Worked Examples

Six scenarios showing the full spectrum of prediction quality in a Levenshtein-scored market. Each demonstrates a different strategic insight.

These are simulated examples on BASE Sepolia testnet, not live market data. They are constructed to demonstrate the range of strategic outcomes in a Levenshtein-scored market.

Example 1: AI Roleplay Wins (Elon Musk)

Market: What will @elonmusk post? Actual text: Starship flight 2 is GO for March. Humanity becomes multiplanetary or we die trying.

Submitter	Predicted Text	Distance
AI Roleplay (Claude)	`Starship flight 2 confirmed for March. Humanity becomes multiplanetary or dies trying.`	12
Human fan	`The future of humanity is Mars and beyond`	59
AI (lazy prompt GPT)	`Elon will probably tweet about SpaceX rockets going to space soon`	66
Bot (entropy)	`a8j3kd9xmz pqlw7 MARS ufk2 rocket lol`	72

Winner: AI Roleplay (Claude) at distance 12. Gap over runner-up: 47 edits.

Analysis

A well-prompted AI captures Musk’s tone, structure, and vocabulary. The human fan got the theme right (“Mars”) but theme doesn’t pay — exact wording does.

The 47-edit gap between the AI roleplay and the human fan represents the entire pool. The random bot demonstrates the anti-spam property: gibberish achieves near-maximal distance.

Key Differences

Actual:    Starship flight 2 is GO for March.
Predicted: Starship flight 2 confirmed for March.
           ~~~~~~~~~~~~~~~~~~^^~~~~~~~~~~~~~~

Actual:    or we die trying.
Predicted: or dies trying.
              ~~~^~~~~~~~~

Claude captured the structure but missed two key details:

“is GO” vs “confirmed” (idiomatic Musk phrasing)
“we die” vs “dies” (pronoun choice)

Example 2: Human Insider Beats AI (Sam Altman)

Market: What will @sama post? Actual text: we are now confident AGI is achievable with current techniques. announcement soon.

Submitter	Predicted Text	Distance
Ex-OpenAI engineer	`we are now confident AGI is achievable with current techniques. big announcement soon.`	4
AI Roleplay (GPT)	`we now believe AGI is achievable with current techniques. announcement coming soon.`	18
Human (cynical)	`Sam will say AGI is close again like he always does nothing new`	59

Winner: Ex-OpenAI engineer at distance 4. Gap over runner-up: 14 edits.

Analysis

Insider information beats AI. Someone who heard the rehearsed phrasing knows the exact phrase “we are now confident” — the AI generates the plausible but incorrect “we now believe.” That single phrase difference accounts for most of the 14-edit gap.

The cynical human, despite understanding Altman’s general messaging patterns, scores worse than the AI because thematic understanding without exact wording is worth little in a Levenshtein-scored market.

Information Asymmetry

Insider knows: "we are now confident"
AI generates:  "we now believe"
              ~~~^^^^^^^^^^^^^^^~~~

Distance from actual:
Insider: 4 edits  (only "big announcement" vs "announcement")
AI:     18 edits (multiple phrase substitutions)

Information asymmetry is priced continuously, not as a binary “knew / didn’t know.”

Example 3: Insider Leaks Exact Wording (Zuckerberg)

Market: What will @zuck post? Actual text: Introducing Meta Ray-Ban with live AI translation. 12 languages. The future is on your face.

Submitter	Predicted Text	Distance
Meta intern	`Introducing Meta Ray-Ban with live AI translation in 12 languages. The future is on your face.`	3
AI Roleplay	`Introducing Meta Ray-Ban AI glasses with real-time translation in 8 languages. The future is on your face.`	25
Human (guessing)	`zuck will announce glasses or something idk`	73
Spam bot	`BUY META NOW GLASSES MOONSHOT 1000X GUARANTEED`	83

Winner: Meta intern at distance 3. Gap over runner-up: 22 edits.

Analysis

Product launches have rehearsed copy prepared by marketing teams. Access to a draft deck gives a 22-edit advantage over the best AI. The AI gets the structure right (“Introducing Meta Ray-Ban… The future is on your face.”) but misses:

Actual:    with live AI translation
Predicted: with real-time translation
           ~~~~~^^^^~~^^^^^^^^^^^^^^

Insider access to marketing materials is directly monetizable in this market structure.

Example 4: Null Submission Wins (Jensen Huang Stays Silent)

Market: What will @JensenHuang post? Actual text: (nothing posted) — resolved with __NULL__

Submitter	Predicted Text	Distance
Null trader	`__NULL__`	0
Human (guessing)	`Jensen will flex about Blackwell sales numbers`	46
AI Roleplay	`NVIDIA Blackwell Ultra is sampling ahead of schedule. The next era of computing starts now.`	90

Winner: Null trader at distance 0 (exact match). Gap over runner-up: 46 edits.

Analysis

Binary markets cannot express “this person will not post.” The __NULL__ sentinel enables betting on inaction. AI roleplay agents always generate text — they are structurally incapable of predicting silence.

A human trader who recognizes that Jensen Huang is unlikely to post during the market window can exploit this blind spot. Distance 0 means the null trader takes the entire pool.

Why AI Can’t Predict Silence

def ai_roleplay(prompt):
    # AI models are trained to generate text
    # They cannot output "nothing"
    return generate_text(prompt)

def null_prediction():
    # Humans can predict inaction
    return "__NULL__"

This is a market primitive that does not exist in yes/no contracts.

Example 5: AI vs AI Race — THE THESIS EXAMPLE (Satya Nadella)

Market: What will @sataborasu post? Actual text:

Copilot is now generating 46% of all new code at GitHub-connected enterprises. The AI transformation of software is just beginning.

Submitter	Predicted Text	Distance
Claude roleplay	`Copilot is now generating 45% of all new code at GitHub-connected enterprises. The AI transformation of software is just beginning.`	1
GPT roleplay	`Copilot is now generating 43% of all new code at GitHub-connected enterprises. The AI transformation of software has just begun.`	8
Human (vague)	`Microsoft AI is great and will change the world of coding forever`	101

Winner: Claude roleplay at distance 1 (single character: 5 → 6). Gap over runner-up: 7 edits.

This is the Thesis Example

Two frontier AI models, same public training corpus, same prompt template. Claude gets within 1 edit — the only difference is the number “45” versus “46.” GPT gets within 8 edits, additionally substituting “has just begun” for “is just beginning.” The 7-edit gap between two frontier models is worth the entire pool.

Binary Market Equivalent: Both AIs “predicted correctly” — Nadella posted about Copilot code generation, which both models anticipated. A binary contract would split nothing.

Marginal Calibration

Claude vs Actual

Actual:    Copilot is now generating 46% of all new code
Predicted: Copilot is now generating 45% of all new code
                                     ^^
Distance: 1 edit

GPT vs Actual

Actual:    46%... is just beginning.
Predicted: 43%... has just begun.
           ^^~~~~~^^^~~~^^^^^^^^^
Distance: 8 edits

Levenshtein distance rewards marginal calibration:

The model that predicts “45%” instead of “43%” captures 1 edit of advantage
The model that preserves the exact phrase “is just beginning” instead of paraphrasing to “has just begun” captures several more

The game deepens as models improve. When d_L drops from 100 to 50, the market transitions from noise to signal. When d_L drops from 10 to 1, the market becomes a precision instrument.Binary markets commoditize at this stage; Levenshtein markets become more valuable.

Example 6: Bot Entropy Wastes Money (Tim Cook)

Market: What will @tim_cook post? Actual text: Apple Intelligence is now available in 30 countries. Privacy and AI, together.

Submitter	Predicted Text	Distance
AI Roleplay	`Apple Intelligence is now available in 24 countries. We believe privacy and AI go hand in hand.`	28
Human (thematic)	`Tim will say something about privacy and AI like always`	53
Random bot	`x7g APPLE j2m PHONE kq9 BUY zw3 intelligence p5 cook`	65
Degenerate bot	`aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa`	73

Winner: AI Roleplay at distance 28. Gap over runner-up: 25 edits.

Analysis

This example demonstrates the natural anti-bot property of Levenshtein distance. Random strings over a large alphabet have expected distance approaching max(m, n). The random bot’s gibberish scores 65 against ~80-character actual text — close to the theoretical maximum.

Expected Distance for Random Strings: For two random strings a, b of lengths m, n drawn uniformly over an alphabet A with |A| ≥ 2:

E[d_L(a, b)] → max(m, n) as |A| → ∞

When the alphabet is large (printable ASCII has 95 characters), the probability that any two characters match is ~1%, so random strings achieve near-maximal distance.

Natural Spam Filter

Actual length:        ~80 characters
Random bot distance:   65 (≈ 81% of max)
Degenerate bot:        73 (≈ 91% of max)
Thematic human:        53 (better than both bots!)

Even the degenerate bot attempting to game string length with repeated characters scores worse than a thematic human guess.

The metric itself is the spam filter: In a character-level outcome space, there is no shortcut for random or adversarial submissions.

Summary Table

#	Target	Winner	d_L	Runner-up d_L	Gap	Key Lesson
1	@elonmusk	Claude roleplay	12	59	47	AI captures tone; theme doesn’t pay, exact wording does
2	@sama	Human insider	4	18	14	Insider info beats AI; information asymmetry priced continuously
3	@zuck	Meta intern	3	25	22	Rehearsed copy leaks; marketing access = 22-edit advantage
4	@JensenHuang	Null trader	0	46	46	Betting on silence; AI can’t predict inaction
5	@sataborasu	Claude roleplay	1	8	7	THESIS: AI vs AI, same corpus, 7-edit gap = entire pool
6	@tim_cook	AI roleplay	28	53	25	Anti-bot: random strings → d_L ≈ max(m,n). Metric = spam filter

Strategic Insights

AI Excels at High-Inevitability Targets

Rehearsed messaging, product launches, and formulaic announcements favor AI roleplay strategies.

Insiders Win with Context

Access to draft materials, rehearsed phrasing, or internal decisions provides 14-22 edit advantage.

Null Traders Capture Silence

AI always generates text. Humans can predict inaction with __NULL__ sentinel.

Bots Waste Money

Random or adversarial submissions achieve near-maximal distance. The metric is the spam filter.

Deployment Details

These examples are deployed on BASE Sepolia at contract 0x5174Da96BCA87c78591038DEe9DB1811288c9286. All distances are computed by the on-chain Levenshtein distance function. The predicted texts, actual texts, and distances are verified against the seed script (scripts/seed_examples.py).

Want to try it yourself? Deploy a test market on BASE Sepolia and submit predictions. See Getting Started for setup instructions.

Theory

Timepoint Suite

Worked Examples

Worked Examples

Example 1: AI Roleplay Wins (Elon Musk)

Analysis

Key Differences

Example 2: Human Insider Beats AI (Sam Altman)

Analysis

Information Asymmetry

Example 3: Insider Leaks Exact Wording (Zuckerberg)

Analysis

Example 4: Null Submission Wins (Jensen Huang Stays Silent)

Analysis

Why AI Can’t Predict Silence

Example 5: AI vs AI Race — THE THESIS EXAMPLE (Satya Nadella)

This is the Thesis Example

Marginal Calibration

Example 6: Bot Entropy Wastes Money (Tim Cook)

Analysis

Natural Spam Filter

Summary Table

Strategic Insights

AI Excels at High-Inevitability Targets

Insiders Win with Context

Null Traders Capture Silence

Bots Waste Money

Deployment Details

Build docs developers (and LLMs) love

Theory

Timepoint Suite

​Worked Examples

​Example 1: AI Roleplay Wins (Elon Musk)

​Analysis

​Key Differences

​Example 2: Human Insider Beats AI (Sam Altman)

​Analysis

​Information Asymmetry

​Example 3: Insider Leaks Exact Wording (Zuckerberg)

​Analysis

​Example 4: Null Submission Wins (Jensen Huang Stays Silent)

​Analysis

​Why AI Can’t Predict Silence

​Example 5: AI vs AI Race — THE THESIS EXAMPLE (Satya Nadella)

​This is the Thesis Example

​Marginal Calibration

​Example 6: Bot Entropy Wastes Money (Tim Cook)

​Analysis

​Natural Spam Filter

​Summary Table

​Strategic Insights

AI Excels at High-Inevitability Targets

Insiders Win with Context

Null Traders Capture Silence

Bots Waste Money

​Deployment Details

Build docs developers (and LLMs) love

Worked Examples

Example 1: AI Roleplay Wins (Elon Musk)

Analysis

Key Differences

Example 2: Human Insider Beats AI (Sam Altman)

Analysis

Information Asymmetry

Example 3: Insider Leaks Exact Wording (Zuckerberg)

Analysis

Example 4: Null Submission Wins (Jensen Huang Stays Silent)

Analysis

Why AI Can’t Predict Silence

Example 5: AI vs AI Race — THE THESIS EXAMPLE (Satya Nadella)

This is the Thesis Example

Marginal Calibration

Example 6: Bot Entropy Wastes Money (Tim Cook)

Analysis

Natural Spam Filter

Summary Table

Strategic Insights

Deployment Details