning target.	Dependent on orchestrator routing. If orchestrator skips threat detector on a document, threat is missed. Requires routing eval coverage.
Precision	Real threats / total alerts raised	>70% soft, improving	Progressive. ~25-30% after triage, ~70-80% after deep analysis. Each stage has its own precision eval. Improves with feedback loop.	Higher at launch. Specialist agents are individually tuned. Cross-validation catches weak cases. But harder to diagnose precision drops (which agent failed?).
Draft quality	Human edit distance on defense responses	Decreasing trend	QA retry loop (pass@k). Single drafter with iterative refinement. Quality depends on one agent's capability.	Adversarial cross-validation. Separate agent stress-tests the defense before human review. Stronger output.
Speed	Ingestion to alert	<4h (triage <5min)	Fast. Frontier LLM runs on ~25-30% of volume. Triage alerts can fire in minutes for deadline-critical items.	Slower. Multiple agent invocations with individual latency. Orchestrator reasoning adds overhead. Less predictable total time.
Adoption	Daily review completion rate	>90%	Simpler output. Contracts team sees a filtered list. Predictable behavior builds trust faster.	Richer output. Cross-validation provides "contractor perspective" which adds value but also complexity to the review.
Guardrail metrics (must not violate)
Citation accuracy	Citations that resolve and quote-match source	99.5% hard	Deterministic verifier in the DAG. Every citation checked by code before surfacing. Failed citations downgraded, never shown.	Same verifier, different position. Runs after evidence linker agent. Same guarantee, but verifier must handle outputs from multiple agent formats.
Deadline accuracy	Computed dates vs hand-verified	100% hard	Code-only. Rules engine with PROVISIONAL flag for missing amendments. LLM never touches date math.	Same engine. Dedicated deadline computer agent is deterministic. Identical guarantee.
Hallucination defense	Fabricated citations caught	100% hard	Structural. Verifier is a mandatory DAG node. No path bypasses it.	Orchestrator-dependent. Orchestrator must always route through verifier. A routing bug could skip it. Needs explicit routing eval.
Sovereignty	Data residency violations	0 hard	Enforced at governance kernel. Every document tagged. Out-of-region data refused at the gate.	Same kernel. Governance wraps the orchestrator. Identical enforcement.
Scope containment	Out-of-scope actions attempted	0 hard	Action allowlist. Capabilities enumerated. "Send" not in the list.	Broader surface. More agents = more capability interfaces to constrain. Each agent needs its own allowlist.
Auditability	Every decision traceable end-to-end	Full chain	Linear DAG. Append-only, hash-chained log. Each node produces one auditable output.	Branching traces. Orchestrator decisions create variable-length traces. Audit log is richer but harder to read.

Recommendation

Design A (Cascade) for v1. The 100% recall constraint makes a fixed, testable pipeline safer than LLM-routed orchestration at launch. Every guardrail is structural, not behavioral.

Evolve toward Design B once 6+ months of labeled data from human decisions provides the eval coverage to trust orchestrator routing. Design B is more aligned with NovaOS's composable agent model and scales better when Shield expands to Horizons 2-3 (Design Gate, Budget Red Team).

These designs are informed by agent orchestration patterns from production AI systems, adapted for Shield's domain constraints (100% recall, sovereign deployment, human-in-the-loop). Both designs share the same governance kernel, deterministic verification layer, and self-improving eval loop.

Threat	Contractor claim for $42.3M (extra cost + 54-day schedule delay)
Owner's position	Strong — 3 procedural weaknesses in contractor's claim identified
Response status	JGX-OWN-0047 sent by Rania Khalid · contractor's notice deadline in 6 days
Evidence pack	EP-047 · 6 documents · audit-mapped
Money saved	Paused the $1.8M steel order (not needed yet) · piling rig rescheduled +2 weeks
Schedule challenge	9 questions drafted — contractor's 54-day delay is unsubstantiated in the P6 narrative

↧ What the agent reads (inputs)

⚙ What the agent does (processing)

⚠ Quality checks (before any output)

↦ Threats detected this week

Contractor building a $42.3M claim against the owner

OWNER'S RESPONSE TO CONTRACTOR RESERVATION — AREA 7 GROUND CONDITIONS REF: JGX-OWN-0047 · DRAFT v1

◆ Shield APPROVED

📄 Evidence Compilation AUTO

📈 Financial & Supply Chain Impact AUTO

📅 Schedule Validation AUTO

👥 PMO Director Review HUMAN GATE

Output 1: Defense record

Output 2: Save money before it's spent

Output 3: Stewardship brief

What changed since yesterday

Actions required today

Defense track record (project to date)

Case study artifacts & the skills that produced them

Product Discovery

Analysis

Product Specs

Prototypes

Context

PM OS — the system behind this case study

DAG architecture overview

1. Ingest

2. Detect

3. Map contract

4. Deadlines

5. Evidence

6. Draft

7. Human gate

LLM Judgment tasks

Code Deterministic tasks

Human Decision tasks

Pipeline walkthrough

Quality gates

LLM Threat detection recall

Code Citation accuracy

Code Deadline accuracy

LLM Hallucination in drafts

LLM Defense position accuracy

Human Draft approval rate

Self-improving loop

Agent flags threat

Human reviews

Decision logged

Agent retunes

Dismissals reduce noise

Edits sharpen drafts

Approvals confirm what matters

End-to-end recap

Design A: Cascade Pipeline

Design B: Orchestrator + Specialist Agents

Shared: Self-Improving Eval Loop

Design Tradeoff: Success Metrics and Guardrails

Recommendation

Defense brief — EV-047 · Contractor claim, Area 7 ground conditions