Future work.

1. Input-output derivation of customer-industry shares

Status: Phase 4. Identified, scoped, deliberately not attempted in the v0.6 pass.

What: replace the hand-authored customer_industries[].share values with values derived from public input-output tables. For each occupation, compute revenue share by joining SOC → industry-of-employment → customer-industry through the BLS Industry-Occupation Matrix and the BEA Use Tables, then re-normalising to sum to 1.

Why: the coefficient notes already flag this as the single largest model-level uncertainty. The shares directly drive flow_i at the heart of the cascade. Moving them from "expert estimate" to "BEA Use Table column" closes the biggest remaining judgement-vs-data gap.

Data:

• US: BLS Industry-Occupation Matrix (832 SOC × 292 NAICS, XLSX, bls.gov/emp/tables) and Inter-industry I-O Matrix (bls.gov/emp/data).
• UK: ONS Blue Book 2024 Supply-Use Tables plus 4-digit SOC × 2-digit SIC employment matrix.
• AU: ABS National Accounts Input-Output Tables (Product Details, latest 2019-20) + Labour Force Detailed (ANZSCO/OSCA × ANZSIC).
• CA: StatCan Symmetric I-O Tables (Table 36-10-0001) + Labour Force by NOC (Table 14-10-0416).

Effort: ~1 week US, ~1 month all four countries. Risk is high because it touches the load-bearing share field - see the deferred-handoff notes in the changelog.

2. Non-US AIOE and AEI coverage via occupation-code crosswalks

Status: Pending crosswalks.

What: AIOE (Felten-Raj-Seamans) is published on US SOC codes; AEI's SOC-major rollup is similarly SOC-anchored even when the country panel is non-US. UK occupations carry SOC 2020, AU uses ANZSCO, CA uses NOC. To get AIOE percentile and AEI aei_soc_major_pct injected into UK/AU/CA occupations, the ETL needs three small crosswalk files mapping each foreign code to US SOC.

Why: the non-US country pages are currently in a "country-level metadata only" state for AEI, and have no AIOE at all. The signals already work for US - extending them to UK/AU/CA gives a 3-of-4 country jump in coverage with a small per-country crosswalk file (~10–12 codes each at our current occupation count).

Data: ONS publishes SOC 2020 ↔ SOC 2010 crosswalks; ABS publishes ANZSCO ↔ ISCO; StatCan publishes NOC ↔ SOC. ISCO ↔ SOC official maps exist via the BLS.

Effort: 1–2 days per country if we restrict to the current 10–12-occupation seed set. Larger if the occupation list grows.

3. AEI time-series anchoring of scenario S-curves

Status: Conceptually clear, waiting on data accumulation.

What: AEI now has five releases published between Feb 2025 and Mar 2026. With multiple snapshots we can fit an empirical near-term adoption curve per SOC major group, and use that fitted curve to constrain the early portion (months 0–24) of every scenario S-curve to match observed adoption velocity.

Why: v0.6 ingests AEI as a static snapshot - valuable, but limited to "this is the level today." A fitted velocity replaces "scenario asserts a midpoint of month 24" with "scenario midpoints are bounded by the velocity AEI is actually showing for this occupation cluster." This is the single biggest near-term certainty win remaining.

Data: aggregate the same soc_pct field across all five published AEI releases (release_2025_02_10 through release_2026_03_24, all on Hugging Face) and fit per-SOC-major-group adoption velocity.

Effort: ~3 days to download, normalise across AEI's slightly-shifting schemas, and write the fit. Plus thinking about how to weight country panels with very different sample sizes.

4. Out-of-sample backtest against observed employment

Status: Possible, modest signal-to-noise.

What: compare the simulator's predicted revenue/employment trajectories for months 0–30 against actual 2023–2026 changes from BLS OEWS, BLS QCEW, ONS LFS, ABS Labour Force Detailed, and StatCan NOC tables. Calibrate scenario parameters where the model systematically misses.

Why: turns the uncertainty bands from "literature-anchored" into "literature-anchored AND retrospectively defensible."

Honest constraint: 2023–2026 occupation-level employment is dominated by post-COVID normalisation, rate-cycle effects, and immigration shifts. AI-attributable change is real but not the dominant signal for most occupations. This kind of backtest works best for the handful of occupations where AEI shows the highest observed usage (translators, copywriters, junior software engineers, customer support) - for the rest, signal-to-noise is too low to be useful.

Effort: ~1 week for a focused 5–10 occupation backtest. Avoid promising universal calibration.

5. Direct occupation-to-occupation cascades

Status: Modelled implicitly via Moretti local-multiplier; not yet endogenised.

What: When software-developer income drops, explicitly shrink the real-estate-agent revenue pipe in proportion to the share of estate-agent revenue that comes from software-developer customers. Today's model captures the aggregate effect via the Moretti spillover (tradeable occupation revenue drops → local-service occupations get a weighted drag), but cannot trace specific occupation-to-occupation channels.

Data: BLS Consumer Expenditure Survey shares - what fraction of each occupation's income flows to which downstream consumption category, broken down by household income.

Effort: ~1 week to ingest CES + write the per-occupation linkage layer. Worth doing after Phase 4 lands because the CES data is more useful when the customer-industry shares are already proper.

6. Income-decile distribution within an occupation

Status: Mean-only.

What: Replace each occupation's single wage_mean with a P10/P25/P50/P75/P90 distribution. Lets the simulator answer "how does AI displacement hit the bottom decile of paralegals vs the top decile?"

Data: BLS OEWS publishes wage percentiles per SOC; ONS ASHE publishes the same for UK; ABS and StatCan equivalents exist.

Effort: ~2 days per country once the ETL is plumbed for percentile fields.

7. Regional variation within a country

Status: National means only - flagged as a live caveat on the methodology page.

What: Per-state / per-MSA wage and employment breakdowns so a software developer in San Francisco and one in Phoenix don't look identical.

Data: BLS OEWS publishes by MSA; ONS publishes by region. The cost is a much larger occupation-payload size (200× MSA blowup) and more careful UI design.

Effort: ~2–3 weeks. Mostly UI work - once you have regional picks, the cascade redraws need to handle regional context cleanly.

8. Policy intervention scenarios

Status: Out of v0 scope.

What: Toggleable scenarios for retraining subsidies, sectoral protections, UBI, or transfer-payment responses. Each shifts adoption curves and demand factors in documented ways.

Why later: the underlying mechanic is well-defined but the political framing is loaded; doing it usefully needs a careful design pass on which interventions are modelled, what evidence anchors their parameters, and how to communicate uncertainty given even more compounded assumptions.