Skip to content
AI Work Index

V4.3 Shadow Model Note

Status: V4.3-shadow task-weighted shadow model

Promoted live

This page explains release impact and readiness. It does not replace the formula spec on Methodology, the threshold detail on Appendix, the schema/download contract on Data, or the citation layer on Research.

Comparison baseline

V4.2

pre-promotion live release used for comparison

Required inputs ready

4/4

present locally for shadow scoring

Published task-native occupations

485

occupations currently scored with the task-native shadow model

Validation comparison

2/3

current match-or-improve gates passing

Published shadow artifacts

The shadow layer is now auditable as data, not just as a readiness note.

Shadow scores

Per-occupation task-adjusted scores and fallback status.

Comparison summary

Score deltas, band flips, and anchor-review counts versus V4.2.

Validation comparison

BLS, family, and cluster comparisons against the live baseline.

Current shadow validation deltas: cluster -0.6667, BLS -0.0471, family -0.2037.

What Changes Already Affect Users

Now live in V4.3

  • Bootstrap uncertainty intervals are published on occupations today.
  • Structural risk and near-term risk are separated in the forecast layer.
  • Task-primitives fields now publish weighted evidence where normalized O*NET task matches exist; sparse occupations remain explicit null.
  • The release and governance surfaces now expose shadow-model readiness instead of hiding it.
  • 485 occupations now have published task-native shadow scores for comparison against the live baseline.

What the audit trail still preserves

  • The V4.3 shadow model has already been promoted into the live structural release.

Remaining Input Gaps

  • All required local shadow-model inputs are now present.

Input Readiness

anthropic task penetration

data/raw/external/anthropic_task_penetration.csv

present

onet task statements

data/raw/external/onet/Task_Statements.txt

present

onet task ratings

data/raw/external/onet/Task_Ratings.txt

present

empirical mobility

data/raw/external/sg_empirical_mobility.json

present

Coverage Snapshot

Occupations

562

current published universe

Direct mapped

521

eligible for the direct coverage gate

Median direct matched task share

100%

current direct-coverage gate basis

Task-weighted share

86%

headline score still untouched

Promotion Gates

GateThresholdActualState

Median matched task weight share across direct-mapped occupations

This gate prevents a sparse task layer from directly changing the headline score before task matching is broadly comparable.

>= 0.61pass

Experimental task-adjusted score matches or improves current validation diagnostics

Requires at least 2 of 3 external checks to match or improve baseline. Current results: cluster directional accuracy 0.3333 vs 1; BLS rho -0.1908 vs -0.1437; family rho -0.4457 vs -0.242.

 at_least_2_of_32/3pass

No implausible anchor label flips without written rationale

8/8 anchors screened; 0 candidates still need editorial sign-off.

 zero_unexplained_flips0pass

What V4.3 Proved

V4.3 proved that task evidence works best as a disciplined exposure upgrade, not as a full wholesale rewrite of the structural formula. The candidate task-native formulas remain published as research scaffolding for V5, not as live rules.

effective_coverage = Σ_t w_it · exposure_t · success_t

net_risk = automation_pressure_i · (1 - λ · concentration_i) · market_modifier_i

What Must Happen Next

  • Keep the full V4.2 vs V4.3 comparison published so the promotion remains auditable.
  • Start V5 as sidecar workstreams: augmentation heterogeneity, empirical mobility, posterior uncertainty, and realized-risk forecasting.
  • Do not absorb multiple new constructs into the live score without separate sidecar validation first.
  • Treat the current empirical mobility prior as supporting evidence until a higher-granularity Singapore transition dataset exists.