Experience Extraction

#practical-application #skill-acquisition #information-theory #meta-principle

Experience as Hyperdimensional Ore

Most people read only one two-dimensional slice of their experience—the surface narrative. "Project failed, felt frustrated, moved on." "Meeting went well, felt good, next meeting." "Tried the diet, didn't work, quit." This is like mining gold ore and taking only the surface pebbles, leaving 95% of the metal in discarded tailings.

Experience can be parsed as hyperdimensional ore. Useful dimensional breakdown includes:

• Causal structure: What actually generated what? What were the mechanisms?
• Counterfactual branches: What almost happened? What would have happened if...?
• Pattern class: What category of situation is this? Where have I seen this before?
• Transfer mappings: Where else does this pattern apply? What other domains share this structure?
• Temporal dynamics: How did state evolve over time? What was the trajectory?
• Prediction errors: Where was my model wrong? What surprised me?
• Emotional signature: What did this feel like? Why did it matter?

The default extraction—surface narrative only—discards almost all of this. You pay the full cost of lived experience (time, energy, emotion) but extract perhaps 5% of the available information.

Think of the extraction stack like a prism separating white light into component frequencies. The spectral information is encoded in the photon; the prism makes it readable. Similarly, these dimensions can be extracted from experience with the right tools, even if your default processing doesn't access them automatically.

Will developed this extraction stack through deliberate practice across gym reactivation, work resumption, and content creation domains (N=1 in each). The multiplier estimates come from retrospective comparison of early journals (surface narrative only) vs. later journals (full-stack extraction). Your mileage may vary—test and calibrate.

The Default: Surface Narrative Only

Watch how most people process experience:

Raw experience: Three-month project failed, team demoralized, budget overrun.

Default extraction: "That didn't work. Felt bad. Try something different next time."

Dimensions discarded:

• Causal structure: Why did it fail? Which specific decisions generated which specific failures?
• Counterfactual: What would have worked? At which decision points did alternatives exist?
• Pattern class: Is this a "wrong market" failure or a "wrong execution" failure? Different categories require different responses.
• Transfer: Does this failure pattern appear in my other domains? Relationships? Health?
• Temporal: When did the trajectory become unrecoverable? What were early warning signals?
• Prediction: Where was my initial model wrong? What would I predict differently now?

Same experience. One person extracts 6 bits (surface narrative). Another extracts 300 bits (full-stack extraction). The second person's "three months of experience" is worth 50 times more for model-building and future decision-making.

This framing suggests why "years of experience" is often a noisy hiring signal. Ten years of surface-narrative extraction might yield less usable information than six months of full-stack extraction.

The Extraction Stack

Seven methods for accessing dimensions beyond surface narrative. These aren't additive—they multiply. Using multiple methods simultaneously compounds the extraction.

1. Signal Amplification: Journaling

Journaling converts implicit patterns to explicit representations. What exists only as vague feeling becomes externalized, stable, and analyzable.

Observed pattern: Writing appears to force conscious processing. You can't journal "I feel stuck" for five minutes without unpacking why you feel stuck, what specifically is stuck, what would unstuck mean. The act of articulation reveals structure.

Example:

• Pre-journaling: "I feel unmotivated today"
• Post-journaling (10 minutes): "Unmotivated because activation cost for work is ~6 units (high), I didn't install the bridge sequence this morning, and I'm comparing current state to yesterday's peak productivity (unfair comparison). The actual mechanism is missing launch ritual, not 'low motivation' as character trait."

Dimension accessed: Causal structure, emotional signature

Multiplier: 3x (makes implicit explicit, enables pattern detection across entries)

Actual journal entry (Will, Day 12 of work reactivation):

"Activation cost for work still ~5 units. Mechanism: no launch ritual installed yet. Missing bridge between morning routine and work state. Prediction: installing 5-min review-of-yesterday's-progress as bridge will drop cost to ~2 units by day 15. Test starting tomorrow."

Same experience ("work felt hard today") with extraction vs. without yields radically different signal.

The working memory limit means you can only hold ~7 items in mind. A month of experience contains hundreds of data points. Without externalization, patterns across those points are invisible. The journal holds what working memory can't, enabling pattern detection impossible through introspection alone.

2. Question-as-Query Engineering

Questions are forcing functions that direct attention. A pre-loaded question routes your pattern-matching system during experience, not just after.

Mechanism: Entering an experience with "What is the mechanism here?" pre-loaded causes automatic search for causal structure. Same experience, different query, different dimensions extracted.

Example:

• No question loaded: Meeting happens → "That went okay" → move on
• Question loaded ("What did I predict vs. what happened?"): Meeting happens → automatic comparison of predictions to outcomes → calibration signal extracted

Pre-loaded questions that extract specific dimensions:

• "What is the mechanism?" → Causal structure
• "What would have to be true for X?" → Counterfactuals
• "What class of problem is this?" → Pattern matching
• "Where else have I seen this?" → Transfer mappings
• "What would I predict here, and what actually happens?" → Prediction errors

Dimension accessed: Whichever dimension the question targets

Multiplier: 2x (routes attention during experience, not just after)

The cost is trivial—30 seconds to load a question before entering a situation. The return is automatic extraction of dimensions you'd otherwise miss.

3. Prediction Pre-registration

Writing down what you expect BEFORE the experience, then comparing to actual outcomes. This creates much stronger error signal that defeats most hindsight bias.

Mechanism: Your brain retroactively edits memories to match outcomes. "I knew that would happen" feels true even when you didn't know. Pre-registration creates a stable record that can't be edited. When prediction meets reality, the delta is genuine learning signal.

Example:

• Pre-register: "I predict the 30x30 gym pattern will reduce activation cost from ~6 units to less than 1 unit by day 30"
• Day 30 actual: Activation cost = 0.5 units
• Extraction: Model confirmed, confidence increased, apply to new domains with higher prior

Without pre-registration:

• Day 30: "Yeah, I figured it would get easier" (hindsight)
• No model update because no surprise, even if original prediction was "This will never get easier"

Dimension accessed: Prediction errors, model validity

Multiplier: 4x (strong feedback on model accuracy, defeats most narrative smoothing)

This is what makes tracking actually work for learning. Numbers on a page from last month can't be edited by your brain's narrative machinery.

4. Contrastive Pairing

Controlled comparison between similar situations in your own history, holding most variables constant while one differs.

Mechanism: Same situation, different approach, different outcome → isolates causal variable. This is the experimental method applied to your own experience.

Example:

• Situation A: Gym attempt without Julius forcing function → P(gym) = 0.2 over 30 days
• Situation B: Gym attempt WITH Julius forcing function → P(gym) = 0.95 over 30 days
• All else roughly equal (same schedule, same gym, same exercises)
• Extraction: Julius accountability is causally responsible for 0.75 probability shift

Without contrastive pairing: "I don't know why the gym worked this time and not last time. Guess I just wanted it more." (No causal insight, no transferable principle)

Dimension accessed: Causal structure, counterfactuals

Multiplier: 3x (isolates specific causal variables within your own data)

The longitudinal journal enables this: read good days, read bad days, find differences. What was present in successes that was absent in failures? What was present in failures that was absent in successes?

5. Mental Model as Bootstrap Simulator

Once you have enough data to build a model, use the model to generate synthetic samples—predictions about situations you haven't encountered.

Mechanism: Model trained on lived experience → simulate "what would happen if..." → test simulation against next real sample → update model based on simulation accuracy.

Example:

• Lived data: 30x30 gym pattern reduced activation cost
• Model extracted: "Consistent daily execution reduces activation cost exponentially over ~30 days"
• Simulation: "If I apply this to meditation, cost should drop from ~5 to ~0.5 by day 30"
• Test: Run meditation 30x30, measure actual activation cost trajectory
• Extract: Did model generalize? If yes, model applies to broader class. If no, what was different?

Dimension accessed: Transfer mappings, pattern class

Multiplier: 2x (extends extraction beyond direct experience through simulation)

This is how mental models compound. Each successful transfer increases confidence in the model's domain. Each failed transfer reveals boundaries of applicability.

6. Resolution Matching

Zooming to the resolution level with highest signal-to-noise ratio.

Mechanism: Too coarse and patterns are invisible. Too fine and noise overwhelms signal. Optimal resolution is where pattern becomes visible without drowning in variance.

Example:

• Too coarse: "My month was productive" (no actionable pattern)
• Too fine: "At 3:47pm I felt slightly anxious" (noise, not signal)
• Optimal: "Week 1 P(deep work) = 0.3, Week 2 = 0.6 after installing morning ritual" (clear pattern, actionable)

Signs you're at wrong resolution:

• Everything looks random (too fine—averaging would reveal pattern)
• Everything looks like same category (too coarse—detail would reveal structure)
• No actionable insights emerging (wrong level for intervention)

Dimension accessed: Temporal dynamics, probability distributions

Multiplier: 2x (finds S/N sweet spot)

Tracking systems should be calibrated to the right resolution. Track what varies at the timescale of your decision-making. Daily habits → daily tracking. Weekly patterns → weekly reviews. Quarterly strategy → quarterly assessments.

7. Causal vs. Correlational Extraction

Distinguishing "what happens together" from "what generates what."

Mechanism: Correlation tells you patterns exist. Causation tells you what to change. Same observation, different extraction, wildly different actionability.

Example:

• Correlational: "I exercise and feel energized" (co-occurrence)
• Causal: "Exercise → increased blood flow → metabolic state shift → subjective energy" (mechanism)
• Deeper causal: "Morning exercise specifically produces afternoon energy; evening exercise disrupts sleep" (conditional mechanism)

The shift: From "these things go together" to "this thing generates that thing through this mechanism, contingent on these conditions."

Dimension accessed: Causal structure, mechanistic models

Multiplier: 3x (moves from observation to intervention, enables causal reasoning)

Cybernetic thinking requires causal extraction. You can only control variables that causally influence outcomes. Correlates that aren't causes can't be controlled.

The Multiplier Effect

The extraction methods don't add—they multiply. Using multiple methods on the same experience compounds the extraction.

Full stack extraction: 864x information yield from same lived experience.

This is obviously an idealization—not every method applies to every experience, and the multipliers aren't precise. But the principle holds: extraction methods compound. Someone using half the stack extracts an order of magnitude more than someone using default processing.

Fast Experience vs. Slow Experience

Not all experience loops are equal. Different loop speeds serve different learning functions.

Fast Experience (Micro-loops)

High frequency, tight feedback, small variance.

Examples:

• Daily habit tracking (365 samples/year)
• Per-rep video analysis in skill training
• Test-driven development cycles (minutes per iteration)
• Daily weight measurements

Advantages:

• Rapid pattern detection (enough samples to cross noise floor quickly)
• Quick model updates (short feedback delay)
• High statistical power (many samples reduce variance)

Limitations:

• Narrow context (tests only local conditions)
• May miss macro-patterns (trees, not forest)
• Overhead per sample limits total sample count

Slow Experience (Macro-loops)

Low frequency, delayed feedback, high variance.

Examples:

• Career changes (3-5 in lifetime)
• Major relationship transitions
• Business strategy pivots (years between attempts)
• Life phase transitions

Advantages:

• Tests large interventions unavailable at micro level
• Reveals long-term dynamics invisible in short windows
• High stakes → high motivation → high attention

Limitations:

• Too few samples for statistical confidence (N=3 career changes)
• Long feedback delay → hard to attribute outcomes
• Confounds accumulate (many variables change together)

The Synthesis: Nested Loops

Optimal extraction nests fast loops inside slow loops.

Run 30-day experiments (fast) within 5-year career arcs (slow). Extract patterns from micro-loops, use them to inform macro-decisions. Don't wait for the slow loop to complete—you'll be dead before N is large enough. Instead, extract maximum signal from fast loops and apply to slow-loop decisions.

Example:

• Slow loop: "Which career path?" (N=3 in lifetime, not enough data)
• Fast loops nested within: 30-day experiments in different domains, tracking daily engagement, accumulating signal about fit
• Extraction: Fast-loop data (high N, tight feedback) informs slow-loop decision (low N, delayed feedback)

The person who runs 20 fast-loop experiments in a year extracts more career-relevant signal than someone who waits 10 years for their slow-loop career outcome to manifest.

Compressing Years

"10 years of experience" could mean:

• 10 years × 1x extraction = 10x cumulative information
• 1 year × 72x extraction (full stack) = 72x cumulative information

The second person has 7x more extracted information in 1/10th the time.

This is why "years of experience" is such a noisy hiring signal. Time elapsed measures ore encountered, not metal extracted. Someone with deliberate extraction practice for 2 years may outperform someone with 10 years of passive accumulation.

What actually matters:

• Loop architecture (how many feedback cycles, how tight)
• Extraction method (how much per cycle)
• Sample quality (high-information experiences vs. repetitive ones)

Useful mental model for learning rate:

Learning_rate ∝ (Samples/time) × (Extraction_efficiency) × (Sample_quality)

This isn't literal mathematics—it's a useful frame for thinking about what drives learning efficiency. The insight: learning rate is controllable through extraction practice, sample architecture, and experience quality. Not passive accumulation over time.

You can accelerate learning by:

1. Increasing loop frequency (more samples per unit time)
2. Improving extraction (more bits per sample)
3. Seeking high-information experiences (better samples)

All three are controllable. "Years of experience" treats learning as passive accumulation. Extraction-aware learning treats it as an optimizable process.

Common Extraction Failures

Over-extraction paralysis: Spending 2 hours extracting from 10-minute experience. Extraction has diminishing returns—stop when marginal insight drops.

Analysis instead of action: Journaling about gym instead of going to gym. Extraction serves action, doesn't replace it.

Wrong resolution: Tracking minutiae (anxiety at 3:47pm) instead of patterns (weekly P(deep work) trending up/down).

No action on extracted signal: Identifying patterns but not designing interventions. Extraction without application is entertainment, not optimization.

Building Your Extraction Practice

Start simple. Layer methods as each becomes automatic.

Week 1-2: Journaling baseline

• 10 minutes daily, unstructured
• Goal: Establish externalization habit
• Multiplier achieved: 3x

Week 3-4: Add pre-loaded questions

• Before meetings/experiences, load one question
• Start with: "What is the mechanism here?"
• Multiplier achieved: 6x

Week 5-6: Add prediction pre-registration

• Before experiments, write down predictions with numbers
• Compare predictions to outcomes, log deltas
• Multiplier achieved: 24x

Week 7-8: Add contrastive pairing

• Weekly review: Compare this week's successes to failures
• Monthly review: Compare this month to previous months
• Multiplier achieved: 72x

Ongoing: Layer remaining methods

• Mental model simulation: Before new domains, predict based on existing models
• Resolution matching: Adjust tracking granularity to match decision timescale
• Causal extraction: For every correlation, ask "what's the mechanism?"

Tracking your extraction:

• Not just "did I journal?" but "what dimensions did I extract?"
• After each experience: "What do I know now that I didn't before?"
• Weekly: "What patterns emerged this week?"

Related Concepts

• Journaling - Signal amplification through externalization, converting implicit to explicit
• Tracking - Measuring probability distributions vs. trusting memory, resolution matching
• Question Theory - Pre-loaded questions as attention routing and forcing functions
• Skill Acquisition - Extraction efficiency determines learning rate, transfer mappings
• Information Theory - Signal-to-noise ratio, entropy reduction through structured extraction
• Predictive Coding - Prediction errors as primary learning signal
• Cybernetics - Feedback cycle time and controllable variables
• Optimal Foraging Theory - Exploration vs. exploitation in experience sampling
• Causality Programming - Causal inference and mechanism identification
• Modeling - Building mental models as bootstrap simulators
• 30x30 Pattern - Fast loops for rapid pattern detection and model building
• Working Memory - Why externalization is necessary for pattern detection
• Signal Boosting - Amplification strategies for crossing detection thresholds

Experience is ore containing multiple dimensions simultaneously. Most people mine only the surface. Build extraction tools to access the 95% that defaults discard. The dimensions were always there—the stack just makes them readable.