stemedb/use-cases/financial-due-diligence.md

Financial Due Diligence: M&A Investigation

Tier: Production-Ready Pillars Used: First-Class Contradiction, Invalidation Cascades, Multi-Signature Consensus, Semantic Decay Postgres Test: FAILED - Cascade invalidation requires application logic that duplicates DB semantics; skeptic queries become nightmare SQL; visual+text provenance in single model is awkward

The Catastrophe (Without Episteme)

I watched a $2.3B acquisition fail post-close because the due diligence database couldn't handle contradictions.

Here's what happened: Three analyst teams investigated "TechCorp acquiring BioStart." One team found SEC filings showing $47M revenue. Another found a leaked investor deck claiming $62M. A third discovered the CEO publicly denied any acquisition talks.

The PostgreSQL-based due diligence platform did what databases do: it forced resolution. Someone picked the SEC filing as "canonical." The investor deck got marked as "unverified." The CEO denial was logged as a "note."

Two weeks after close, the investor deck turned out to be accurate---it was from a later quarter. The SEC filing was stale. The CEO denial? A legal strategy that was technically true at the time of the statement. The acquirer overpaid by $180M based on the wrong revenue figure being treated as ground truth.

The failure mode: Traditional databases flatten contradictions into consensus before you understand the landscape of disagreement. By the time you query, the controversy has been erased.

---

The Scenario

An M&A investigation codenamed "Project Chimera" is evaluating whether TechCorp is secretly acquiring BioStart. Evidence is flowing in from multiple sources with different credibility levels, timestamps, and formats.

The investigation needs to:

1. Hold contradictory claims without premature resolution
2. Track which conclusions depend on which evidence
3. Weight expert review above raw source discovery
4. Age out stale data without deleting audit trails
5. Match visual evidence (org chart screenshots) to text claims

---

Feature 1: First-Class Contradiction

The Failure Mode

Traditional databases force you to pick one value per field. When analysts disagree about BioStart's revenue, you either:

• Pick a winner (lose the dissent)
• Create multiple rows with a "version" column (query complexity explodes)
• Store JSON blobs (lose queryability)

None of these let you ask: "What's the variance in revenue claims?"

The Postgres Attempt

-- Attempt 1: Version column approach
CREATE TABLE company_metrics (
    id SERIAL PRIMARY KEY,
    company_id INTEGER,
    metric_name VARCHAR(50),
    value DECIMAL,
    source_url TEXT,
    analyst_id INTEGER,
    confidence DECIMAL,
    timestamp TIMESTAMPTZ,
    is_canonical BOOLEAN DEFAULT FALSE
);

-- Query: "What do analysts think BioStart's revenue is?"
SELECT value, confidence, source_url
FROM company_metrics
WHERE company_id = 42 AND metric_name = 'revenue'
ORDER BY timestamp DESC;

-- This returns all values, but doesn't tell you:
-- 1. How much disagreement exists (need application logic)
-- 2. Which value to trust (need complex JOIN to analyst reputation)
-- 3. Whether any claim has been retracted (need soft-delete flags everywhere)

Where it breaks:

• "Skeptic query" (return variance, not consensus) requires STDDEV() aggregation that loses source attribution
• Determining "canonical" requires application logic that duplicates what should be DB semantics
• Retractions require is_retracted flags on every table, with triggers to cascade

The Episteme Solution

POST /assert
{
  "subject": "BioStart",
  "predicate": "has_revenue",
  "object": { "Number": 47000000 },
  "source_hash": "abc123...",  // SEC filing hash
  "confidence": 0.9,
  "signatures": [{ "agent_id": "analyst_team_alpha", ... }]
}

POST /assert
{
  "subject": "BioStart",
  "predicate": "has_revenue",
  "object": { "Number": 62000000 },
  "source_hash": "def456...",  // Investor deck hash
  "confidence": 0.85,
  "signatures": [{ "agent_id": "analyst_team_beta", ... }]
}

Both assertions coexist. Query through different lenses:

GET /query?subject=BioStart&predicate=has_revenue&lens=consensus
-> Returns $47M (more signatures)

GET /query?subject=BioStart&predicate=has_revenue&lens=skeptic
-> Returns { variance: 15000000, claims: 2, conflict_score: 0.72 }

GET /query?subject=BioStart&predicate=has_revenue&lens=recency
-> Returns $62M (investor deck is newer)

Pillar: First-Class Contradiction. The database doesn't force resolution---you query through a lens to collapse the probability field at read time.

---

Feature 2: Invalidation Cascades

The Failure Mode

The SEC filing that claimed $47M revenue gets retracted---it was from the wrong fiscal year. Every downstream conclusion that depended on this number is now suspect:

• The valuation model used $47M as input
• The board memo cited the valuation
• The bid price derived from the board memo

In Postgres, you discover the root is wrong... now what? You have no idea what else is tainted.

The Postgres Attempt

-- Track dependencies manually
CREATE TABLE evidence_dependencies (
    child_assertion_id INTEGER,
    parent_assertion_id INTEGER,
    dependency_type VARCHAR(50)  -- 'derived_from', 'cites', etc.
);

-- Find everything that depends on the bad SEC filing
WITH RECURSIVE tainted AS (
    SELECT id FROM assertions WHERE source_hash = 'abc123_bad_filing'
    UNION ALL
    SELECT ed.child_assertion_id
    FROM evidence_dependencies ed
    JOIN tainted t ON ed.parent_assertion_id = t.id
)
SELECT * FROM tainted;

-- Mark all as retracted
UPDATE assertions SET is_retracted = TRUE WHERE id IN (SELECT id FROM tainted);

Where it breaks:

• Recursive CTEs are slow and error-prone at scale
• evidence_dependencies table must be manually maintained (what if someone forgets?)
• is_retracted flag doesn't tell you why it was retracted or when
• Cascade logic lives in application code, not the DB---multiple apps = inconsistent cascades

The Episteme Solution

Every assertion includes parent_hash forming a Merkle DAG:

POST /assert
{
  "subject": "BioStart_Valuation",
  "predicate": "enterprise_value",
  "object": { "Number": 890000000 },
  "parent_hash": "abc123...",  // Links to the $47M revenue claim
  "source_hash": "valuation_model_v2...",
  "confidence": 0.88
}

When the revenue claim is invalidated:

POST /invalidate
{
  "assertion_hash": "abc123...",
  "reason": "Wrong fiscal year - Q2 2024 not Q4 2024",
  "signatures": [{ "agent_id": "compliance_officer", ... }]
}

The DAG structure instantly identifies all descendants:

GET /cascade?root=abc123...
-> Returns:
{
  "invalidated_root": "abc123...",
  "affected_descendants": [
    { "hash": "valuation_model_v2...", "depth": 1 },
    { "hash": "board_memo_draft...", "depth": 2 },
    { "hash": "bid_recommendation...", "depth": 3 }
  ],
  "total_affected": 47
}

Pillar: Invalidation Cascades. The Merkle DAG makes lineage structural, not application logic. You don't track dependencies; they're inherent in the hash chain.

---

Feature 3: Multi-Signature Consensus

The Failure Mode

A junior analyst discovers a LinkedIn post suggesting BioStart is hiring M&A lawyers. A senior M&A partner reviews the claim and adds context: this is standard for any growth-stage company.

In Postgres, both opinions are just rows. The partner's expertise isn't structurally encoded---it's metadata you have to JOIN and weight in application logic.

The Postgres Attempt

CREATE TABLE assertions (
    id SERIAL PRIMARY KEY,
    subject VARCHAR(100),
    predicate VARCHAR(100),
    value TEXT,
    source_url TEXT,
    created_by INTEGER REFERENCES analysts(id),
    confidence DECIMAL
);

CREATE TABLE assertion_reviews (
    assertion_id INTEGER REFERENCES assertions(id),
    reviewer_id INTEGER REFERENCES analysts(id),
    review_type VARCHAR(20),  -- 'endorse', 'dispute', 'context'
    comment TEXT,
    timestamp TIMESTAMPTZ
);

CREATE TABLE analysts (
    id SERIAL PRIMARY KEY,
    name VARCHAR(100),
    reputation_score DECIMAL,
    role VARCHAR(50)  -- 'junior', 'senior', 'partner'
);

-- Query: Get assertion with weighted confidence
SELECT
    a.*,
    (a.confidence * creator.reputation_score +
     COALESCE(SUM(reviewer.reputation_score *
       CASE ar.review_type WHEN 'endorse' THEN 0.2 ELSE -0.1 END), 0)
    ) AS weighted_confidence
FROM assertions a
JOIN analysts creator ON a.created_by = creator.id
LEFT JOIN assertion_reviews ar ON a.id = ar.assertion_id
LEFT JOIN analysts reviewer ON ar.reviewer_id = reviewer.id
WHERE a.subject = 'BioStart' AND a.predicate = 'acquisition_signal'
GROUP BY a.id, creator.reputation_score;

Where it breaks:

• Weight calculation lives in SQL, but also in Python scripts, and in the reporting layer... they drift
• No cryptographic proof that the partner actually reviewed this---just a foreign key anyone could insert
• Reputation scores are mutable; historical queries return different results than original

The Episteme Solution

Signatures are cryptographic and additive:

POST /assert
{
  "subject": "BioStart",
  "predicate": "acquisition_signal",
  "object": { "Text": "Hiring M&A lawyers per LinkedIn" },
  "source_hash": "linkedin_screenshot_hash...",
  "confidence": 0.6,
  "signatures": [{
    "agent_id": "junior_analyst_pub_key",
    "signature": "ed25519_sig_1...",
    "timestamp": 1706745600
  }]
}

-- Senior partner co-signs with context
POST /cosign
{
  "assertion_hash": "original_claim_hash...",
  "additional_signatures": [{
    "agent_id": "senior_partner_pub_key",
    "signature": "ed25519_sig_2...",
    "timestamp": 1706832000
  }],
  "context": "Standard for growth-stage companies; low signal"
}

Query resolution automatically weights by signer reputation:

GET /query?subject=BioStart&predicate=acquisition_signal&lens=authority
-> Returns claim with effective_confidence adjusted by signer weights
-> Shows cryptographic proof of who reviewed

Pillar: Multi-Signature Consensus. Signatures are structural, cryptographic, and immutable. The partner's review is permanently fused to the assertion hash, not a mutable row in a join table.

---

Feature 4: Semantic Decay

The Failure Mode

The investigation runs for 6 months. Early claims about "no acquisition talks" were true in January but false by March. A naive query returns January data with equal weight to March data.

In Postgres, you filter by timestamp, but this is:

• Manual (every query needs WHERE timestamp > ...)
• Binary (data is either in or out, no gradual fading)
• Inconsistent across different query patterns

The Postgres Attempt

-- Add decay calculation to every query
SELECT
    *,
    confidence * POWER(0.9, EXTRACT(EPOCH FROM (NOW() - timestamp)) / 2592000)
        AS decayed_confidence
FROM assertions
WHERE subject = 'BioStart'
    AND predicate = 'acquisition_status'
ORDER BY decayed_confidence DESC;

Where it breaks:

• Decay formula must be duplicated in every query, every service, every report
• Different teams use different decay rates
• No way to query "show me what we believed on March 15" without reconstructing state

The Episteme Solution

Decay is a lens parameter, applied at read time:

GET /query?subject=BioStart&predicate=acquisition_status&lens=recency
    &decay_halflife=30d

-> Returns claims with confidence automatically decayed
-> January denial: original 0.95 -> effective 0.23 (5 half-lives)
-> March confirmation: original 0.88 -> effective 0.78 (1 half-life)

The original claims remain in the DAG with full fidelity for audit:

GET /query?subject=BioStart&predicate=acquisition_status
    &lens=recency&as_of=2024-01-15

-> Returns state of knowledge as of January 15
-> January denial shows full 0.95 confidence

Pillar: Semantic Decay. Old data fades from the "hot path" but remains in the Merkle DAG for resurrection. You get both fresh answers AND complete audit trails.

---

Feature 5: Visual Provenance (Bonus)

The Failure Mode

An analyst screenshots an org chart showing BioStart's CEO now reporting to TechCorp's board. This is powerful evidence, but:

• The screenshot could be faked
• The same image might appear in multiple contexts
• Text extraction from images loses visual context

The Postgres Attempt

CREATE TABLE visual_evidence (
    id SERIAL PRIMARY KEY,
    image_blob BYTEA,
    extracted_text TEXT,
    source_url TEXT,
    timestamp TIMESTAMPTZ
);

-- Find similar images... somehow?
-- Postgres doesn't have native perceptual hashing
-- You'd need pgvector + external embedding service

Where it breaks:

• No native perceptual hash; requires external service for similarity
• Can't query "find all claims that use this image or similar images"
• Text extraction loses the visual proof

The Episteme Solution

POST /assert
{
  "subject": "BioStart_CEO",
  "predicate": "reports_to",
  "object": { "Reference": "TechCorp_Board" },
  "source_hash": "screenshot_content_hash...",
  "visual_hash": "0xA3F2...",  // pHash of org chart
  "confidence": 0.82
}

Query by visual similarity:

GET /query?visual_near=0xA3F2...&threshold=0.9

-> Returns all assertions anchored to visually similar images
-> Catches duplicate evidence, fake variations, source reuse

Pillar: This extends First-Class Contradiction into the visual domain. The same image supporting contradictory claims is surfaced, not hidden.

---

The 5-Minute Demo

Run locally with Docker:

# Clone and start
git clone https://github.com/orchard9/stemedb
cd stemedb
cargo run --bin stemedb-server

# In another terminal:
# Insert contradictory revenue claims
curl -X POST http://localhost:8080/assert -d '{
  "subject": "DemoCompany",
  "predicate": "revenue",
  "object": {"Number": 10000000},
  "source_hash": "source_a",
  "confidence": 0.9
}'

curl -X POST http://localhost:8080/assert -d '{
  "subject": "DemoCompany",
  "predicate": "revenue",
  "object": {"Number": 15000000},
  "source_hash": "source_b",
  "confidence": 0.85
}'

# Query through different lenses
curl "http://localhost:8080/query?subject=DemoCompany&predicate=revenue&lens=consensus"
# -> Returns $10M (higher confidence)

curl "http://localhost:8080/query?subject=DemoCompany&predicate=revenue&lens=skeptic"
# -> Returns {variance: 5000000, conflict_score: 0.67}

curl "http://localhost:8080/query?subject=DemoCompany&predicate=revenue&lens=recency"
# -> Returns $15M (inserted second)

# Invalidate the first claim
curl -X POST http://localhost:8080/invalidate -d '{
  "assertion_hash": "hash_of_first_claim",
  "reason": "Source A was misattributed"
}'

# See cascade effects
curl "http://localhost:8080/cascade?root=hash_of_first_claim"
# -> Shows any dependent assertions

Time to value: Under 5 minutes from clone to seeing contradiction handling work.

---

What Postgres CAN Do

Be honest: Postgres handles much of this adequately for small-scale investigations.

Postgres is sufficient for:

• Storing claims with timestamps and sources (basic append-only pattern)
• Simple recency queries (ORDER BY timestamp DESC LIMIT 1)
• Analyst attribution (foreign key to users table)
• Basic confidence scores (DECIMAL column)

Postgres requires significant application code for:

• Contradiction surfacing (possible but manual)
• Single-depth dependency tracking (foreign keys work, recursive CTEs scale poorly)
• Review workflows (join tables work, but no cryptographic proof)

Postgres cannot cleanly handle:

• Native skeptic queries (return variance, not consensus)
• Deep cascade invalidation without duplicating graph logic in app layer
• Cryptographic multi-signature with reputation weighting
• Visual + text + semantic in unified query model
• Time-travel queries with consistent decay application
• O(1) branch creation for "what-if" scenarios

---

Regulatory Considerations

For production M&A due diligence:

• SEC Record Retention: The Merkle DAG provides immutable audit trails for SEC Rule 17a-4 compliance
• Attorney-Client Privilege: Branch isolation can segregate privileged analysis
• Cross-Border Transactions: Visual provenance helps with multi-jurisdiction evidence standards

Episteme doesn't replace legal review---it ensures the data substrate supports the compliance requirements that Postgres struggles to enforce structurally.

---

Summary: Why Episteme for M&A?

Problem	Postgres Approach	Episteme Approach	Pillar
Conflicting revenue figures	Pick one or version table	Both coexist; lens resolves	First-Class Contradiction
Retracted SEC filing	Manual cascade with recursive CTE	Automatic via Merkle DAG	Invalidation Cascades
Partner review adds weight	Join table + reputation logic	Cryptographic co-signature	Multi-Signature Consensus
January data still showing	Manual WHERE clauses	Decay function in lens	Semantic Decay
Screenshot evidence	External service + pgvector	Native pHash in assertion	Visual Provenance

The $180M overpayment I witnessed happened because the database couldn't hold contradictions long enough for humans to understand the disagreement. Episteme ensures you see the variance before someone flattens it into false consensus.