Module shares 

Decayed-knowledge materialization: knowledge_shares and doe_scores temporary tables consumed by every WS-B analysis.

Knowledge share model

Each developer’s knowledge of a file decays exponentially with time since their last contribution. The decay constant (220 days, halving ≈ 5 months) comes from Jabrayilzade et al., ICSE-SEIP 2022 (arXiv 2202.01523 §3.1).

Contribution weight is scaled by AI attribution:

• Human commit → weight 1.0
• AI-assisted commit → weight 0.7 (heuristic motivated by arXiv 2507.08160, which shows GenAI-heavy commits distort expertise models)
• AI-authored commit → weight 0.3 (same source)

Reviewer credit: commits carrying Co-Authored-By: or Reviewed-By: trailers award each reviewer W_REVIEWER (0.5) of the author weight per Jabrayilzade 2022 (reviewers = ½ author weight) and Rigby & Bird, ESEC/FSE 2013 (review transfers 66–150% of authorship knowledge). Commits touching more than 10 files are excluded from reviewer credit (Rigby & Bird’s >10-file exclusion rule: large sweeping commits do not meaningfully transfer file-level knowledge to reviewers).

DOE (Degree of Expertise) model

DOE per author×file is computed from the linear model by Cury & Avelino, SBES’24 (arXiv 2408.08733):

doe = 5.28223
    + 0.23173 × ln(1 + adds)
    + 0.36151 × fa
    − 0.19421 × ln(1 + num_days)
    − 0.28761 × ln(size.max(1))

Where:

• adds = lifetime SUM(loc_added) by this author for this path
• fa = 1.0 if this author created the file (change_type = 'added'), 0.0 otherwise
• num_days = days since this author’s last touch of the file, measured against the repo’s newest commit (recency, per the DOE definition)
• size = HEAD SUM(sloc) from complexity_metrics (clamped ≥ 1; the formula has ln(size) without a +1 guard, so clamping prevents ln(0))

Expert threshold: doe >= 1.0 AND doe >= 0.75 × max_doe_for_file. This normalization convention is adopted from Avelino’s DOA work; the DOE paper itself leaves the threshold unstated.

Constants

DECAY_DAYS

Exponential decay constant in days. Source: Jabrayilzade et al., ICSE-SEIP 2022 (arXiv 2202.01523 §3.1).

W_AI_ASSISTED

AI-assisted commit knowledge weight (same source as W_AI_AUTHORED).

W_AI_AUTHORED

AI-authored commit knowledge weight. Heuristic motivated by arXiv 2507.08160 (GenAI distorts expertise models). See also W_AI_ASSISTED.

W_REVIEWER

Reviewer knowledge weight relative to author. Sources: Jabrayilzade et al. 2022 (reviewers = ½ author weight); Rigby & Bird, ESEC/FSE 2013 (review transfers 66–150% of authorship).

Functions

materialize_knowledge_shares

Materialises two temporary tables into db: