Tuesday, December 30, 2025 Leave a Comment

The Art of the Corporate "Spin": Deconstructing Deception in the Tobacco Industry.

This is a short follow up from the previous blog.

How does a company talk its way out of a crisis? In 1996, the tobacco industry faced a "crack in the dam" when the Liggett Group settled a massive class-action lawsuit, breaking the industry’s decades-long "united front."

Philip Morris CEO Geoff Bible had to address his global workforce. Behind the scenes, the speech went through four rigorous drafts. By comparing the first draft (**gb1**) to the final version (**gb4**), we can see a masterclass in how language is "hardened" to manufacture confidence and deflect blame.

## 1. The Foundation: Brooke Heller’s Thesis

In her seminal work, *"Manipulative Language in Corporate Discourse,"* Brooke Heller performed a forensic linguistic analysis of these drafts. She discovered that manipulation isn't just about what is added—it’s about what is **erased**.

* **The Deletion of "Addiction":** Heller found that the word "addiction" appeared in early drafts but was surgically removed by the final version, replaced with euphemisms to avoid legal liability.

* **Dehumanizing Whistleblowers:** Early drafts acknowledged that opponents were "entitled to their views." The final version stripped this away, reframing whistleblowers not as people, but as "fodder for disinformation."

* **The Pep Rally Effect:** Heller argues the final text was no longer a corporate update; it was a "pep rally" designed to internalize a sense of ethical superiority among employees.

## 2. The Data: Measuring the "Hardening" Effect

When I apply the **Harvard General Inquirer (GI)** categories to these two versions, the shift from a defensive legal posture to an aggressive "moral" one becomes visible in the numbers.

### What decreased in the final draft?

The final version saw a massive drop in "vulnerable" language. By removing **Hostile (-7)** and **NegAff (-6)** words, the speech was "cleaned" of the raw anger and negativity that might make a leader look desperate. Crucially, **Weak (-4)** and **Fail (-3)** categories were eliminated, closing the "cracks in the dam" linguistically before they could be felt by the audience.

### What increased in the final draft?

The most fascinating shifts occurred in how "strength" was re-inserted:

* **Affirmation and Virtue (+8 Yes, +4 Virtue):** The speaker leans heavily into moral certainty, concluding with a new paragraph explicitly claiming Philip Morris is an "ethical company."

* **The "If" Paradox (+9):** Interestingly, conditional language increased. This suggests **Strategic Ambiguity**—using "if/then" scenarios to make bold promises about the future without being tethered to concrete, legal facts.

* **Certainty and Action (+4 IAV, +3 Know):** Interpretive Action Verbs (IAV) like "understand" and "know" were added to force the audience into a shared perspective.

## 3. Summary: From Defense to Defiance

The evolution from **gb1** to **gb4** represents a total pivot in corporate strategy.

| Feature | Initial Draft (gb1) | Final Draft (gb4) |

| **Tone** | Cautious, Explanatory | Defiant, Moralistic |

| **Logic** | Defensive (Explaining the law) | Offensive (Attacking the "Sham") |

| **Ethos** | Corporate leadership | Moral crusaders for "Freedom" |

As my analysis shows, the final version is significantly more "extreme" because it replaces **uncertainty with manufactured certainty**. It is a document that has been polished to remove accountability, using simplified logic and high-intensity emotional language to bypass the listener's critical faculties.

By preserving these analyses, we gain a vital tool for the future: the ability to recognize when "corporate values" are being used as a shield for systemic deception.

### Executive Summary: The Evolution of Corporate Persuasion

**Case Study: Philip Morris CEO Geoff Bible’s "State of the Company" Speech (1996)**

This table summarizes the linguistic transformation of a high-stakes corporate message from its first draft (**gb1**) to its fourth and final version (**gb4**). It tracks how "defensive" language was replaced with "moral defiance."

| --- | --- | --- | --- |

| **Opponent Framing** | Acknowledged as people with "views." | Labeled as a "Sham" and "Desperation move." | **Dehumanization:** Stripping credibility from critics to unify the internal team. |

### Quantitative Indicators of "Extreme" Persuasion

*Data derived from Harvard General Inquirer (GI) Analysis*

#### 📉 What was DELETED (The "Cracks in the Dam")

The final draft surgically removed words that suggested vulnerability or negative reality:

* **Weak (-4) & Fail (-3):** Eliminated any suggestion that the company was losing.

* **Legal (-2):** Reduced the "lawyerly" tone to sound more human and relatable.

* **Negate (-3):** Replaced negative statements with affirmative ones.

#### 📈 What was ADDED (The "Hardened" Front)

The final draft increased categories that manufacture a sense of control and virtue:

* **Yes (+8) & Virtue (+4):** A massive spike in affirmative and moralistic language.

* **IAV (+4) & Know (+3):** Interpretive Action Verbs (e.g., "we understand," "we know") were used to dictate how the audience should interpret reality.

* **Persist (+4) & Active (+2):** Strengthened the language of determination and movement.

### The "Heller" Insights: What was hidden?

Linguistic researcher Brooke Heller identified that the most deceptive parts of this evolution were the **omissions**:

1. **Term Deletion:** The word **"Addiction"** was present in early drafts but entirely removed from the final version to avoid legal triggers.

2. **Financial Cloaking:** References to **dropping stock prices** were deleted to maintain a false sense of total stability.

3. **Syntactic Simplification:** The text was rewritten into punchier, shorter sentences to lower "cognitive load," making the manipulative logic easier for the listener to accept without questioning.

**Conclusion:** The shift from Draft 1 to Draft 4 is not just an edit; it is a psychological realignment. It demonstrates how corporate discourse can move from a state of **crisis management** to a state of **manufactured moral triumph** through precise linguistic engineering.

IN ADDITION:

The "Frank Statement to Cigarette Smokers" (1954) is essentially the **ancestor of the deceptive patterns** found in the 1996 Bible drafts. While separated by over 40 years, the "Frank Statement" uses the exact same linguistic clusters I identified to manage a massive public health crisis—the initial link between smoking and lung cancer.

Here is how the "Frank Statement" aligns with my "High-Deception Profile" and Brooke Heller’s findings:

### 1. The "Moral Shield" (High Virtue & PosAff)

Just like the final Philip Morris draft (**gb4**) suddenly emphasizes being an "ethical company," the 1954 statement anchors itself in moral authority.

* **Virtue/Yes Cluster:** The statement claims, *"We accept an interest in people’s health as a basic responsibility, paramount to every other consideration"*. This is the ultimate "Virtue" play, designed to pre-emptively block the "Hostile" accusations of critics.

* **PosAff Cluster:** It frames tobacco as having given *"solace, relaxation and enjoyment to mankind"* for 300 years, using positive affect to distract from the "Negate" findings of doctors.

### 2. The Sanitization of Failure (Low Weak & Fail)

The 1954 document is a masterclass in **Aggressive Omission**.

* **Scrubbing Weakness:** Despite a drop in cigarette consumption and stock prices following "Cancer by the Carton" reports, the statement contains zero language suggesting the industry is in trouble.

* **Removing Fail/Legal:** It reframes scientific "Failure" (the inability to disprove cancer links) as a lack of "conclusive" proof. It replaces specific legal or medical terminology with the softer, more human category of "deep concern".

### 3. The Ambiguity Paradox (High If & Know)

The "Frank Statement" uses **Strategic Ambiguity** even more effectively than the Bible speech.

* **High "If" (Conditionals):** It uses conditionals to create doubt: *"Even though its results are inconclusive,"* it shouldn't be dismissed, but it should be questioned by *"distinguished authorities"*. This creates the **"If" paradox** I noted—using the language of uncertainty to appear "balanced" while actually manufacturing doubt.

* **High "Know" (Certainty):** Simultaneously, it uses certainty language (the **Know/IAV cluster**) to state there is *"no proof"* and *"no agreement"*. It forces the interpretation that the science is a "theory" rather than a fact.

### Summary of Cross-Document Extrapolation

| GI Cluster Profile | Frank Statement (1954) | Bible Draft (gb4, 1996) |

### Conclusion

The 1954 "Frank Statement" proves my theory: **Drafts are not strictly necessary to detect deception if the "Linguistic Signature" is present.** This document matches my "High-Deception Profile" almost perfectly. It shows that the tobacco industry has used the same "Category Clusters"—spiking **Virtue** and **If** while scrubbing **Weak** and **Fail**—for over half a century to socially engineer public doubt.

[Deception in the 1954 Frank Statement]

This video provides historical context on how the 1954 "Frank Statement" was the opening move in a decades-long campaign of organized public deception.

Monday, December 29, 2025 Leave a Comment

Quantifying Deception: How Philip Morris Refined a Speech Through Four Drafts

When Words Become Weapons: A Computational Analysis

In 1996, Philip Morris CEO Geoffrey Bible faced a crisis. Whistleblowers were coming forward, the FDA was investigating, and a major tobacco company had just broken ranks to settle lawsuits. He needed to address employees worldwide. What he said mattered—but more importantly, *how* he revised what he said reveals the machinery of corporate manipulation.

Brooke Heller's Discovery: Deception Through Revision

In her 2007 Master's thesis, linguist Brooke Heller made a remarkable argument: **you cannot find deception in individual words, but you can find it in the pattern of crafting and re-crafting text across drafts.**

Heller analyzed four sequential drafts of Bible's April 9, 1996 speech using discourse analysis techniques. She examined semantic changes (how meaning shifts on the page) and pragmatic changes (how meaning shifts in real-world context). Her findings were striking:

What Heller Found Through Close Reading:

**Systematic Omissions:**

- The word "addiction" appeared in early drafts but vanished by the final version

- Acknowledgments of being "slow to respond" were deleted

- References to stock price impacts from the Liggett settlement disappeared

**Strategic Evasions:**

- Discussion shifted from whistleblowers as credible scientists to dismissing their "statements"

- "Addiction" was replaced with the euphemism "to keep people smoking"

- Focus redirected from defending against accusations to attacking the accusers

**Language Softening:**

- "Dead wrong" became "In truth"

- "Destroy significant value" became "temporarily erode significant value"

- "Must continue to expect further trouble" became "can expect more leaks"

**Dehumanization:**

- "Three individuals who have given these affidavits" → "the three who have given these statements" → "former employees"

Heller applied Galasinski's typology of deceptive language (omission, evasion, explicit commission, implicit commission) and Grice's maxims of cooperative conversation. Her conclusion: the editing process itself reveals intentional manipulation.

But Heller's analysis was qualitative—close reading of selected passages. Could we validate her findings quantitatively?

My Experiment: Counting the Changes

I wanted to test Heller's thesis computationally. If the revision process really did show systematic manipulation, it should be measurable.

**Method:**

1. I took Draft 1 (GB1) and the Final Version (GB4)

2. I ran both through a word counter using the Harvard General Inquirer dictionary

3. The General Inquirer categorizes words into psychological and sociological categories (Hostile, Virtue, Strong, Weak, etc.)

4. I calculated the difference: which word categories increased and which decreased from first to final draft.

**The Question:** Would the quantitative patterns support Heller's qualitative findings?

The Results: Stunning Confirmation

### Words REDUCED from Draft 1 → Final Draft:

The top categories that decreased:

1. **Hostile** (-7): Aggressive language toned down

2. **NegAff** (-6): Negative emotional words reduced

3. **Strong** (-5): Strong words in aggressive contexts removed

4. **Weak** (-4): Language suggesting weakness eliminated

5. **Vice** (-3): Morally negative terminology decreased

6. **Submit** (-3): Language of subordination removed

7. **If** (-3): Some conditional hedging reduced

8. **Fail** (-3): Failure language eliminated

9. **Negate** (-3): Negative statements reduced

10. **Legal** (-2): Certain legal terminology decreased

### Words INCREASED in Final Draft:

The top categories that increased:

1. **If** (+9): Conditional statements dramatically increased (!)

2. **Yes** (+8): Affirmative language strengthened

3. **Strong** (+7): Strong words in confident contexts added

4. **PosAff** (+5): Positive emotional language increased

5. **Virtue** (+4): Moral and ethical language added

6. **IAV** (+4): Interpretive Action Verbs increased (belief, understanding)

7. **Persist** (+4): Determination language strengthened

8. **Know** (+3): Certainty language added

9. **Begin** (+2): Initiative language increased

10. **Active** (+2): Active voice strengthened

## What This Reveals: The Anatomy of Manipulation

The quantitative data perfectly validates Heller's qualitative findings. Here's what the numbers show:

### 1. The Emotional Arc Transformation

Draft 1: Defensive → Hostile → Vulnerable

Final Draft: Confident → Virtuous → Determined

The speech underwent a complete emotional reframing. Hostile and negative language was systematically removed, while positive, virtuous, and confident language was added.

### 2. The Legal Protection Strategy

Notice the **"If" paradox**: conditionals both decreased (-3) and increased (+9) substantially. How?

- **Removed**: Defensive conditionals that admitted vulnerability ("if our stock has taken a beating")

- **Added**: Protective conditionals that create legal wiggle room ("if we cannot get an impartial opportunity")

This is sophisticated legal defensiveness masked as reasonableness.

### 3. The Moral Reframing

The increase in **Virtue** (+4) and **PosAff** (+5) wasn't accidental:

- Added: "We are an ethical company"

- Added: "We are principled people who are honest and straight-dealing"

- Added: "We believe kids should not smoke"

This moral positioning counters the accusations without directly addressing them.

### 4. The Power Dynamics Shift

**Submit** (-3) and **Weak** (-4) decreased while **Persist** (+4) and **Strong** (+7 in confident contexts) increased. Philip Morris edited out any language suggesting they were reactive, constrained, or defensive, replacing it with language of determination and agency.

### 5. The "Strong" Paradox Explained

"Strong" appears in both the decreased AND increased categories. This reveals sophisticated reframing:

- Strong-aggressive words removed (hostility toward opponents)

- Strong-confident words added (assertiveness about their position)

Same intensity, different target—from attacking to asserting.

## Specific Examples That Match the Data

Let me show how the quantitative patterns manifest in actual text changes Heller documented:

**Hostile (-7) & NegAff (-6):**

- Draft 1: "Dead wrong"

- Final: "In truth"

- Draft 1: "hypocritical politicians"

- Final: "some hypocritical politicians"

**Weak (-4) & Submit (-3):**

- Draft 1: "I know that some of you have felt that we have been a little slow on this one and, perhaps, we have been"

- Final: [Deleted entirely]

**Fail (-3):**

- Draft 1: "destroy significant value"

- Final: "temporarily erode significant value"

**Virtue (+4) & PosAff (+5):**

- Draft 1: [Not present]

- Final: "We are an ethical company... principled people who are honest and straight-dealing"

**If (+9) - Conditional hedging:**

- Draft 1: "We can not be shooting back wildly"

- Final: "Sometimes that will mean an immediate response. Other times it will mean waiting..."

**IAV (+4) - Interpretive Action Verbs:**

- Added: "We believe" "We understand" "We know"

- These verbs control interpretation—they frame what things "mean"

## Why This Matters

This analysis proves that **deception operates at the level of process, not just content.**

Individual sentences in the final speech might seem reasonable. But the *pattern* of revision reveals intentional manipulation:

1. **Systematic removal** of vulnerability markers

2. **Systematic addition** of confidence markers

3. **Consistent direction** across multiple linguistic dimensions

4. **Strategic coherence** in the editing choices

This isn't random refinement—it's calculated reconstruction of reality.

## The Broader Implications

### For Tobacco Control Research:

Rather than searching documents for smoking-gun admissions, researchers should analyze *drafts* of key documents. The editing process reveals intent more clearly than final text.

### For Detecting Corporate Deception:

This method could be applied to:

- Political speech drafts

- Corporate crisis communications

- Legal document revisions

- Financial disclosures

- Any high-stakes persuasive text

### For Understanding Manipulation:

The combination of:

- Qualitative close reading (Heller's approach)

- Quantitative categorical analysis (this computational approach)

- Creates a powerful method for detecting deceptive intent in iterative documents.

## The "Manipulation Vector"

What we've created here is essentially a **manipulation vector**—the direction and magnitude of change across emotional and rhetorical dimensions:

- **Direction**: From defensive/negative → confident/virtuous

- **Magnitude**: Substantial changes in 20+ linguistic categories

- **Consistency**: All changes serve the same strategic goal

This vector points unambiguously toward calculated manipulation.

## Conclusion: The Machine Behind the Message

Geoffrey Bible's final speech sounds reasonable, even principled. He expresses concern about youth smoking, defends the company's ethical standards, and promises to fight fairly in court.

But the four drafts tell a different story. They reveal a systematic process of:

- Removing evidence of vulnerability

- Erasing acknowledgment of problems

- Adding moral self-justification

- Creating legal protection through careful hedging

The quantitative analysis confirms what Heller discovered through close reading: **deception isn't in the words—it's in the pattern of revision.**

When Philip Morris said "we are right," they meant it. But the drafts show they knew exactly what they needed to hide, soften, and reframe to make that claim believable.

The machine behind the message is now visible. And it's no less damaging for being well-oiled.

---

## Methodology Notes

**Heller's Approach:**

- Four sequential drafts of April 9, 1996 speech

- Discourse analysis (semantic and pragmatic)

- Galasinski's deception typology

- Grice's conversational maxims

- Paragraph-by-paragraph comparison using Draft Analysis Program

**Computational Approach:**

- Harvard General Inquirer word categorization

- Raw frequency counts across 180+ categories

- Difference calculation (GB1 - GB4)

- Top 10 increases and decreases identified

- Pattern analysis across emotional/rhetorical dimensions

**Data Validation:**

The quantitative findings independently confirmed Heller's qualitative analysis, suggesting robust patterns of manipulation visible through multiple analytical lenses.

---

*Brooke Heller's complete thesis: "Manipulative Language in Corporate Discourse: A Case Study of Deception in a Major Tobacco Industry Speech" (University of Georgia, 2007)*

*Source documents available through the Legacy Tobacco Documents Library*

Saturday, December 27, 2025 Leave a Comment

Beating a Decade-Old Deception Detection Benchmark:

80% Verbal-Only Accuracy with Raw GI Counts and FIGS

*Posted on December 28, 2025, by Tom Berger – A hobbyist dive into forensic linguistics and interpretable ML*

In the high-stakes world of automated deception detection, few benchmarks have endured like the University of Michigan's multimodal trial dataset.

Just last month (November 2025), lead researcher Rada Mihalcea and collaborators snagged the ICMI Ten-Year Impact Runner-up Award for their pioneering work on real-life courtroom deception spotting.

It's a well-deserved nod—their 2015 papers, "Deception Detection using Real-life Trial Data" and "Verbal and Nonverbal Clues for Real-life Deception Detection," clocked multimodal accuracies of 75-82%, outpacing human baselines (~54%) and setting a gold standard for fusing linguistics with gestures in adversarial settings like trials. But here's the rub: That was *ten years ago*.

No major updates, despite the explosion in LLMs and tree ensembles. As someone tinkering with open-source tools for fun (and a bit of forensic curiosity), I wondered: Can a stripped-down, verbal-only pipeline—using a 1960s dictionary, raw word counts, and a greedy tree-summing model—top their results? Spoiler: Yes. I hit 80% accuracy on lie detection with *five features*. No gestures, no black-box fusion. Just words, counts, and interpretability.

This post breaks it down for the technically inclined (think NLP/ML folks hunting deception signals) while keeping the gist accessible: Sometimes, less is more, and raw data beats polished percentages. Grab the code/data tweaks from my GitHub [link TBD], and let's dissect.

## The Michigan Benchmark: Multimodal Muscle on Real Trial Data

The Michigan team's setup is a masterclass in multimodal ambition. They scraped ~121 video clips (59 deceptive, 62 truthful) from public U.S. court trials—defendants and witnesses under oath, with real motives (freedom, convictions) that lab-elicited "mock crimes" can't touch.

Transcripts were run through LIWC (Linguistic Inquiry and Word Count), pulling 100+ normalized features like cognitive processes (e.g., % hedges like "approximately"), emotional tone, and first-person pronouns—hallmarks of cognitive load in lies per psycholinguistics lore.

Nonverbals? They hand-annotated gestures (e.g., self-adaptors like fidgeting, which spike in stress) across categories like emblems and illustrators, yielding binary flags.

Classifiers (likely SVMs or early RFs—papers are light on details) fused these at feature or decision levels: Verbal alone ~70%, gestures ~65%, multimodal fusion pushing 75-82% accuracy (and AUCs ~0.85-0.90).

No explicit cross-validation mentioned (they used train-test splits, per the 2015 EMNLP/NAACL proceedings), but the real-world stakes make it a beast for generalization testing.

Human eval? ~54% accuracy—our squishy brains lag behind silicon when spotting evasion in flat testimony. The dataset's public (h/t Michigan's LIT lab), so I downloaded it verbatim: 121 segments, balanced labels, raw transcripts + gesture binaries. Goal: Beat 82% verbal+nonverbal with *verbals only*, using off-the-shelf tools.

## My Counterpunch: Minimalism Meets Raw Power

I flipped the script—opposite everything. Where Michigan went broad and normalized, I went lean and literal. Here's the pipeline:

### Data Prep

- **Verbal Features**: Ditched LIWC's percentage outputs (e.g., % of words in "negate" category). Percentages dilute repetition signals— a liar's five "no recollections" in a 50-word clip?

That's 10% negation, but raw count=5 screams denial harder than a normalized 0.10. Used Kris Kyle's CLA (Python word counter from UH Mānoa) with a custom Harvard General Inquirer (GI) dictionary—vintage 1960s, but with 11k+ tags across Lasswell power/affiliation buckets and GI's interpretive verbs/states. Pulled raw frequencies for *five* categories that popped in exploratory PCA:

- **Card_GI**: Cardinal numbers (e.g., "one," "half mile")—liars lowball specifics.

- **Sv_GI**: State verbs (e.g., "recall," "exhausted")—internal hedging overload.

- **Polit_2_GI**: Political/ideological refs (e.g., "police," "ally")—avoidance of authority.

- **Quan_GI**: Quantity assessors (e.g., "approximately," "period")—vague abundance.

- **Notlw_Lasswell**: Negations/denials (e.g., "no," "unsuccessful")—defensive caps.

Why these? GI's granularity (vs. LIWC's broader buckets) snags forensic nuances like power evasion; raw counts preserve cognitive effort (more words = more fabrication tax).

- **No Nonverbals (Yet)**: Ignored their hard-won gesture codes—binary 0/1 for adaptors, etc. (I'll fuse 'em next; binaries might drag if not scaled, per my tests).

- **Total Features**: 5 raw ints per segment. Michigan? 100+ normalized floats + binaries. Less is interpretable.

### Modeling: FIGS for Greedy, Readable Trees

- **Why FIGS?** I've cartwheeled through CART/boosted trees and XGBoost (JMP plugin FTW) for years, but Fast Interpretable Greedy-tree Sums (FIGS) is the sleeper hit:

Additive ensemble of shallow trees (max_rules=5 here), summing leaf "Val" scores for probs (1=lie). Greedy splits on impurity, but ultra-readable—no XGBoost opacity. Trained on 121 samples, random_state=100 for repro.

FIGS output looks like this:

> FIGS-Fast Interpretable Greedy-Tree Sums:

> Predictions are made by summing the "Val" reached by traversing each tree

> ------------------------------

Card_GI <= 0.500 (Tree #0 root)

Sv_GI <= 6.500 (split)

Polit_2_GI <= 0.500 (split)

Val: 0.765 (leaf)

Val: 0.340 (leaf)

Quan_GI <= 9.500 (split)

Val: 0.022 (leaf)

Val: 0.819 (leaf)

Val: 0.222 (leaf)

Notlw_Lasswell <= 1.500 (Tree #1 root)

Val: -0.077 (leaf)

Val: 0.246 (leaf)

-----------------------------------

A simple, readable, additive tree.

- **Validation**: 5-fold CV, repeated 3x (15 folds total)—Michigan's splits were static; CV catches overfitting in imbalanced deception data. Metrics: Accuracy, precision/recall/F1 on lies (minority class), ROC-AUC.

# FIGS

model = FIGSClassifier(max_rules=5)

cv_scores = cross_val_score(model, X, y, cv=5, scoring='f1_weighted') # Repeat 3x

```

## Results: 80% Lie Recall, Verbal-Only—And Interpretable Rules

Boom: 49/61 lies flagged (>0.5 prob threshold), ~80% accuracy overall (vs. Michigan's 75% verbal, 82% fused). F1 ~0.67 avg across folds (dips to 0.52 in noisy ones, spikes to 0.90+), AUC ~0.65-0.92—volatile but domain-realistic for small N. Humans? Still ~54%. My verbal-only edges their multimodal without the annotation grind.

The rules? Gold for debugging:

- Root: Low Card_GI (sparse numbers) → Cascade to high Sv_GI (state verbs like "shock") + low Polit_2_GI → Val=0.765 (lie: detached internals, no power refs).

- High Quan_GI (hedges) → Val=0.819 (vague filler).

- High Notlw (>1.5 negations) → +0.246 (defensive boost).

Sum 'em: Lies hit >0.5; truths <0.3. E.g., on Kennedy's Chappaquiddick statement (bonus test): Last two segments (fabricated "dives") score 0.765—high Sv ("exhausted"), low Card. Portability win.

Kennedy Statement segmented into 5 pieces and tested with FIGS 5 rules:

figs5raw lie

0.34 0

0.222 0

0.34 0

0.765 1

Metric	Michigan Verbal	Michigan Multimodal	My Verbal-Only (FIGS CV Avg)
Accuracy	~70%	75-82%	80% (lies: 80%)
F1 (Lies)	N/A	~0.78	0.67
Features	100+ (LIWC %)	+ Gestures	5 (GI raw)
Val Method	Train-test	Train-test	5-fold x3 CV

## Why It Works: Raw > Normalized, FIGS > Fusion Bloat

Michigan's fusion is elegant but brittle—LIWC %s smooth out liar verbosity (e.g., rambling quantifiers dilute to 5%, but raw=9 flags hedging). GI's verb/state focus (Sv_GI) nails interpretive leakage better than LIWC's psych buckets; raw counts amplify repetition as effort proxy. FIGS? Its additive sums yield causal-ish rules (e.g., "low specifics + high internals = evasion") without SVM hyperparameters—perfect for forensic explainability (e.g., "This testimony hedges 12x? Prob lie=0.82").

They innovated data; I iterated methods. Ten years on, no CV or raw explorations? Room to build. (Pro tip: Binaries tanked my tests to ~65%—scale 'em next?)

## Next: Gesture Fusion and Beyond

Adding Michigan's gesture binaries to my GI raws—expect 85%? Stay tuned. Fork the repo, run on tobacco trials (UCSF docs library's a deception motherlode), or hit VERBALIE for forensic interviews. Deception detection's ripe for open-source revival—let's crowdsource better benchmarks.

Thoughts? Drop a comment or PR. Code/data: [GitHub link]. #DeceptionDetection #elastictruth #InterpretableML

Tuesday, December 16, 2025 Leave a Comment

Trial Transcript Analysis Of
Susan Neill-Fraser Case by GROK xAI

The Trial of Shadows: Unraveling Lies and Loose Ends
in the Neill-Fraser Case.

The 2010 Supreme Court of Tasmania trial of Susan Neill-Fraser (SNF) for Bob Chappell's murder—without a body, just a sabotaged yacht and vanishing partner—spans 1590 pages of theater-like drama: evasive witnesses, color-coded dinghies, and alibis that deflate like punctured inflatables. Justice Peter Blow presided over a case hinging on circumstantial threads—blood traces, cut ropes, a luminol-glow dinghy—woven into a tapestry of doubt.

Prosecutors painted SNF as a calculating killer; defense cried frame-up via "grey dinghy" ghosts and dodgy cops. No smoking gun, but anomalies abound: fabricated timelines, phantom intruders, and a witness whose tales of chicken-wire plots feel ripped from a bad noir script. Here's the juiciest—lies that scream "consciousness of guilt," hedged denials, and plot holes wide as the Derwent.

#### The Bunnings Bombshell: Alibi from Aisle to Absurdity

SNF's post-2 p.m. Australia Day gap? Filled with a "hours-long browse" at Bunnings hardware—detailed aisles, slow drives, fading light—meant to anchor her home by dusk. But CCTV? Zilch.

Confronted, it shrinks: "Longer than I thought?" No? "Mixed-up days—shocked trauma!" By May interview: Abandoned. Crown's Tim Ellis SC called it a "convocation of lies" (p1419), a deliberate diversion while forensics brewed.

Judge Blow charged jurors: If false, it's guilt's whisper (p1535). Defense? "Honest confusion"—but why invent routes, knots, no purchases? A table of her timeline tango:

Version/Date	Claim	Contradiction/Exposed By	Deceptive Flavor
Jan 27 Statutory Decl. (p61)	Left yacht ~2 p.m., Bunnings till "getting dark" (~8 p.m.), home for calls.	Closes at 6 p.m.; no CCTV (viewed full afternoon, p62-63).	Over-detailed (aisles, speed) for "casual browse"—preemptive filler?
Feb 5 Notes (p758-760)	Arr. 4:20 p.m., parked Brooker side, no trailer, slow driver, home pre-8:30 call.	Trailer on in prior photos; no footage (p761).	Reactive: Shortens stay post-6 p.m. reveal, adds "girl with dark hair" tie-up witness.
Mar 4 Interview (p69)	"Adamant" went; confusion on time, but "must've been me" at yacht 4 p.m. (dinghy sighting).	Retracts: "Mistake—earlier trip" (p1192 trial).	Hedged pivot: "I think... possibly" slips in, buys time amid probes.
May 5 Caution (p1176-77)	No Bunnings; walked home, guilty for leaving Bob sans dinghy.	Jurors hear "strong belief" it happened (p1252).	Emotional deflection: "Trauma fog"—but why persist till footage? Crown: Obstruction (p1441).

This "farrago" (p71) cost police hours sifting tapes—classic misdirection, per CCA later.

#### Dinghy Drama: Grey Ghosts or White Lies?

SNF's white Quicksilver tender (blue trim, ~10 ft) was her "usual" ride—Bob "not nimble," so she hogged it (p874). Left tied at Royal Yacht Club post-~4 p.m. drop-off. But sightings multiply: A "large grey" intruder vessel haunts Four Winds from 3:55 p.m. (Conde: "Battleship grey," pointy bow, scuffed—not hers, p444,1560) to 5 p.m. (Lorraine: "Dark, small," no motor, p1561; P36: "Mid-grey, large," floating stern, p906). Press plea? For "grey with blue trim" (p1051)—SNF's? Police assumed hers, grilling her timings without color reveal (p985-86,1001). Defense: Red herring ignored—why no yacht club dragnet? (p1474). Crown: "Coincidence? Absurd—it's hers, misdescribed" (p1504). Blow: Jury weighs if "another dinghy" fits intruders (p1557-59). Anomalies: Luminol blood-glow inside (p667-74), but no Chappell DNA match—transfer? And Hughes' midnight "female" rower (p391-98): Slow, seated mid-dinghy (not transom), northeast-bound—SNF denies (p1204). Hedging? "Can't remember color/seat" (p396)—convenient fog.

Sighting	Description	Ties to SNF?	Anomaly/Red Flag
3:55 p.m. (Conde, p364)	Large battleship-grey inflatable, midships port, taking water.	Assumed hers; she IDs as white/blue (p822).	Pointy bow, worn—not new Quicksilver (p970); police skip follow-up (p974).
~5 p.m. (Lorraine, p439)	Dark small tender, stern-tied, no motor noticed.	"Whitish cream-yellow" (p940)—differs; faded/old.	Untested: Why no outboard probe? Fits "intruder" better than hers (p832).
~5 p.m. (P36, p1561)	Large mid-grey, tightly inflated, short-rope float.	Sketch stern-placed (p906); no Quicksilver markings.	Anonymous—why no ID hunt? Defense: Sabotage vessel (p1473).
11:30 p.m.-12 a.m. (Hughes, p391)	Inflatable, single female outline, slow outboard NE to yachts.	SNF: "Not me" (p70); mid-seated, color unknown.	Dark—no details; path matches her "usual" route (p68). Guilt hedge?

Police fixated: Told SNF "dinghy like yours" at 3:55—tricking "Must be me" (p821)—despite grey clues (p825). Trail cold; no grey owner traced (p956).

#### Triffett's Twisted Tales: Murder Blueprints or Vendetta?

Enter Phillip Triffett (p539-): Ex-friend, handy-man fixer for SNF's 1990s yacht. Alleges mid-90s plots: First, drown brother Patrick off Electrona (toolbox weights, sink via bilge—eerie *Four Winds* echo, p72).

Then, post-Christmas 1996: "Bob's mean, dangerous—wrap in chicken wire, weigh, sink yacht" (p557-58). Motive? Inheritance feud, Chappell's "stinginess." SNF denies utterly (p1235: "Absolutely untrue"); calls him "fraudulent" post-2009 favor (dropped ammo charges via her plea, p848-59). Cross: Convictions galore (assaults, firearms, p565-67); "disposal expert" rep (bodies? p569).

Defense: Fabricated for leniency—post-Chappell vanishing, he shops tale Jan 28 (p1485). Crown: Prophetic blueprint—yacht kill, body-dump, scuttle (p13-14). Blow: Weigh credibility; no prejudice if relevant to intent (p45). Chilling anomaly: Bilge sabotage mirrors his "plan"—coincidence or confab?

#### Hedging in the Hot Seat: Evasions and "Fog"

Transcript scans for qualifiers faltered (tech glitch?), but cross-exams drip avoidance: SNF's "I think... possibly" on Bunnings routes (p1286); "Can't remember" dinghy seats/colors (p396); Triffett chats? "Don't recall" (p1236). Preemptive? Early statements idealize ("Steady relationship," p35)—crumbles under strain probes. Trial hedging: "Trauma shock" excuses timeline flips (p1535). Ellis: "Lies from fear of truth" (p1534). Gunson: "Honest errors." Cumulative? Cognitive load of invention.

#### Verdict Vibes: Guilt's Echo or Echo Chamber?

Blow's charge (p1540+): Lies (Bunnings/home-alone) signal guilt if not "shock"—but grey dinghy? "Possible intruder vessel" (p1559). Jury convicts Nov 2010. Appeals? Dismissed, but 2025 inquiries probe DNA fraud, police tunnel-vision. Fascinating fracture: SNF's words web a cage—evolving alibis, denied demons—yet loose threads (grey ghost, Triffett's timing) tease "what if?" A yacht of deceit, adrift in doubt. Dive deeper? Full CCA at austlii.edu.au.

This site uses cookies from Google to deliver its services - Click here for information.

ElasticTruth