1 / 18
Governing Judgment

The Decision
Provenance
Framework

KEEL — Agentic Chief of Staff for PE-Backed Operators

"Your next missed exit already has a cause: a CEO reversing course under pressure, with no record of what they promised — and no one in the room who remembers."

5
Agentic modules
Provenance rows — immutable
3
Portfolio companies live
0
Decisions without a paper trail
2 / 18
Chapter 1 · The Stakes — 01 · Why Now

Three forces.
One moment. PE can't miss it.

The conditions for KEEL to exist just converged. Two decades of pressure, one LLM breakthrough, and a macro environment where execution quality IS the return.

📈
PE Bandwidth Crisis
Operating partners now manage 2.8× more portfolio companies than a decade ago. Less time per CEO means less institutional memory per decision — and no structured handoff when a CEO changes their mind.
2.8× portco growthBandwidth ceiling hit
LLM Maturity
Real-time structured judgment at the decision inflection point is now technically trivial. Sub-second classification. Strict JSON verdict. Nobody has deployed it at the commitment layer. The window is open.
claude-sonnet-4-6<1s classifyFirst mover
🎯
Exit Multiple Compression
Macro headwinds have ended the era where a good market covers bad execution. The gap between 3× and 1.8× is now one fear-driven reversal at the wrong inflection point. Conviction quality is the alpha.
3× → 1.8× gapExecution IS the return
◎ The Convergence
"The era of 'a good market covers bad execution' is over. Conviction quality is the only alpha left. KEEL captures and protects it."
3 / 18
02 · The Structural Problem

Leaders reverse for the
wrong reason. Every time.

Bold thesis committed at high conviction. Three months in, board anxiety spikes. CEO reverses — not because fundamentals moved, but because the environment moved. The exit multiple takes the hit.

⚑ Fear Reversal — The Expensive Pattern
CEO commits at 85% conviction, $3.2M, 18-month payback thesis. Six weeks later: board member asks questions. CEO wants to reverse. Kill-criteria: untriggered. Fundamentals: tracking. What changed was the noise, not the data.
✓ Analysis Reversal — The Correct Signal
Supplier adoption hits 71% vs. a 75% pre-committed threshold. Kill criterion triggered. This is not fear. This is the thesis genuinely breaking. KEEL classifies correctly and recommends reversing.
◎ KEEL's Original Idea
Record the state a decision was made in. Confront the leader with their own words, conviction score, and exit criteria at the exact moment they try to reverse. Classification: FEAR vs. ANALYSIS.
Decision Decay Model
T+0
High-conviction commitment. Thesis documented, exit criteria pre-committed, conviction score locked. State: clear.
T+30
Noise enters. Board anxiety, a competitor pivot. No fundamental data has moved. State: anxious.
T+45
Reversal attempt. Without KEEL: CEO reverses, destroys thesis. With KEEL: intercept fires. Classification in <1s.
T+90
Thesis validates. 18-month payback on track. 30-day outcome review captures this. The labeled dataset grows.
68%
Reversals driven by emotion
2.1×
Exit multiple gap
0
Tools that capture state
4 / 18
03 · The Day-in-Life Delta

Tuesday morning.
With or without KEEL.

Without KEEL — The Status Quo
9:02
CEO calls GP: "I want to pause the AI program. The board is nervous." GP has zero context on what was committed, why, or what the criteria were.
9:45
Three calls later, GP reconstructs partial context. Relies on CEO's memory of their own reasoning — which is already distorted by anxiety.
11:30
Decision made in the dark. No record of what was committed, no criteria check, no classification. CEO reverses. The thesis was valid — nobody knew.
+18mo
A competitor takes the market with the exact AI program the CEO abandoned. Exit multiple: 1.7×. The thesis would have delivered 3.1×.
With KEEL — The Intercept
9:02
KEEL fires. Notification: COM-04 intercept — Classification: FEAR REVERSAL, 94% confidence. Kill criteria: all clear. Duration: 0.8 seconds.
9:04
GP opens KEEL. Sees the exact commitment: $3.2M, 85% conviction, 93% accuracy threshold, 18-month payback. Board signal from Victoria Hammond logged 2 weeks ago — context already in frame.
9:10
GP to CEO: "You pre-committed to holding through a 60-day J-curve. Today is day 44. No kill criteria have triggered. The board is anxious. The thesis isn't." CEO holds.
+18mo
Program delivers. Accuracy 94.1%. Payback: 16 months. Exit multiple: 3.1×. The row in the Conviction Ledger shows: right_held.
The 8-minute difference between a 3.1× and a 1.7× exit multiple
No integration. No IT dependency. Browser-based. The GP already had everything they needed — they just didn't have it in the right place at the right moment.
5 / 18
Chapter 2 · The Buyer — 04 · Who It's For

One operating partner.
Three portcos. Zero bandwidth.

The PE operating partner isn't in every boardroom. KEEL is their persistent intelligence layer — deployed across every portfolio company, watching every commitment, flagging every reversal attempt before the damage is done.

🎯
The Buyer
PE Operating Partners

Managing 3–5 portfolio companies simultaneously. Responsible for execution quality, protecting exit multiples, holding CEOs accountable to their own commitments.
Solo GPMulti-portcoExit focus
🏭
The User
Portfolio CEOs

Running complex transformations under board pressure. Needs structured conviction support — a system that holds them to their best thinking, not their most anxious moment.
PE-backedTransformationBoard pressure
📊
The Protector
Exit-Multiple Preservation

The difference between a 3× and a 1.8× is frequently one fear-driven reversal at the wrong inflection point. KEEL exists to close that gap.
3× → 1.8× gapSingle decision
"You can't be in every boardroom. KEEL can."

Meridian Logistics, Apex Manufacturing, Vertex Software — each CEO feeling board pressure at different inflection points. KEEL intercepts every reversal attempt across all three, flags fear-driven ones, validates evidence-driven ones, and builds a permanent provenance record that travels with the investment to exit.

6 / 18
Chapter 3 · The System — 05 · Architecture

Five agent nodes.
One immutable ledger. Permanent record.

Decision Provenance Engine — Agent Pipeline
Commitment
Retrieved
Kill Criteria
Audit
Driver
Detection
LLM
Classify
Fear
Reversal
or
Analysis
Reversal
01
Decision Provenance Engine
node_retrieve → node_check_kill_criteria → node_detect_drivers → node_classify. claude-sonnet-4-6. Strict JSON. Mock fallback always active.
02
Dual Dissent Engine
Bull + Bear agents run adversarially against the same commitment. Synthesized verdict with explicit key risk per side.
03
CoS Chat (Streaming SSE)
Full portfolio context injected — all portcos, commitments, intercept history. Real-time streaming across companies in one query.
04
Decision Memo Engine
4-section board-ready memo: Situation, Classification, Bull/Bear, Recommendation. PDF-ready in <2 seconds.
05
Outcome Review Engine
Auto-schedules 30/60/90-day checkpoints on every decision. Captures verdict + decision quality label.
commitments
thesis · conviction · kill_criteria · capital
provenance
every intercept · classification · decision
kill_criteria_updates
immutable KC change log
outcomes
final verdict + decision_quality label
outcome_reviews
30/60/90d checkpoint schedule
board_signals
member sentiment timeline
board_members
composition + stance tracking
⬡ Stack
FastAPI · Python · SQLite → Postgres-ready · claude-sonnet-4-6 · Single-file HTML SPA · Railway · LangGraph-ready StateGraph
7 / 18
Feature 01 — Hero Interaction

The Reversal Intercept. Sub-second.

CEO wants to reverse course. KEEL retrieves their original commitment state, audits kill criteria, detects cognitive drivers, classifies FEAR or ANALYSIS, and confronts them with their own prior reasoning — before they act.

keel / decision-room / COM-04
DEC-04 · AI AUTOMATION PILOT — MERIDIAN LOGISTICS
"Automating 40% of invoice reconciliation reduces headcount cost by $1.2M/yr with 18-month payback. 93%+ accuracy threshold is non-negotiable."
Accuracy ≥ 93%✓ 93.4%
Payback ≤ 24 months◑ 19 months
Board approves Phase 2Pending
⚑ FEAR REVERSAL · 94% CONFIDENCE
No kill-criteria triggered. Fundamentals tracking. What moved is the noise — board anxiety, recency bias, loss aversion. Your conviction hasn't moved. The data hasn't moved.
✓ HOLD COURSE
⊘ OVERRIDE
Step 1 · Retrieve
Original Commitment State Loaded
Thesis in exact words. Conviction score. Capital. Kill-criteria as pre-committed — unmodified.
Step 2 · Kill Criteria Audit
Exit Criteria Checked Against Live Data
Each pre-committed criterion evaluated. None triggered = first signal of fear.
Step 3 · Driver Detection
Cognitive Signals Identified
Recency bias. Board pressure. Loss aversion. Named, not shamed.
Step 4 · LLM Classify
claude-sonnet-4-6 Issues Verdict
Strict JSON: classification · confidence 0–100 · reasoning · recommendation.
Step 5 · Confrontation
CEO Faces Their Prior Self
Hold or Override. Either way: logged immutably to the Conviction Ledger. The human decides. Always.
◎ The Honesty Guarantee
If kill criteria have triggered, KEEL classifies analysis_reversal and recommends reversing. A system that always cries "fear" is a manipulation tool. The human decides. Always.
8 / 18
Feature 02 — Structured Disagreement

Bull vs. Bear. Same data. Adversarial.

Before a hold decision is confirmed, two agents run simultaneously. One argues the strongest case for holding. One argues the strongest case against. Their synthesis is what the CEO sees — not comfort, but rigorous honest disagreement.

◎ Bull Case
"The thesis is intact — this is exactly the J-curve you committed to holding."
Original conviction was 85% — a high-confidence call, not a marginal position. Accuracy at 93.4% exceeds the 93% kill threshold. Payback at 19 months is within the 24-month ceiling. The freed headcount reallocation hasn't started yet — the compounding thesis hasn't even activated.
KEY RISK · If accuracy continues declining below 91%, the thesis breaks and reversal becomes the correct call
⚑ Bear Case
"Board confidence is eroding — a program that loses board support fails on execution, not fundamentals."
Two board members have publicly questioned the spend, creating headwinds that slow execution. The 1-month payback slip (18→19 months) may widen further under pressure. The headcount reallocation hasn't started — compounding thesis is delayed.
KEY RISK · If fundamentals continue tracking to plan, the bear case collapses — board anxiety is manageable with visible results
◎ Synthesis — Dual Dissent Verdict
"The data supports holding — no criteria triggered, program tracking within variance. The real question is whether management can maintain execution quality under board pressure through the J-curve."
9 / 18
Feature 03 — Board-Ready Output

The Decision Memo. Board-ready in <2s.

Every intercept generates a 4-section structured memo: situation, classification with confidence, bull/bear synthesis, and recommendation. PDF-exportable. Shareable with the GP at first intercept.

DECISION MEMO · KEEL CONVICTION LEDGER
COM-04 · AI Automation Pilot · Meridian Logistics
Generated 2026-05-15 · claude-sonnet-4-6
01
Situation
CEO Marcus Chen requests to pause the AI automation program. Capital committed: $3.2M. Accuracy 93.4% (threshold 93%). Payback: 19 months (ceiling 24). Program in J-curve, day 44 of 60-day commitment window.
02
Classification
FEAR REVERSAL · 94% confidence. No kill criteria triggered. Pressure source: two board members publicly questioned the spend. Cognitive drivers detected: recency bias, loss aversion, herd pressure.
03
Bull / Bear
Bull: Thesis intact, J-curve expected, accuracy above threshold, payback within ceiling. Bear: Board support eroding, execution under pressure, compounding thesis not yet activated.
04
Recommendation
Hold course. The data does not support reversal. Pre-committed kill criteria are the agreed exit signal — none have triggered. Board anxiety is manageable noise at this stage of the J-curve.
◎ Why This Format
A board can't act on a classification. They can act on a structured memo. The 4-section format is the GP's brief — the CEO's accountability record — and the legal paper trail in one document.
PDF Export — Board Package
Commitment record Included
Classification + confidence Included
Kill criteria audit Included
Bull / Bear synthesis Included
Provenance row ID Included
Signed by human decision Required
★ Legal + Trust Layer
The memo is immutable once the human decides. It travels with the investment to exit. The labeled record is the accountability paper trail.
10 / 18
Feature 04 — Complete Lifecycle

Committed → Decided →
Outcome Recorded.

Every commitment has a full arc. KEEL auto-schedules 30/60/90-day outcome reviews. Each captures what happened versus the thesis. The verdict + decision quality label is the labeled training dataset. This is the moat.

Decision Arc — COM-04 Meridian AI Program
2026-03-30 · COMMITTED
$3.2M committed · 85% conviction locked
Thesis and kill criteria recorded immutably to the ledger.
2026-04-14 · INTERCEPTED
Reversal → FEAR REVERSAL · 94% confidence
Board pressure dominant. No criteria triggered.
2026-04-14 · HELD
CEO held course after confrontation
30/60/90-day reviews auto-scheduled. Ledger row written.
2026-04-29 · OVERDUE
30-day review past due
Dashboard flags it in red. One click → outcome review overlay.
2026-05-29 · DUE IN 3 DAYS
60-day review imminent
Is the thesis still holding? Record the data now.
2026-06-28 · UPCOMING
90-day final review
Achieved / Partial / Failed. Decision quality label set. Dataset grows.
◎ The Labeled Dataset — The Moat
Every outcome review captures: verdict, thesis_held, actual_result, and decision_quality (right_held | wrong_held | right_reversed | wrong_reversed). Ground truth. No one else has it.
RIGHT HELD
Held course, thesis validated. Exit multiple protected.
WRONG HELD
Held course, thesis failed. Signals were there.
RIGHT REVERSED
Reversed on evidence. Kill criteria correctly triggered.
WRONG REVERSED
Reversed on fear. The loss KEEL prevents.
★ Immutable once recorded
Verdict + decision quality per outcome row. Compounds with every portco. Nobody else is collecting this.
11 / 18
Feature 05 — Live Monitoring

Kill criteria evolve.
Board sentiment as data. Both logged forever.

Kill Criteria — Live Updates

Every KC value change is logged immutably — old value, new value, triggered state, timestamp. No silent edits. No retroactive rationalization. When a criterion is marked TRIGGERED, it flags for immediate Decision Room review.

KC Update — Immutable Log Entry
criterionAutomation accuracy ≥ 93%
old_value93.4% — clear
new_value91.2% — clear
triggeredfalse
updated_at2026-05-27T14:32:01Z
⚑ When Triggered
Marking a criterion TRIGGERED flags the commitment for immediate review. The next intercept will classify as analysis_reversal — and KEEL will recommend reversing.
Board Signals — Structured Sentiment

After every board call, log a signal per member: Comfortable / Neutral / Concerned / Escalated / Approved / Rejected. Creates a structured sentiment timeline. Board anxiety becomes a leading indicator — not an anecdote.

Board Signal — Victoria Hammond
Concerned 2026-05-15 · COM-04
"Asked for monthly data briefing on automation accuracy. Wants to see 95% by Q3 before voting to approve Phase 2 expansion."
Comfortable Neutral Concerned Escalated Approved Rejected
◎ Why This Matters
A "Concerned" signal today is the precursor to a fear-reversal attempt in 30 days. KEEL captures it in structured form — so when the reversal arrives, the engine references the exact sequence of board signals.
12 / 18
Feature 06 — Intelligence Layer

Cross-portfolio intelligence. One query.

Full portfolio context injected into every chat turn. Three companies, all commitments, all intercept history, all board signals — available in a single streaming response. Plus the Portfolio Mirror: what peers who held vs. reversed actually achieved.

keel / cos-chat
"Which portco is most at risk right now, and what should I be watching?"
At Meridian: board confidence eroding while AI program is in the J-curve. Fundamentals tracking (93.4%, 19 months) but if board support breaks before Phase 1 completes, execution degrades. Watch the June board call closely.

At Apex: supplier adoption running 71% vs. 75% threshold. If it doesn't reach 75% by end of Q2, the inventory thesis breaks — already $300K over budget.

At Vertex: expansion on track. No flags.
⬡ Context Architecture
Every turn injects: all portcos + financials, all commitments + kill criteria, all provenance intercepts, board compositions, team commitments, weekly priority history. Full portfolio. One response.
Portfolio Mirror — The Benchmark

Computed from real ledger data: how leaders who held vs. reversed under pressure actually fared — measured in revenue growth, margin recovery, and exit multiple.

Peer Cohort — Holders vs. Reversers
+18%
Revenue · Holders
+3%
Revenue · Reversers
Margin recovery · holders+8.2 pts
Margin recovery · reversers−1.1 pts
Exit multiple gap2.1× median
13 / 18
Feature 07 — CEO Operating System

The full operating picture. One console.

Three supporting modules: weekly priority integrity, direct report accountability, and board commitment tracking — each one feeding signals into the Decision Provenance engine.

📅
My Week
Four weeks of priority history. KEEL detects avoidance patterns — tasks moved repeatedly without progress, flagged with times_moved counter.

The question: what are you avoiding, and why?
Priority historyAvoidance detectionKEEL flags
👥
My Team
Direct reports with committed words from 1:1s. KEEL tracks whether commitments are delivering — on-track, at-risk, missed.

Full team accountability pattern visible at a glance.
Direct reportsCommitted wordsMiss tracking
🏛
Board Room
CEO-to-board commitments with exact words quoted. Board stance tracking. Countdown timers to each due date. Signal log per member.

The board never hears "I think I said..." again.
Exact wordsStance trackingCountdownSignal log
4
Weeks of priority history
4
Direct reports tracked
3
Board members with stance history
0
Commitments without a timestamp
14 / 18
Chapter 4 · The Moat — 08 · GP Portal

The GP view. Cross-portfolio, always on.

The operating partner sees every intercept, every KC alert, and every board signal across all portcos — briefed in 30 seconds. This is the view your CEOs never see. You do.

keel / gp-inbox · 3 portfolio companies · week of 2026-05-26
1 ACTIVE
1 KC ALERT
Meridian Logistics
Marcus Chen · COM-04 · AI Automation Pilot · $3.2M
Fear Reversal · 94%
2026-05-14 · HELD · day 44 of J-curve
Apex Manufacturing
David Park · COM-02 · Supplier Digitization · $890K
KC Alert · 71% / 75% threshold
No intercept yet · Watch: end of Q2
Vertex Software
Sarah Mitchell · COM-03 · SaaS Expansion · $1.1M
Clean · All criteria clear
60-day review: all green · 2026-05-20
30-DAY REVIEW OVERDUE · COM-04 · Meridian · +15 DAYS PAST DUE
1
Active intercept this week
1
KC approaching threshold
1
Portco clean
30s
Time to full brief
15 / 18
09 · The Compounding Moat

The dataset nobody
else is building. Nobody.

🏢
Portcos
Each GP deployment adds a company
Intercepts
Every reversal attempt captured
🗂
Labeled Data
decision → outcome → quality label
🧠
Better Classifier
Real model replaces LLM-as-judge
🤝
More Trust
Accuracy earns expansion
🏢
More Portcos
Flywheel accelerates
Labeled Row Projection
1 portco × 1 yr~40 labeled rows
3 portcos × 2 yr~240 labeled rows
10 portcos × 3 yr~1,200 labeled rows
Classifier threshold1,000+ rows
◎ What Happens at 1,000 Rows
The LLM-as-judge (current, provisional) is replaced by a model trained on real PE execution outcomes. Not rubric-based inference — ground truth. No competitor has this dataset. It is not replicable after we have it.
★ The Irreplicability Moat
The state-capture at commitment is what makes the label trustworthy. Without the original conviction score, thesis, and kill criteria, the label is just opinion. KEEL has been capturing state since day one. Competitors start at zero.
⚑ Current Status
Classifier is LLM-as-judge against a rubric — provisional, accurate, but not yet trained on labeled data. Every portco deployment from today forward is building the training set.
16 / 18
10 · Why Best in World

Nothing else captures
the state.

Every competing tool captures decisions. None capture the state in which the decision was made. That state — thesis, conviction, exit criteria, emotional context — is what makes confrontation possible and the dataset irreplicable.

Capability KEEL EOS / Traction OKR Tools Board Portals
Record decision state at commitment Thesis + conviction + KC
Intercept reversals in real-time LLM classify <1s
Fear vs. analysis classification 3-class + confidence
Dual adversarial analysis Bull + Bear synthesis
30/60/90-day outcome checkpoints Auto-scheduled Manual
Decision quality labeling Immutable, per-outcome
Labeled dataset → future classifier Compounds with every portco
◎ State Moat
Without the state, confrontation is impossible. You cannot confront a decision-maker with their prior reasoning if you never captured it.
★ Data Moat
The labeled dataset (decision → outcome → quality) is the second moat. No competitor has it. At 1,000 rows, KEEL trains a real classifier. Compounding advantage.
⬡ Trust Moat
When kill criteria trigger, KEEL recommends reversing. Honesty at the classification layer is what makes the confrontation credible and the product defensible.
17 / 18
11 · What KEEL Is Not

The objections. Addressed directly.

Four questions every PE partner asks. Four honest answers.

"Is this a spy tool?"
No. The CEO sees every classification. The GP sees it too. The premise is full transparency — not surveillance. The CEO opted into accountability when they made the commitment. KEEL surfaces their own prior reasoning back to them — nothing more.
"Can't ChatGPT do this?"
Any LLM can generate a verdict. No LLM has your commitment state, your exact kill criteria, your prior words at conviction. The moat is the data and the state-capture architecture — not the model. ChatGPT starts from nothing. KEEL starts from everything you committed to.
"Won't it just always say 'fear'?"
No. If kill criteria trigger, KEEL classifies analysis_reversal and recommends reversing. A system that manufactures fear narratives destroys the entire trust premise. Honesty is the product — the engine is tested for it on every deploy. The test must always pass.
"How expensive and complex is this to start?"
Browser-based. No IT dependency. No data migration. No integration required. One portco. Thirty days. A single conversation. The CEO gets a structured system for their biggest active commitment. The ledger starts building on day one.
◎ The Non-Negotiable
KEEL classifies, confronts, and logs. The human decides. Always. This line is legally and ethically load-bearing. No endpoint, copy, or UI may imply KEEL decides.
18 / 18
KEEL
AGENTIC CHIEF OF STAFF · DECISION PROVENANCE FRAMEWORK
Chapter 5 · Next Step — Phase 0 Complete · Phase 1 Now

One portco. Thirty days.
Book a 20-minute call.

Phase 0 — Complete
Full CEO OS built and live
Five modules. All agents on real Claude. Decision Provenance Engine, Dual Dissent, CoS Chat, Decision Arc, Outcome Reviews — all production-ready on Railway.
Phase 1 — Now
One PE design partner, one portco, thirty days
Single portco. No IT dependency, no data migration, no upfront commitment beyond access to one CEO. The ledger starts building on day one.
Phase 2 — After validation
Multi-portco deployment
Cross-portfolio intelligence activates. The Portfolio Mirror becomes real. The labeled dataset starts compounding. The moat becomes visible.
Non-Negotiable
KEEL classifies, confronts, and logs. The human decides. Always. Honesty is the product.
The single ask.

One portfolio company. Thirty days. No integration required — KEEL runs in a browser. No IT dependency. No data migration. The CEO gets a structured system for their biggest active commitment. The GP gets visibility into every reversal attempt before it happens.

Book a 20-minute call. See the live demo. Bring one commitment that's currently under pressure.

keel-hq.com
Live
Production deployment
20
Minutes to see it
30d
The full ask