Research Library

Limitations: Observed Claude usage is not a full labour-market census and is still a platform-specific measure.

Repo use: Active live source in the audited exposure ensemble and a major input to the future task-native direction.

Limitations: Commentary rather than peer-reviewed estimation; proposes collecting new price/quantity data rather than a ready-made occupation-level measure.

Repo use: Motivates the V7 demand-persistence proxy. The proxy measures recent labour-demand persistence (momentum, vacancies, scarcity), not output-price elasticity, and does not capture dimensionality — it is a partial response to this critique, not a resolution.

Limitations: Uses US outcome data and a platform-linked usage measure, so it still needs Singapore-specific interpretation.

Repo use: Primary candidate reference for promoting the shadow model beyond readiness-only governance.

Limitations: Experimental and workflow-specific evidence still needs careful translation into occupation-level scoring.

Repo use: Useful as a V5.1 calibration reference rather than as a direct V4.x score ingredient.

Limitations: The repo uses a simplified concentration buffer rather than the paper's full firm-task empirical setting; the buffer magnitude (lambda = 0.20) is heuristic, not estimated from the paper's coefficients.

Repo use: Primary scientific justification for the live V7 task-concentration exposure buffer (concentration offsets labour-demand losses via within-job task reallocation).

Limitations: Still a global exposure measure rather than a Singapore outcome model.

Repo use: Included because its ISCO alignment improves crosswalk robustness for the ensemble.

Limitations: The study is Denmark-specific and examines early effects; longer-run displacement remains unresolved.

Repo use: Supports the repo's choice to keep structural risk separate from near-term or realised-risk interpretations.

Limitations: Provides strong augmentation evidence but still within constrained experimental settings.

Repo use: Supports the repo direction of modelling augmentation as its own construct rather than as a mirror image of automation.

Limitations: Adoption speed alone does not identify realised labour displacement or productivity effects.

Repo use: Used to justify separating structural pressure from near-term realised-risk proxies.

Limitations: Cross-country evidence informs the transition design, but the repo still needs a Singapore-specific mobility dataset.

Repo use: Guides the schema for observed transition priors and future ranking logic in the transition layer.

Limitations: The effect size is workflow-specific and should not be generalized into a universal augmentation constant.

Repo use: Candidate reference for replacing a single structural augmentation heuristic with workflow-aware priors.

Limitations: The paper is early and US-oriented, and it does not by itself provide Singapore labour-market calibration.

Repo use: Used as a reference for the repo's future task-native model direction, not as a direct live source key.

Limitations: Like other exposure indices, it measures capability overlap rather than realised displacement.

Repo use: Kept in the live ensemble because it adds an LLM-specific construct not covered by AIOE alone.

Limitations: Designed as a broad labour-exposure framing rather than a Singapore occupation-level calibrated outcome model.

Repo use: Directly tied to the theta-based bottleneck implementation in the current scorer.

Limitations: Measures theoretical exposure rather than realised AI use or job outcomes.

Repo use: Tied directly to the live AIOE source key and still used in the canonical exposure ensemble.

Limitations: Provides the conceptual bridge from AI capabilities to work content, but not current observed usage or Singapore-specific outcomes.

Repo use: Referenced as the foundational exposure framework behind later AIOE-style occupation measures.

Limitations: Aggregate exposure/complementarity shares at the economy level, not occupation-by-occupation scores; uses the IMF exposure framework rather than this repo’s 4-source ensemble.

Repo use: The most directly relevant external benchmark for a Singapore AI-exposure product. Not yet used for formal calibration; flagged as a convergent-validity target more relevant than US BLS.

Limitations: This is US evidence and should not be interpreted as Singapore occupation outcome truth.

Repo use: Lives in the repo as both a validation benchmark and a clearly labeled external proxy.

Limitations: The benchmark is model-centric and software-task heavy, so it is not a direct occupation impact measure.

Repo use: Useful for the forecast layer, not for the core structural score.

Limitations: This is a measurement agenda rather than an occupation-scoring formula.

Repo use: Best reference for the repo's longer-run shift from heuristic confidence toward richer uncertainty and task-native measurement.

Limitations: Measures annotator instability, not which annotation is correct; ensemble averaging across independent measures dilutes but does not remove the error.

Repo use: Disclosed in Known Limitations: affects the Eloundou and ILO components of the exposure ensemble.

Limitations: O*NET is US-based and enters the Singapore product mainly as explanatory or crosswalk context.

Repo use: Not a direct structural-score input, but essential for the supporting task layer and future task-native scoring.

Limitations: US workforce reweighting; the direction of bias for Singapore depends on local adoption composition, which is not measured at occupation level.

Repo use: Disclosed in Known Limitations: affects the Anthropic observed-usage component of the exposure ensemble.

Limitations: It is a summarized skills view, not a direct task-by-task automation score.

Repo use: Best public structured skill overlay for the US country layer.

Limitations: US payroll data; the Feb 2026 update concedes effects are significant only from 2024 under the broadest macro controls. Entry-margin evidence, not occupation-level displacement totals.

Repo use: Primary citation for the seniority modifiers and for the hiring-margin limitation: entrant-facing and incumbent-facing risk are different objects.

Limitations: Occupation/industry classification changes reduce perfect longitudinal comparability.

Repo use: Useful for showing who is exposed and for planning transition context, not for changing the structural pressure score.

Limitations: Internal workplace evidence is highly specific to Anthropic and should not be generalized into a country labour market without adjustment.

Repo use: Useful as a contextual adoption case study and a reminder that augmentation can change task mix, learning speed, and workflow design.

Limitations: OEWS is US-specific and must not be interpreted as a global wage distribution.

Repo use: Best public source for the US wage layer and the richest local wage context currently available.

Limitations: Designed for career guidance rather than AI-risk scoring.

Repo use: Adds the human-readable occupation narrative layer for the US market.

Limitations: ORS is not a labour-market outcome series; it is a requirements survey used for bottleneck and friction context.

Repo use: The best current official analogue for transition friction and human-bottleneck validation in the US layer.

Limitations: Historical US evidence; no AI-era occupation-level reinstatement rate yet exists that could serve as a score input.

Repo use: The canonical quantified evidence for why displacement-only scores overstate long-run risk; cited in the central limitation.

Limitations: Singapore is not in the PIAAC/STEP samples, so the bias cannot be directly corrected; high-income economies show smaller divergence from US task content.

Repo use: The canonical citation for the SSOC-ISCO-SOC-O*NET crosswalk caveat; motivates benchmarking against Singapore-native task data (SkillsFuture Skills Framework).

Limitations: US transition-network data; Singapore-specific mobility networks are not yet measured, so the repo applies the insight via its structural transition-capacity scores.

Repo use: Grounds the high-risk x few-exits quadrant surfacing (rankings and occupation pages): escape-route quality, not pressure alone, shapes outcomes.

Limitations: Pre-AI-era programmes, mostly Europe/US; Singapore schemes (SkillsFuture, SCTP, CCP) have published only selection-biased administrative outcomes, no causal evaluation.

Repo use: Grounds the caveat that transition-capacity scores measure structural skill adjacency, not evidence that retraining works.