organvm — Public Process
← Back to dissertation overview

Chapter 3: Methodology

Precision Over Volume Chapter 3 of 7 @4444J99 March 04, 2026
7.8k words 31 min
mcdacareer-pipelinedesign-sciencemixed-methodsportfolio-optimization

CHAPTER 3 | METHODOLOGY

3.1 Research Design and Method

3.1.1 Design Science Research

This study employs a mixed-methods design science research (DSR) approach as formalized by Hevner, March, Park, and Ram (2004). Design science research contributes to the knowledge base through the construction and evaluation of artifacts — systems, models, methods, or instantiations — that address important organizational problems. Unlike behavioral science, which seeks to discover truth through the observation of existing phenomena, design science seeks to create value through the design of novel artifacts that extend the boundaries of human capability.

Hevner et al. (2004) establish seven guidelines for rigorous design science research:

  1. Design as an artifact: The research must produce a viable artifact in the form of a construct, model, method, or instantiation. The precision pipeline satisfies this requirement as a production-deployed software system comprising 30+ Python scripts, 15,000+ lines of code, and 1,554 automated tests.

  2. Problem relevance: The research must address an important and relevant problem. The application volume crisis documented in Chapter 2 — with cold application conversion rates below 2%, 51,330 tech layoffs in Q1 2026, and no existing system combining MCDA with network analysis and portfolio optimization — establishes problem relevance.

  3. Design evaluation: The artifact must be rigorously evaluated. This study evaluates the pipeline through three methods: formal mathematical proof (Sections 3.4–3.8 and Chapter 4), systematic competitive analysis (Chapter 4, Section 4.3), and empirical pipeline data analysis (Chapter 4, Section 4.4).

  4. Research contributions: The research must provide clear contributions. Four specific contributions are identified: integrated MCDA for applicant-side decisions, network-MCDA unification, Kelly criterion application to career strategy, and closed-loop outcome learning.

  5. Research rigor: Rigorous methods must be applied in both construction and evaluation. Mathematical proofs use established theorems; competitive analysis follows a systematic 12-dimension taxonomy; empirical analysis uses era-separated cohort comparison.

  6. Design as a search process: The design process is iterative. The v1-to-v2 evolution documents the iterative refinement cycle: v1 tested the volume hypothesis, which failed empirically (60 cold applications, 0 interviews), motivating the v2 precision redesign.

  7. Communication of research: The research must be communicated to both technology-oriented and management-oriented audiences. This thesis serves both audiences through formal mathematical treatment (technology) and practical interpretation (management).

3.1.2 Mixed-Methods Integration

The study integrates three analytical methods:

Method 1: Formal Mathematical Analysis. Each major component of the pipeline is formalized using established mathematical frameworks and evaluated for optimality properties. The WSM scoring engine is analyzed under the Keeney-Raiffa additive value axioms. The Kelly criterion is applied to derive optimal effort allocation. The pipeline state machine is modeled as an absorbing Markov chain. Reachability analysis is formulated as sensitivity analysis on the network proximity dimension. Time decay is analyzed against exponential, step-function, and power-law models.

Method 2: Systematic Competitive Evaluation. A 12-dimension capability taxonomy, derived from the theoretical framework (Chapter 2, Section 2.12), is used to evaluate 60+ existing products, platforms, and academic prototypes. Each product is scored on a binary (0/1) or partial-credit (0/0.5/1) scale for each dimension. The methodology follows the structured literature review protocol of Kitchenham and Charters (2007), adapted for commercial product evaluation.

Method 3: Empirical Pipeline Data Analysis. The production pipeline’s 1,000+ entries are analyzed using era-separated cohort comparison. Entries submitted before the precision pivot date (March 4, 2026) constitute the “volume era” cohort; entries submitted after constitute the “precision era” cohort. Metrics compared include composite score distributions, score component analysis, pipeline velocity (time in each state), and (where available) conversion outcomes.

3.1.3 Methodological Justification

The reliance on mathematical proof and competitive analysis rather than purely empirical outcome comparison is a deliberate methodological choice driven by two constraints:

First, the precision era began on March 4, 2026. At the time of writing, insufficient post-pivot outcome data has accumulated for statistically powered empirical comparison. Mathematical proof provides an alternative form of validation that is independent of sample size: if a theorem holds, it holds for all instances, not just the observed sample.

Second, ethical and practical constraints prevent a controlled experiment. The applicant cannot simultaneously pursue the same opportunity using both precision and volume strategies (the treatment and control conditions are mutually exclusive at the individual level). The pre-post design (volume era vs. precision era) is the strongest available quasi-experimental design given these constraints, but its statistical power depends on the number of post-pivot outcomes. Mathematical proof bridges this gap by establishing optimality properties that hold regardless of empirical sample size.

3.2 Population and Sample

3.2.1 Pipeline Population

The pipeline population consists of all entries created in the precision pipeline system from initial deployment (January 2026) through the data collection cutoff (March 31, 2026). This population encompasses:

3.2.2 Track Distribution

Entries span 9 application tracks:

Track Description Expected Conversion Weight Vector
job Technology employment 2–30% (channel-dependent) WEIGHTS_JOB (network = 0.20)
grant Creative/research funding 2–15% (program-dependent) WEIGHTS (network = 0.12)
residency Artist/researcher residency 3–10% WEIGHTS
fellowship Academic/professional fellowship 5–20% WEIGHTS
writing Writing commissions/publications 5–25% WEIGHTS
prize Awards and competitions 2–8% WEIGHTS
consulting Freelance/consulting engagements 10–40% WEIGHTS
program Academic programs/cohorts 5–20% WEIGHTS
emergency Emergency financial support 10–30% WEIGHTS

The dual weight vector system (WEIGHTS and WEIGHTS_JOB) reflects the differential importance of network proximity across tracks. For the job track, where the referral multiplier is most extensively documented (8x, per ERIN, 2024), network proximity receives a weight of 0.20 (the highest single weight after mission_alignment at 0.25). For creative tracks, where network effects are less dominant (panel review processes introduce substantial randomness, per Pier et al., 2018), network proximity receives 0.12.

3.2.3 Era Separation

The precision pivot date (March 4, 2026) serves as the natural boundary for cohort separation:

Volume era (pre-pivot): Entries submitted before March 4, 2026, processed under v1 rules: lower qualification threshold, no explicit network analysis, no reachability analysis, daily standup emphasizing throughput and volume pressure, shorter stale thresholds (7 days).

Precision era (post-pivot): Entries submitted on or after March 4, 2026, processed under v2 rules: 9.0 qualification threshold, network proximity as a first-class scoring dimension, reachability analysis, mode-based governance, relaxed stale thresholds (14 days), daily workflow emphasizing research and relationship cultivation.

The get_entry_era() function in pipeline_lib.py operationalizes this separation by classifying entries based on their timeline.submitted date relative to the pivot date constant PRECISION_PIVOT_DATE = "2026-03-04".

3.2.4 Sampling Considerations

This study analyzes a census rather than a sample: all pipeline entries are included in the analysis. No sampling was necessary because the entire population is available and computationally tractable. However, the population itself is a convenience sample of opportunities visible to and selected by a single applicant, and the findings are therefore subject to the limitations discussed in Section 3.15.

3.3 System Architecture: v1 versus v2

3.3.1 v1 Architecture (Volume-Optimized)

The v1 system, deployed from January through early March 2026, implemented a volume-optimized tracking architecture with the following characteristics:

3.3.2 v2 Architecture (Precision-Optimized)

The v2 system, deployed beginning March 4, 2026, implements a precision-optimized decision engine with the following architectural innovations:

Table 3.1: v1 versus v2 Feature Comparison

Dimension v1 (Volume) v2 (Precision)
Scoring dimensions 8 9 (+ network_proximity)
Weight vectors 1 (uniform) 2 (creative, job)
Network analysis None 6-signal time-decayed proximity
Qualification threshold 5.0–5.5 9.0 (mode-adjustable)
Max active entries Unlimited 10
Max weekly submissions Unlimited 1–2
Stale threshold 7 days 14 days
Stagnant threshold 14 days 30 days
Outcome learning None Bayesian (70/30 blend)
Relationship cultivation None cultivate.py workflow
Reachability analysis None Per-entry gap analysis
Mode governance None Precision/Volume/Hybrid
Research pool Flat directory Separated (research_pool/)
Identity positions None 5 canonical positions
Content architecture Flat blocks Storefront/Cathedral tiers
Time decay None Step-function (30/90/180d)
Company cap None 1 per organization
Daily workflow 5hr output-focused 2hr research, 2hr relationships, 1hr application

3.3.3 Architectural Rationale

The v2 architecture was not designed ab initio from theoretical principles; it evolved iteratively from v1 in response to empirical failure. The inciting event was the submission of 60 cold applications in four days under v1 rules, which produced zero interviews. This outcome, documented in the system’s own conversion logs, triggered a systematic review of the theoretical literature (Chapter 2) and the redesign described in this chapter.

Each v2 feature can be traced to a specific theoretical justification:

v2 Feature Theoretical Basis
9-dimension scoring MCDA (Fishburn, 1967; Dawes, 1979)
Dual weight vectors Domain-specific weight optimization
Network proximity scoring Granovetter (1973), Lin (2001), Burt (1992)
9.0 qualification threshold McCall reservation wage (1970)
Max 10 active entries Kelly criterion bet sizing (Kelly, 1956)
Bayesian outcome learning Bayesian updating (Berger, 1985)
Time-decayed signals Burt (2000), Ebbinghaus (1885)
Multi-track portfolio Markowitz diversification (1952)
Storefront/Cathedral content Shannon channel capacity (1948)
Relationship cultivation Cialdini reciprocity (2006)

3.4 The Weighted Sum Model Scoring Engine

3.4.1 Formal Specification

The pipeline’s scoring engine implements the Weighted Sum Model as defined in Chapter 2, Section 2.4.2. For a pipeline entry a, the composite score V(a) is computed as:

V(a) = Sum_{i=1}^{9}(w_i * s_i(a))

where:

Proof of boundedness. Since each s_i is in [1, 10] and each w_i > 0 with Sum(w_i) = 1:

V_min = Sum(w_i * 1) = Sum(w_i) = 1.0 V_max = Sum(w_i * 10) = 10 * Sum(w_i) = 10.0

Therefore V(a) is in [1.0, 10.0] for all entries a and all weight configurations satisfying the constraints. This boundedness is preserved regardless of the specific weight values, providing a stable, interpretable score range.

3.4.2 The Nine Dimensions

Each scoring dimension is computed from data fields in the pipeline YAML entry. The computation is deterministic: identical entry data always produces identical dimension scores.

Dimension 1: mission_alignment (w_creative = 0.25, w_job = 0.25). Quantifies the degree of alignment between the applicant’s expertise and the target’s stated requirements. Computed from keyword overlap between the entry’s requirement fields and the applicant’s canonical competency tags. Score range: 1 (no alignment) to 10 (complete alignment).

Dimension 2: evidence_match (w_creative = 0.20, w_job = 0.20). Quantifies the degree to which the applicant can provide direct evidence (work samples, portfolio items, case studies) addressing the target’s requirements. Computed from the presence and quality of matched materials in the submission block library.

Dimension 3: track_record_fit (w_creative = 0.15, w_job = 0.15). Quantifies the relevance of the applicant’s prior experience to the target role. Computed from a combination of signals including organization type match, industry alignment, seniority level fit, and skill overlap. The dimension is further decomposed into four sub-signals: relevance breadth, depth match, position seniority alignment, and differentiator presence.

Dimension 4: network_proximity (w_creative = 0.12, w_job = 0.20). Quantifies the strength and freshness of the applicant’s relationship to the target organization. This dimension is described in full detail in Section 3.5.

Dimension 5: strategic_value (w_creative = 0.10, w_job = 0.10). Quantifies the long-term strategic value of the opportunity beyond immediate conversion. Considers factors including skill development, portfolio expansion, network expansion, and positioning for future opportunities.

Dimension 6: financial_alignment (w_creative = 0.08, w_job = 0.05). Quantifies the degree to which the opportunity’s compensation aligns with the applicant’s financial requirements. Includes benefits cliff awareness (SNAP limit at $20,352, Medicaid at $21,597, Essential Plan at $39,125) to flag opportunities whose compensation would trigger loss of public benefits.

Dimension 7: effort_to_value (w_creative = 0.05, w_job = 0.03). Quantifies the ratio of expected effort (application preparation, portal completion, follow-up) to expected value (conversion probability * outcome value). Higher scores indicate higher value per unit of effort.

Dimension 8: deadline_feasibility (w_creative = 0.03, w_job = 0.01). Quantifies the feasibility of meeting the application deadline given current workload and preparation status. Computed from the deadline.date field relative to the current date, with consideration of effort classification and current active entry count.

Dimension 9: portal_friction (w_creative = 0.02, w_job = 0.01). Quantifies the submission friction of the target’s application portal. Portal friction scores are loaded from strategy/market-intelligence-2026.json, which assigns friction ratings to known portals (Greenhouse, Lever, Ashby, Workday, iCIMS, etc.) based on field count, required document types, and manual entry requirements. Lower friction portals receive higher scores (friction is a cost dimension, inverted).

3.4.3 Weight Configuration

The weight vectors are defined in strategy/scoring-rubric.yaml:

Creative/Grant/Residency weights (WEIGHTS):

mission_alignment:   0.25
evidence_match:      0.20
track_record_fit:    0.15
network_proximity:   0.12
strategic_value:     0.10
financial_alignment: 0.08
effort_to_value:     0.05
deadline_feasibility: 0.03
portal_friction:     0.02
Sum:                 1.00

Job track weights (WEIGHTS_JOB):

mission_alignment:   0.25
evidence_match:      0.20
network_proximity:   0.20
track_record_fit:    0.15
strategic_value:     0.10
financial_alignment: 0.05
effort_to_value:     0.03
deadline_feasibility: 0.01
portal_friction:     0.01
Sum:                 1.00

The key difference between weight vectors is the network_proximity weight: 0.12 for creative tracks versus 0.20 for the job track. This reflects the differential referral multiplier effect documented in Chapter 2 (Section 2.5): in the job market, referrals produce an 8x hiring rate advantage (ERIN, 2024), while in creative funding, panel review processes introduce substantial reviewer variance (Pier et al., 2018), diluting the network effect.

3.4.4 Axiom Verification for the Career Application Domain

The WSM is provably optimal when four axioms hold (Keeney & Raiffa, 1976). This section verifies each axiom for the career application domain:

Axiom 1: Commensurability. All dimensions must be measured on a common scale. The pipeline scores all dimensions on the same integer scale [1, 10]. This satisfies commensurability by construction — no normalization is required because the measurement scale is uniform.

Axiom 2: Mutual Preferential Independence (MPI). The preference ordering over any subset of dimensions must be independent of the levels of the remaining dimensions. For the pipeline’s 9 dimensions:

MPI is a reasonable assumption for the pipeline’s dimensions because they measure genuinely distinct aspects of the application opportunity. The potential exception is the interaction between network_proximity and effort_to_value (higher network proximity may reduce effort through referral streamlining), but this interaction is captured within the effort_to_value dimension’s computation rather than requiring a multiplicative interaction term.

Axiom 3: Additivity. The value function must be additive (no interaction effects). As argued above, the dimensions are designed to measure independent aspects of fit, and any interactions are captured within individual dimension computations rather than between dimensions.

Axiom 4: Monotonicity. Preference must be monotonic in each criterion. All 9 dimensions are “benefit” criteria: more mission_alignment, more evidence_match, more network_proximity, etc. are always preferred. Portal_friction is the most subtle: the raw construct (friction) is a cost, but the dimension score inverts the scale so that higher scores indicate lower friction, maintaining monotonicity.

3.4.5 The Qualification Threshold

The auto-qualify threshold of 9.0/10 functions as McCall’s (1970) reservation wage in the scoring domain. The threshold defines the minimum composite score below which opportunities are not pursued, regardless of other factors.

Threshold calibration. The 9.0 threshold is set at the 89th percentile of the feasible scoring range [1.0, 10.0]. This reflects the precision strategy’s emphasis on pursuing only the highest-fit opportunities. The threshold is mode-adjustable:

Mode auto_qualify_min max_active max_weekly_submissions
Precision 9.0 10 2
Volume 7.0 30 10
Hybrid 8.0 15 5

The mode system operationalizes the McCall comparative statics: when financial pressure requires faster conversion, the threshold is lowered (volume mode); when runway permits selectivity, the threshold is raised (precision mode).

Cold application impossibility. Under the job track weight configuration, a cold-network entry (network_proximity = 1) cannot reach 9.0 even with perfect scores on all other dimensions:

V_max_cold = 0.25(10) + 0.20(10) + 0.20(1) + 0.15(10) + 0.10(10) + 0.05(10) + 0.03(10) + 0.01(10) + 0.01(10) = 2.5 + 2.0 + 0.2 + 1.5 + 1.0 + 0.5 + 0.3 + 0.1 + 0.1 = 8.2

Since 8.2 < 9.0, it is mathematically impossible for a cold-network job-track entry to reach the qualification threshold. This is not a bug; it is a designed structural enforcement: the weight configuration requires network cultivation as a prerequisite for job-track qualification.

For creative tracks (network_proximity weight = 0.12):

V_max_cold_creative = 0.25(10) + 0.20(10) + 0.15(10) + 0.12(1) + 0.10(10) + 0.08(10) + 0.05(10) + 0.03(10) + 0.02(10) = 2.5 + 2.0 + 1.5 + 0.12 + 1.0 + 0.8 + 0.5 + 0.3 + 0.2 = 8.92

At 8.92, cold-network creative entries barely fall below the 9.0 threshold even with perfect scores on all other dimensions. This creates a softer constraint for creative tracks, where panel-based review introduces enough randomness to justify occasional cold submissions for exceptionally well-matched opportunities.

3.5 Network Proximity Scoring with Time Decay

3.5.1 The Network Proximity Dimension

Network proximity is the fourth scoring dimension and the most architecturally significant innovation of the v2 system. It quantifies the strength and freshness of the applicant’s relationship to the target organization, operationalizing the social capital theories of Granovetter (1973), Lin (2001), and Burt (1992) into a numerical scoring dimension.

3.5.2 Ordinal Scale

Network proximity is measured on a 5-level ordinal scale mapped to integer scores:

Level Score Description Empirical Basis
Cold 1 No relationship to anyone at the organization Baseline: 2% conversion (Indeed, 2025)
Acquaintance 4 Shared community, event, or platform connection Weak tie: ~5% conversion (Granovetter, 1973)
Warm 7 Active communication, mutual connection with introduction Moderate weak tie: ~15% conversion (Rajkumar et al., 2022)
Strong 9 Direct referral from current employee or champion Referral: 28–30% hire rate (ERIN, 2024)
Internal 10 Direct relationship with hiring manager Internal: 40–60% conversion (SHRM, 2024)

The non-linear spacing (1, 4, 7, 9, 10) reflects the non-linear relationship between tie strength and conversion probability. The gap between cold (1) and acquaintance (4) is larger than the gap between strong (9) and internal (10), consistent with Rajkumar et al.’s (2022) finding of an inverted U-shaped relationship between tie strength and job transmission probability: the largest marginal gains come from moving from cold to weak-tie connections, with diminishing returns at higher tie strengths.

3.5.3 Six-Signal Aggregation

The network proximity score is computed as the maximum across six independent signals, each of which can independently establish a minimum score:

Signal 1: Explicit relationship_strength. The entry’s network.relationship_strength field provides the base score via the ordinal map above. This is the primary signal, set through manual assessment or automated hydration (enrich.py --network).

Signal 2: Referral channel. If conversion.channel == "referral", the minimum score is 8, reflecting a known referral pathway regardless of the explicit relationship_strength setting.

Signal 3: Follow-up responses (time-decayed). If any follow_up[] entry has response in ("replied", "referred"), the minimum score depends on the freshness of the response:

Response Age Minimum Score Tier
<= 30 days 7 Fresh
<= 90 days 5 Aging
<= 180 days 3 Stale
> 180 days 0 (no boost) Expired
No date 7 (benefit of doubt) Legacy

This time decay reflects Burt’s (2000) observation that professional network value decays substantially within 3–6 months of last contact. The “benefit of doubt” rule for entries without dates preserves backward compatibility with legacy data created before the time decay system was implemented.

Signal 4: Mutual connections. If network.mutual_connections >= 5, the minimum score is 5. This reflects Lin’s (2001) social capital theory: a concentration of mutual connections at a single organization indicates embedded social resources.

Signal 5: Outreach actions (time-decayed). Completed outreach actions (outreach[].status == "done") contribute to the score if they are recent (within 60 days):

Signal 6: Organizational density. The number of other pipeline entries at the same organization provides an indirect measure of familiarity:

The max operator aggregation means that the highest signal wins. This is a deliberate design choice: a single strong signal (e.g., a referral channel) should not be diluted by the absence of other signals. An entry with a direct referral (Signal 2: min 8) but no mutual connections (Signal 4: no boost) receives a score of 8, not an average of 8 and 1.

3.5.4 Time Decay Model

The time decay model uses a step function with four tiers:

_NETWORK_DECAY = {
    "response_fresh": 30,    # days - full boost
    "response_aging": 90,    # reduced boost
    "response_stale": 180,   # minimal boost
    "outreach_stale": 60,    # done outreach older than this gives no boost
}

Theoretical justification for step-function over exponential. As discussed in Chapter 2 (Section 2.5.5), the exponential decay model f(t) = f_0 * e^{-lambda*t} is the most theoretically principled time decay model. However, the step function is more appropriate for the pipeline’s context for three reasons:

  1. Date resolution. Pipeline entries record dates at day-level granularity, not hour or minute level. The precision of the date data is insufficient to support continuous decay; a step function matches the data resolution.

  2. Behavioral meaningfulness. The tiers correspond to behaviorally meaningful distinctions: “contacted last week” (fresh) is categorically different from “contacted three months ago” (aging), which is categorically different from “contacted six months ago” (stale). The step function captures these categorical distinctions directly.

  3. Implementational transparency. The step function is immediately interpretable by the user: “your contact with this person was 45 days ago, which puts it in the ‘aging’ tier.” Exponential decay would produce a continuous value (e.g., 0.71) that requires additional explanation.

The error introduced by using a step function rather than continuous exponential decay is bounded by the tier width. Within each tier, the maximum error is the difference between the step function value and the exponential value at the tier boundary. For practical purposes, this error is dominated by the much larger uncertainty in the relationship between contact freshness and actual conversion probability.

3.6 Pipeline State Machine as Absorbing Markov Chain

3.6.1 State Machine Definition

The pipeline’s status progression is defined as a finite state machine (FSM) with 10 states and defined transitions. Formally:

Definition. Let M = (S, T, s_0, F) where:

The valid transitions, as defined in pipeline_lib.py:

VALID_TRANSITIONS = {
    "research":     {"qualified", "withdrawn"},
    "qualified":    {"drafting", "staged", "deferred", "withdrawn"},
    "drafting":     {"staged", "qualified", "deferred", "withdrawn"},
    "staged":       {"submitted", "drafting", "deferred", "withdrawn"},
    "deferred":     {"staged", "qualified", "drafting", "withdrawn"},
    "submitted":    {"acknowledged", "interview", "outcome", "withdrawn"},
    "acknowledged": {"interview", "outcome", "withdrawn"},
    "interview":    {"outcome", "withdrawn"},
    "outcome":      {},
    "withdrawn":    {},
}

The state machine enforces forward-only progression through the primary pipeline path (research -> qualified -> drafting -> staged -> submitted -> acknowledged -> interview -> outcome) with lateral transitions to deferred (hold) and backward transitions from drafting/staged/deferred back to qualified (re-evaluation). The terminal states outcome and withdrawn have no valid outgoing transitions.

3.6.2 Absorbing Markov Chain Formulation

The pipeline state machine can be modeled as an absorbing Markov chain — a stochastic process where some states are absorbing (once entered, never left) and from every non-absorbing (transient) state, it is possible to reach at least one absorbing state.

Definition. An absorbing Markov chain is characterized by a transition probability matrix P that can be written in canonical form:

P = Q R
  0 I

where:

Transient states (t = 8): research, qualified, drafting, staged, submitted, acknowledged, interview, deferred

Absorbing states (r = 2): outcome, withdrawn

3.6.3 The Fundamental Matrix

Theorem (Kemeny & Snell, 1960). The fundamental matrix N = (I - Q)^{-1} exists and has the following interpretations:

Theorem. The absorption probability matrix B = NR gives:

These results provide computable answers to practically relevant questions:

3.6.4 Estimation of Transition Probabilities

The transition probabilities in Q and R are estimated from the observed transition frequencies in the pipeline’s data. For each pair of states (s_i, s_j) where T(s_i, s_j) = 1:

p_{ij} = (number of observed transitions from s_i to s_j) / (total transitions from s_i)

At the current stage of the pipeline, some transitions have sparse observations (particularly in the post-pivot precision era). The fundamental matrix analysis is therefore presented as a methodological capability rather than a fully calibrated empirical model. As the pipeline accumulates more transition data, the Q and R matrices will be refined, and the fundamental matrix predictions can be validated against observed absorption patterns.

3.6.5 Practical Applications of the Markov Chain Model

The Markov chain formulation enables several practical analyses:

Expected pipeline length. The row sums of N predict how many state transitions a typical entry undergoes before reaching a terminal state. This informs workload planning: entries that are expected to traverse many states (high row sum) require more sustained attention.

Bottleneck identification. High values in specific columns of N indicate states where entries “pile up” — they are visited many times before progressing. If N_{drafting} is high, it suggests that the drafting -> staged transition is a bottleneck requiring intervention.

Withdrawal versus outcome probability. The B matrix decomposition reveals whether entries from a given starting state are more likely to reach outcome (success or rejection) or withdrawal (applicant-initiated exit). High withdrawal probability from a specific state may indicate that the state’s requirements are too demanding or that the information available at that state frequently reveals deal-breaking issues.

3.7 Reachability Analysis as Sensitivity Analysis

3.7.1 Motivation

Reachability analysis addresses a specific operational question: “For a given pipeline entry, what is the minimum relationship level needed to push the composite score above the qualification threshold?” This question is fundamental to the relationship cultivation workflow, as it determines which entries are worth investing cultivation effort in and how much improvement is needed.

3.7.2 Formal Definition

Definition. For an entry a with current dimension scores s = (s_1, …, s_9), weight vector w = (w_1, …, w_9), and qualification threshold theta, the entry is reachable if there exists a network proximity level l in {cold, acquaintance, warm, strong, internal} such that:

V(a s_4 = l_score) = Sum_{i != 4}(w_i * s_i) + w_4 * l_score >= theta

where l_score is the integer score corresponding to level l (cold=1, acquaintance=4, warm=7, strong=9, internal=10) and dimension 4 is network_proximity.

Computation. The minimum required network proximity score is:

s_4required = (theta - Sum{i != 4}(w_i * s_i)) / w_4

The entry is reachable if s_4_required <= 10 (the maximum network proximity score). The reachable_with level is the minimum ordinal level whose score exceeds s_4_required.

3.7.3 Reachability Scenarios

Table 3.7: Reachability by Network Level and Track

Starting Score (all other dims = 8) Network Required (Job Track, w=0.20) Network Required (Creative, w=0.12)
7.0 (9.0 - 6.4) / 0.20 = 13.0 -> UNREACHABLE (9.0 - 7.04) / 0.12 = 16.3 -> UNREACHABLE
7.5 (9.0 - 6.8) / 0.20 = 11.0 -> UNREACHABLE (9.0 - 7.44) / 0.12 = 13.0 -> UNREACHABLE
8.0 (9.0 - 7.2) / 0.20 = 9.0 -> STRONG (9) (9.0 - 7.84) / 0.12 = 9.7 -> INTERNAL (10)
8.5 (9.0 - 7.6) / 0.20 = 7.0 -> WARM (7) (9.0 - 8.24) / 0.12 = 6.3 -> WARM (7)
9.0 Already above threshold (9.0 - 8.64) / 0.12 = 3.0 -> ACQUAINTANCE (4)

This analysis reveals the practical implication of the weight structure: for job-track entries with moderate scores on other dimensions (8/10), a warm relationship (score = 7) is sufficient to reach the threshold, but a cold application (score = 1) leaves the entry far below qualification. The reachability analysis thus provides specific, actionable guidance for cultivation effort: “Build a warm connection at this organization, and the entry qualifies.”

3.7.4 Reachability as Sensitivity Analysis

Reachability analysis is formally a single-variable sensitivity analysis on the network proximity dimension. It answers the question: “Holding all other dimensions constant, how sensitive is the qualification decision to changes in network proximity?”

The sensitivity of the composite score to network proximity changes is:

dV/ds_4 = w_4

This is constant (the WSM is linear in each dimension), so the sensitivity is directly proportional to the network proximity weight: 0.20 for job track, 0.12 for creative tracks. A one-unit increase in network proximity score changes the composite by 0.20 (job) or 0.12 (creative).

3.8 Bayesian Outcome Learning

3.8.1 Motivation and Architecture

The Bayesian outcome learning system closes the feedback loop between submission outcomes and scoring weights. Without outcome learning, the scoring system operates on expert-assigned weights that may not optimally predict conversion outcomes. With outcome learning, the system converges toward empirically optimal weights as outcome data accumulates.

3.8.2 Implementation

The outcome learning system (outcome_learner.py) operates in four stages:

Stage 1: Data Collection. The collect_outcome_data() function scans closed/ and submitted/ entries for those with outcome fields and pre-outcome dimension scores. Each qualifying entry yields a record containing: entry_id, outcome (accepted/rejected/withdrawn/expired), composite_score, dimension_scores (dict), track, and identity_position.

Stage 2: Dimension Accuracy Analysis. The analyze_dimension_accuracy() function compares average dimension scores for accepted versus rejected entries. For each dimension i:

delta_i = mean(s_i outcome = accepted) - mean(s_i outcome = rejected)

The signal is categorized as:

Stage 3: Weight Recommendation. The compute_weight_recommendations() function adjusts weights by redistributing 0.02 weight units from overweighted to underweighted dimensions. The adjustment is conservative (0.02 per cycle) to prevent oscillation. After adjustment, weights are renormalized to sum to 1.0.

Stage 4: Bayesian Blending. The get_weights() function in score.py blends calibrated weights with base weights:

w_posterior = 0.70 * w_prior + 0.30 * w_calibrated

This 70/30 blend ratio reflects a conservative Bayesian prior: the expert-assigned weights (informed by the literature review) are given 70% influence, while the empirical calibration receives 30%. This ratio ensures that:

  1. The system does not overfit to small samples (when n < 10, calibration is suppressed entirely)
  2. Expert knowledge from the theoretical framework is preserved
  3. The system converges gradually rather than jumping to empirically-derived weights that may reflect noise

After blending, weights are renormalized: w_final = w_posterior / Sum(w_posterior), ensuring they continue to sum to 1.0.

3.8.3 Convergence Properties

Theorem. Under the assumption that the calibrated weights w_calibrated converge to a fixed point w* (the empirically optimal weights) as the sample size increases, the blended weights w_posterior converge to:

w_converged = 0.70 * w_prior + 0.30 * w*

This is not the empirically optimal w* but a blend that incorporates expert judgment. The rationale for not converging to pure w* is that expert-assigned weights encode theoretical knowledge (e.g., the importance of network proximity based on the Granovetter/Rajkumar literature) that may not be fully captured in the pipeline’s own outcome data, which is limited to a single user’s experience.

Safeguards. Three safeguards prevent degenerate weight configurations:

  1. Minimum outcomes: Calibration is suppressed when n < MIN_OUTCOMES_FOR_CALIBRATION (10)
  2. Minimum weight: No dimension weight can fall below 0.01 (ensuring all dimensions contribute)
  3. Normalization: Weights are always renormalized to sum to 1.0 after any adjustment

3.9 The Relationship Cultivation Subsystem

3.9.1 Architecture

The relationship cultivation subsystem (cultivate.py) bridges the gap between scoring and action by translating reachability analysis results into concrete relationship-building tasks. The subsystem imports analyze_reachability() from score.py and identify_referral_candidates() from warm_intro_audit.py.

3.9.2 Candidate Selection

The get_cultivation_candidates() function identifies entries where network improvement could push the composite score above the qualification threshold. The selection criteria are:

  1. Entry has an actionable status (research, qualified, drafting, staged)
  2. Entry’s current composite score is below the threshold (9.0 in precision mode)
  3. Entry is reachable (there exists a network proximity level that would push it above the threshold)
  4. Entry is not deferred or withdrawn

Candidates are sorted by the “gap” — the difference between the current composite and the threshold — with the smallest gaps (closest to qualifying) prioritized, as these require the least cultivation effort to unlock.

3.9.3 Action Suggestion

The suggest_actions() function generates concrete, score-aware cultivation recommendations. For a given candidate, it:

  1. Identifies the current network proximity level
  2. Computes the target level needed to reach the threshold
  3. Translates the target level into specific actions (e.g., “LinkedIn connect would move network 1->4, adding +0.6 pts for creative track, +0.8 pts for job track”)
  4. Estimates the score impact of each action

3.9.4 Theoretical Grounding

The cultivation workflow operationalizes three theoretical principles:

Cialdini’s reciprocity principle: The workflow emphasizes providing value before requesting referrals. Cultivation actions are structured in a sequence: (1) research the contact and their work, (2) provide value (share relevant content, make introductions, comment substantively on their work), (3) establish rapport through non-transactional interaction, (4) make the referral request only after reciprocity debt is established.

Granovetter’s weak ties theory: The workflow targets acquaintance-level connections that can serve as bridges to target organizations, consistent with the finding that moderately weak ties are the most effective for job transmission (Rajkumar et al., 2022).

Burt’s structural holes: The workflow prioritizes building connections that span structural holes between the applicant’s current professional network and target organizations, maximizing the non-redundant information access that Burt (1992, 2004) identifies as the primary benefit of structural hole-spanning connections.

3.10 Mode Switching and Adaptive Governance

3.10.1 Three Operational Modes

The pipeline operates in one of three modes, each defined by a set of threshold parameters:

Parameter Precision Volume Hybrid
auto_qualify_min 9.0 7.0 8.0
max_active 10 30 15
max_weekly_submissions 2 10 5
stale_days 14 7 10
stagnant_days 30 14 21

3.10.2 Theoretical Justification

The mode system operationalizes the McCall (1970) comparative statics (Chapter 2, Section 2.6.2):

Mode selection is governed by market conditions and applicant circumstances. The get_pipeline_mode() function loads the current mode from strategy/market-intelligence-2026.json, which includes a pivot_date, review_date, and revert_trigger specifying the conditions under which mode should be reassessed.

3.10.3 Mode as Constraint, Not Override

A critical design principle is that mode can only raise thresholds, not lower them below the mode’s floor. The _mode_adjusted_threshold() function in agent.py ensures that autonomous pipeline actions never violate mode constraints:

def _mode_adjusted_threshold(base_threshold, mode_thresholds):
    """Mode can only raise, not lower."""
    mode_min = mode_thresholds.get("auto_qualify_min", base_threshold)
    return max(base_threshold, mode_min)

This prevents the agent from lowering standards below what the current mode permits, even if other conditions might suggest relaxation.

3.11 Data Collection Methods and Instruments

3.11.1 Pipeline YAML as Data Instrument

Each pipeline entry is stored as a YAML file containing structured fields that serve as the primary data collection instrument. The schema (pipeline/_schema.yaml) defines required and optional fields across several categories:

Identification fields: id, title, track, status, tags Target fields: organization, role_title, application_url, portal, location Fit assessment: score, dimensions, identity_position, notes Network fields: relationship_strength, mutual_connections, contact_name, contact_role Timeline fields: discovered, qualified, submitted, deadline Submission fields: materials_attached, blocks_used, cover_letter, portal_fields Follow-up fields: follow_up[] (date, method, contact, response, notes) Outreach fields: outreach[] (type, status, date, channel, contact, note) Conversion fields: channel, outcome, outcome_date, stage_reached

3.11.2 Market Intelligence Database

The market intelligence database (strategy/market-intelligence-2026.json) provides time-varying strategic parameters including:

This database is updated manually from 112 sources reviewed quarterly, with a freshness warning system (--staleness flag) that identifies data older than 6 months.

3.11.3 Signal Collection

Conversion and signal data are collected in three structured YAML files:

3.12 Data Analysis Methods

3.12.1 Score Distribution Analysis

Composite scores are analyzed by era (volume vs. precision), track, and identity position. The analysis computes:

3.12.2 Conversion Funnel Analysis

The conversion funnel is analyzed at each pipeline stage:

3.12.3 Competitive Analysis Methodology

Each competitive product is evaluated against the 12-dimension capability taxonomy using a structured assessment protocol:

  1. Product documentation review (website, help center, API documentation)
  2. Free trial or demo evaluation (where available)
  3. Published customer reviews and case studies
  4. Academic publications describing the system’s methodology (for academic prototypes)

Scoring uses a three-level scale: 0 (capability absent), 0.5 (partial implementation), 1.0 (full implementation matching the taxonomy definition). Two independent raters assess each product, with disagreements resolved through discussion.

3.13 Validity and Reliability

3.13.1 Internal Validity

Internal validity — the degree to which observed effects can be attributed to the intervention rather than confounding variables — is addressed through multiple strategies:

Mathematical proof eliminates confounding by establishing results that hold under all conditions satisfying the axioms, not just the observed data. If the WSM boundedness proof holds, it holds for all possible entries, not just those in the current pipeline.

Era-separated cohort comparison controls for temporal confounding to the extent that the volume era and precision era differ primarily in the pipeline strategy applied, not in the external market conditions. However, the pivot date (March 4, 2026) also coincides with evolving market conditions, creating a potential confound that cannot be fully controlled.

Automated scoring eliminates scorer bias: identical entry data always produces identical scores, regardless of when or by whom the scoring is triggered. This addresses the consistency advantage that Dawes (1979) identifies as the primary benefit of linear models.

3.13.2 External Validity

External validity — the degree to which findings generalize beyond the study context — is limited by the single-user deployment (L1), the specific labor market context (L5), and the specific track mix. The mathematical proofs are user-independent (they hold for any applicant facing the same structural problem), providing a form of generalizability that does not depend on the empirical sample. The competitive analysis results are generalizable to the extent that the 12-dimension taxonomy captures the relevant capabilities for career decision support.

3.13.3 Reliability

Reliability — the degree to which results are reproducible — is ensured by:

3.14 Ethical Considerations

3.14.1 Data Privacy

All pipeline data is generated through the author’s own career activities. No third-party personal data is collected, stored, or analyzed. Contact names appearing in pipeline entries are stored locally and are not transmitted to any external service.

No human subjects were studied; the pipeline analyzes the author’s own decision-making process and its outcomes. IRB review is not required for self-study research that does not involve other human participants.

3.14.3 Algorithmic Fairness Considerations

The scoring system is used for self-selection (the applicant deciding which opportunities to pursue), not for evaluating other people. As discussed in Chapter 2 (Section 2.4.6, algorithmic fairness), this substantially changes the fairness calculus: there is no disparate impact concern in the traditional sense because the system scores opportunities, not individuals.

However, structural bias can enter the system through dimensions that reflect existing social capital inequality. The network_proximity dimension inherently advantages applicants with larger or better-positioned professional networks. The system’s design mitigates this concern in two ways: (1) the relationship cultivation workflow provides a mechanism for building network capital rather than merely measuring it, and (2) the mode system allows threshold adjustment when existing network capital is insufficient.

3.14.4 AI-Assisted Research Ethics

Large language models were used as research assistants for literature discovery, citation verification, mathematical exposition, and draft generation. All substantive claims, proofs, and interpretations are the author’s own. The methodology follows the “AI-conductor” model: human direction, AI-assisted generation, human review and editorial control. This approach is documented transparently in the acknowledgments and is consistent with emerging academic standards for AI-assisted research (Nature, 2023; Science, 2023).

3.15 Limitations

3.15.1 Methodological Limitations

L1: Single-user deployment. The pipeline serves one applicant. All empirical results derive from a single user’s career activities. Multi-user validation would strengthen generalizability claims.

L2: Early precision era. Limited post-pivot outcome data constrains empirical validation. The methodological choice to emphasize mathematical proof addresses this limitation.

L3: Absence of controlled experiment. The pre-post quasi-experimental design cannot rule out temporal confounds (market conditions changing simultaneously with the strategy pivot).

L4: Expert-assigned weights. Weights are assigned through expert judgment informed by market research, not derived through formal AHP pairwise comparison. The Bayesian outcome learning system provides a convergence mechanism.

3.15.2 Technical Limitations

L5: Static dimension set. The 9 dimensions are fixed by design. No mechanism exists for automatically discovering new relevant dimensions from data.

L6: Maximum operator for network signals. The max-aggregation of 6 network signals discards information about signal concordance. A weighted combination might better capture the joint effect of multiple weak signals.

L7: Step-function time decay. The step function introduces discontinuities at tier boundaries. Entries one day apart at a tier boundary receive different scores despite having essentially identical freshness.

These limitations are acknowledged not as fatal flaws but as design choices with known trade-offs. Each represents a simplification that prioritizes implementational clarity and interpretability over theoretical purity — a trade-off that Dawes (1979) suggests is the correct one for practical decision support systems.

Related Repositories