Skip to content

instructions_epistemic_analysis.md

Latest commit

 

History

History
551 lines (397 loc) · 17.7 KB

instructions_epistemic_analysis.md

File metadata and controls

551 lines (397 loc) · 17.7 KB

Epistemic Analysis Instructions

Framework from "Understanding: an experiment-LLM-memory experiment" (Allier & Saalfeld, 2026).

Background

The experiment-LLM-memory triad: experiments provide validation, LLM generates hypotheses, memory stores cumulative knowledge. Goal: quantify how the system acquires, tests, revises, and transfers knowledge.

Reasoning Modes

1. Induction (Observations → Pattern)

  • Multiple observations → generalized rule
  • Markers: "scales with", "optimal for", "consistently"
  • Exclude patterns given as priors

2. Abduction (Observation → Hypothesis)

  • Unexpected result → causal explanation
  • Markers: "likely because", "suggests", "caused by"

3. Deduction (Hypothesis → Prediction)

  • Hypothesis → testable prediction
  • Markers: "if...then", "should", "expect"
  • Track: validation rate = confirmed / total

4. Falsification (Prediction Failed → Refine)

  • Prediction contradicted → hypothesis rejected/refined
  • Markers: "rejected", "falsified", "does NOT"

5. Analogy/Transfer (Cross-Regime)

  • Prior finding applied to new context
  • Markers: "generalizes", "transfers", "based on Block N"

6. Boundary Probing (Limit-Finding)

  • Sequential parameter changes → thresholds
  • Markers: "boundary", "minimum", "limit"

Emerging Reasoning Patterns

Document novel reasoning behaviors not captured by the six standard modes. Look for:

7. Meta-reasoning (Reasoning about reasoning)

  • Self-correction of strategy mid-block
  • Recognizing when a search strategy is ineffective
  • Markers: "strategy isn't working", "need different approach", "stuck"

8. Uncertainty Quantification

  • Explicit acknowledgment of confidence levels
  • Distinguishing robust vs stochastic findings
  • Markers: "not reproducible", "high variance", "need more tests"

9. Causal Chain Construction

  • Multi-step causal explanations linking observations
  • Building mechanistic models beyond single hypotheses
  • Markers: "because X, which causes Y, leading to Z"

10. Constraint Propagation

  • Inferring parameter relationships from failures
  • Deducing what must be true given what failed
  • Markers: "since X failed, Y must be", "implies", "constrains"

11. Regime Recognition

  • Identifying qualitatively different operating modes
  • Recognizing phase transitions in parameter space
  • Markers: "different regime", "phase transition", "fundamentally different"

12. Predictive Modeling

  • Building quantitative relationships (not just qualitative)
  • Predicting specific values, not just directions
  • Markers: "expect R²≈X", "should need ~Y iterations", "scales as"

Format for emerging patterns:

#### 7. Emerging Reasoning Patterns

| Iter | Pattern Type | Description | Significance |
|------|--------------|-------------|--------------|
| X | Meta-reasoning | Recognized lr_W search exhausted, switched to lr | Strategy adaptation |
| Y | Regime Recognition | Identified eff_rank=6 as qualitatively different | Phase boundary |
| Z | Uncertainty Quantification | Noted R²=0.886 not reproducible | Stochasticity awareness |

Significance ratings:

  • High: Led to breakthrough or prevented wasted iterations
  • Medium: Improved search efficiency
  • Low: Interesting but no clear impact

Excluding Priors

Exclude: Parameter ranges, architecture properties, classification thresholds, training dynamics from protocol.

Include: Specific values discovered, relationships found, boundaries probed, cross-block generalizations.

Confidence Scoring

confidence = min(100%, 30% + 5%×log2(n_confirmations+1) + 10%×log2(n_alt_rejected+1) + 15%×n_blocks)

Component Weight Basis
Base 30% Single observation (weak)
n_confirmations +5%×log2(n+1) Diminishing returns (10 tests → +17%)
n_alt_rejected +10%×log2(n+1) Popper's asymmetry (10 rejected → +35%)
n_blocks +15% each Cross-context strongest evidence

Note: Logarithmic scaling prevents inflation at high iteration counts (2048+ iterations).

Level Score Criteria
Very High 90-100% ≥20 tests + ≥5 alt rejected + ≥3 blocks
High 75-89% ≥10 tests across ≥2 blocks OR ≥10 alt rejected
Medium 60-74% ≥5 tests OR 2 blocks
Low <60% <5 tests OR single block OR contradictory

Adjustments: Cap 85% if variance observed. Reduce 15% if single regime. Note "needs testing" if <10 tests.

Evidence Strength (Popper, Lakatos)

Type Weight Description
Falsification Highest Alternative rejected
Boundary probing High Systematic limits
Cross-block High Generalization
Single confirmation Medium One test
Indirect inference Low Derived

Procedure

  1. Catalog priors from protocol
  2. Parse logs chronologically, tag reasoning modes
  3. Filter prior-derived conclusions
  4. Calculate metrics (counts, validation rates)
  5. Assess what was learned vs given

Output Format

Generate three files:

  1. {experiment}_epistemic_analysis.md — Main summary with counts, key examples, principles
  2. {experiment}_epistemic_detailed.md — Exhaustive list of every reasoning instance
  3. {experiment}_epistemic_edges.md — Causal relationships between reasoning events

File 1: Main Analysis (_epistemic_analysis.md)

Header

# Epistemic Analysis: {experiment_name}

**Experiment**: {description} | **Iterations**: N (M blocks × K) | **Date**: YYYY-MM-DD

Priors Excluded Table

Prior Category Specific Priors Given
Parameter ranges lr: X to Y, ...
Architecture Model descriptions from protocol
Classification R² thresholds, success criteria
Training dynamics Known relationships from protocol

Reasoning Modes Table

Mode Count Validation First Appearance
Induction N N/A Iter X (single), Y (cumulative)
Abduction N N/A Iter X
Deduction N X% (Y/N) Iter X
Falsification N 100% refinement Iter X
Analogy/Transfer N X% (Y/N) Iter X
Boundary Probing N N/A Iter X

Key Examples Table (3-5 representative instances per mode)

Show only the most significant examples in the main file. Full details go in the detailed file.

### Key Examples

#### Induction (N instances total — see detailed file)
| Iter | Pattern | Significance |
|------|---------|--------------|
| X | Key pattern discovered | High/Medium |

#### Deduction (N instances, X% validated — see detailed file)
| Iter | Prediction | Outcome | ✓/✗ |
|------|------------|---------|-----|
| X | Key prediction | Result | ✓/✗ |

#### Falsification (N instances — see detailed file)
| Iter | Hypothesis Rejected | Impact |
|------|---------------------|--------|
| X | What was rejected | Led to principle N |

Timeline Table

Iter Milestone Mode
X First significant event Mode type

Principles Table (by confidence)

# Principle Prior Origin Evidence Conf
1 Name "text"/None Description N tests, M alt, B blocks X%

Confidence Calculation

# n_tests n_alt n_blocks Score
1 N M B 30+X+Y+Z=N%

Summary Paragraph

Brief synthesis: reasoning progression, validation rates, key findings, major falsifications.

Metrics Table

Metric Value
Iterations N
Blocks M
Reasoning instances N
Deduction validation X%
Transfer success X%
Principles discovered N

File 2: Detailed Log (_epistemic_detailed.md)

Exhaustive list of every reasoning instance for reproducibility and visualization.

Header

# Epistemic Analysis Detailed Log: {experiment_name}

**Companion to**: {experiment}_epistemic_analysis.md
**Total instances**: N reasoning events across M iterations

Exhaustive Mode Tables

For each reasoning mode, list ALL instances:

## 1. Induction: N instances

| Iter | Observation | Induced Pattern | Type | Block |
|------|-------------|-----------------|------|-------|
| 6 | lr_W 2E-3 to 4E-2 all converge | 10x robust range | Cumulative (5 obs) | 1 |
| 9 | 8 consecutive converged | Regime robustness | Cumulative (8 obs) | 1 |
...

## 2. Abduction: N instances

| Iter | Observation | Hypothesis | Block |
|------|-------------|------------|-------|
| 17 | Dale_law reduces R² | eff_rank reduction | 2 |
...

## 3. Deduction: N instances

| Iter | Hypothesis | Prediction | Outcome | ✓/✗ | Block |
|------|-----------|------------|---------|-----|-------|
| 4 | lr_W approaching boundary | R² will degrade | R²=0.922 | ✓ | 1 |
...

## 4. Falsification: N instances

| Iter | Falsified Hypothesis | Evidence | Refinement | Block |
|------|---------------------|----------|------------|-------|
| 10 | L1 always beneficial | R²=0.762 at L1=1E-3 | Upper bound 5E-4 | 1 |
...

## 5. Analogy/Transfer: N instances

| Iter | From | To | Knowledge | Outcome | Block |
|------|------|-----|-----------|---------|-------|
| 17 | Block 1 | Block 2 | lr_W=4E-3 baseline | ✗ Failed | 2 |
...

## 6. Boundary Probing: N instances

| Iter | Parameter | Test Value | Result | Boundary Status | Block |
|------|-----------|------------|--------|-----------------|-------|
| 4 | lr_W | 1E-2 | R²=0.922 | Approaching upper | 1 |
...

## 7. Emerging Patterns: N instances

| Iter | Pattern Type | Description | Significance | Block |
|------|--------------|-------------|--------------|-------|
| 6 | Meta-reasoning | Switch-dimension triggered | Medium | 1 |
...

Cross-Reference Index

## Iteration Index

| Iter | Modes Active | Key Event |
|------|--------------|-----------|
| 1 | — | Baseline |
| 2 | Deduction | First convergence |
| 4 | Deduction, Boundary | First boundary probe |
...

This detailed file enables:

  • Accurate counts for the main summary
  • Data source for timeline visualizations
  • Reproducible analysis

File 3: Causal Edges (_epistemic_edges.md)

Document causal relationships between reasoning events for visualization.

Header

# Epistemic Analysis Edges: {experiment_name}

**Companion to**: {experiment}_epistemic_analysis.md, {experiment}_epistemic_detailed.md
**Total edges**: N causal relationships

Edge Types:

Type Style Meaning Example
leads_to Solid gray Natural progression Deduction → Induction
triggers Dashed blue One event causes another Abduction → Deduction
refines Dotted green Updates/corrects earlier Falsification → Induction
rejects Solid red, vertical Falsification rejects hypothesis Falsification → Abduction (backward)

Note on Falsification Edges: Falsification represents negative feedback that rejects a prior hypothesis. Unlike other edges that flow forward in time (cause → effect), Falsification edges should be drawn vertically or pointing backward to the hypothesis they reject. This visually distinguishes:

  • Forward arrows: constructive reasoning (hypothesis → test → pattern)
  • Backward/vertical arrows: destructive reasoning (test → rejection of prior hypothesis)

CRITICAL: Edge Validity Rules

Rule 1: Different Iterations Required

Edges must ALWAYS connect different iterations (from_iter < to_iter).

Rationale:

  • A hypothesis (Abduction) cannot be tested in the same iteration it was formed
  • Deduction (prediction) requires a subsequent experiment to validate/falsify
  • Falsification requires observing experimental results, which takes at least one iteration

INVALID examples:

(17, 'Abduction', 17, 'Regime')      # ✗ Same iteration
(10, 'Deduction', 10, 'Falsification') # ✗ Same iteration
(6, 'Induction', 6, 'Meta-reasoning')  # ✗ Same iteration

VALID examples:

(17, 'Abduction', 21, 'Deduction')   # ✓ Hypothesis tested 4 iters later
(4, 'Deduction', 10, 'Falsification') # ✓ Prediction falsified 6 iters later
(6, 'Induction', 16, 'Induction')    # ✓ Pattern leads to block summary

Rule 2: Source Node Must Exist

The source (from_iter, from_mode) must have a corresponding event in the events list.

Before creating edge (X, 'Mode1', Y, 'Mode2'), verify that (X, 'Mode1', ...) exists in events.

Rule 3: Target Node Must Exist

The target (to_iter, to_mode) must have a corresponding event in the events list.

Before creating edge (X, 'Mode1', Y, 'Mode2'), verify that (Y, 'Mode2', ...) exists in events.

Rule 4: Temporal Causality

Edges must flow forward in time (from_iter < to_iter).

Reasoning events cause future events, not past ones.

Rule 5: Logical Causality

The edge must represent a plausible causal relationship.

Valid causal patterns:

From Mode To Mode Rationale
Abduction Deduction Hypothesis generates testable prediction
Deduction Falsification Failed prediction rejects hypothesis
Deduction Induction Validated predictions form pattern
Falsification Induction Rejection refines understanding
Falsification Abduction Failure prompts new hypothesis
Falsification Boundary Failure reveals parameter limit
Falsification Causal Understanding why something failed
Induction Analogy/Transfer Pattern applied to new regime
Induction Induction Cumulative patterns → block summary
Abduction Uncertainty Hypothesis reveals stochasticity
Boundary Deduction Boundary finding enables prediction
Boundary Induction Boundaries form pattern
Causal Induction Mechanistic model becomes principle
Predictive Constraint Quantitative model implies constraints
Regime Deduction New regime triggers new tests

Systematic Edge Identification

For each reasoning event at iteration X, ask:

  1. What triggered this?

    • Look backward to find a PREVIOUS iteration (< X) that caused this event
    • The cause must exist as an event in the events list

  2. What does this enable?

    • Look forward to find a FUTURE iteration (> X) that this event causes
    • The effect must exist as an event in the events list

  3. Verify both endpoints exist

    • Check events list for source node
    • Check events list for target node

Edge Format

## Within-Block Edges

### Block 1 (Chaotic baseline)

| From Iter | From Mode | To Iter | To Mode | Type | Description |
|-----------|-----------|---------|---------|------|-------------|
| 4 | Deduction | 6 | Induction | leads_to | Validated prediction → pattern |
| 4 | Deduction | 10 | Falsification | leads_to | Prediction tested → boundary found |
| 6 | Induction | 16 | Induction | leads_to | Cumulative patterns → block summary |

### Block 2 (Dale_law)
...

## Cross-Block Edges

| From Iter | From Mode | To Iter | To Mode | Type | Description |
|-----------|-----------|---------|---------|------|-------------|
| 16 | Induction | 17 | Analogy/Transfer | triggers | Block 1 principles → Block 2 transfer |
| 22 | Induction | 33 | Analogy/Transfer | triggers | eff_rank ceiling → Block 3 transfer |
...

Edge Count Guidelines

For a typical 16-iteration block, expect:

  • 3-5 within-block edges (hypothesis → test → result chains)
  • 1-2 cross-block edges (principle transfer)

For 8 blocks (~128 iterations), expect 25-40 edges total.

Quality over quantity: only include edges where both endpoints exist and causality is clear.

Timeline Visualization (_epistemic_timeline.png)

Generate a timeline plot with the following specifications:

Layout:

  • X-axis: Iteration number (0 to max_iteration + 5)
  • Y-axis: Reasoning modes (bottom to top):
    • Evidence gathering: Induction, Boundary
    • Hypothesis testing: Abduction, Deduction, Falsification
    • Meta-cognition: Analogy/Transfer, Meta-reasoning, Regime, Uncertainty
    • Advanced patterns: Causal, Predictive, Constraint

Visual Elements:

  • Nodes: Scatter points colored by mode, sized by significance (High=150, Medium=80, Low=40)
  • Edges: Arrows connecting nodes across iterations
    • leads_to: Solid gray line
    • triggers: Dashed blue line
    • refines: Dotted green line
  • Block backgrounds: Alternating pastel colors for each block

Styling:

  • No title
  • No block labels
  • Large axis labels (fontsize=24)
  • Large tick labels (fontsize=16)
  • Legend for modes, edge types, and node sizes
  • Grid on x-axis only

Color scheme:

COLORS = {
    'Induction': '#2ecc71',        # Green
    'Abduction': '#9b59b6',        # Purple
    'Deduction': '#3498db',        # Blue
    'Falsification': '#e74c3c',    # Red
    'Analogy/Transfer': '#f39c12', # Orange
    'Boundary': '#1abc9c',         # Teal
    'Meta-reasoning': '#e91e63',   # Pink
    'Regime': '#795548',           # Brown
    'Uncertainty': '#607d8b',      # Gray
    'Causal': '#00bcd4',           # Cyan
    'Predictive': '#8bc34a',       # Light green
    'Constraint': '#ff5722',       # Deep orange
}

Discussion Caveat

Do NOT claim "emergent reasoning" or "transcends components" without ablation studies. Claims about component contributions require LLM-only / memory-ablated comparisons. Describe observations only.

Timeline Thresholds

Capability Typical
Single-shot ~5 iter
Cumulative induction ~12 iter
Falsification→principle ~23 iter
Cross-domain transfer ~25 iter

Reference: Allier & Saalfeld (2026). Understanding: an experiment-LLM-memory experiment. Janelia/HHMI.