# Large Labor Model v5.0 — Changelog

**Released:** 2026-04-19 (Phase 12 historical backfill + late fixes applied 2026-04-20)
**Previous version:** v3.2

---

## Headline changes

- **Historical backfill 1970 → 2014 (Phase 12, added 2026-04-20).** v5.0 now ships 585 new `territory_replaceability` records covering 1970–2014, produced by a dedicated Phase 12 pipeline: two parallel honest-deployment research instances (GPT-5.4 + Claude Opus 4.7) reconstructed the 1970–2025 capability-vector trajectory with URL-cited milestone anchors; programmatic reconciliation applied honest-deployment discipline (`mid = min(A_mid, B_mid)`); per-occupation replaceability computed via logistic crossover against current task decompositions; employment-weighted per territory. Fixes a late-caught regression where the canvas displayed 0% replaceability for every territory before 2015.
- **Ground-up rebuild, not a patch.** Every scoring column in v5 was recomputed from a new Phase 4 task decomposition, a dual-recalibrated 2026 capability vector, and an empirically fit replacement formula. v3.2 served only as post-hoc sanity check; never staged as input to any v5 phase.
- **Canonical ISCO-08 codes for all 480 occupations.** v3.2 had inherited ~200 wrong ISCO codes and 89 fabricated codes from v3.0 (e.g., "Chief Executives" at 1111 — the real 1111 is Legislators; "Physicists and Astronomers" at 2112 — the real 2112 is Meteorologists). v5 re-anchors every entry to the ILO ISCO-08 hierarchy, with specialised splits (e.g., `2512-ml`, `2261-orthodontist`) using a documented parent + suffix convention.
- **Replaceability and Replacement are now cleanly separated.** Replaceability (the 2026 score) answers "if swapped to an autonomous system today, does it work? — tech × economics × availability." Replacement (the displacement curves) answers "how fast does that capability translate into actual workforce displacement?" Regulatory and human-preference frictions live in the Phase 6 replacement model, not in the replaceability score. The distinction was corrected in Phase 11 after a review-caught conflation (see §2.6b in the editorial process record).
- **Post-recalibration 2026 capability vectors.** After two independent Phase 11 recalibrators (Claude Opus 4.7 and GPT 5.4) applied a benchmark-to-deployment honest lens, 2026 capability values were lowered: **C_R 0.87 → 0.76, C_G 0.62 → 0.57, P_A 0.78 → 0.75, Phi_S 0.60 → 0.46, Phi_U 0.15 (held), S_E 0.42 → 0.35**. Ordering preserved. The shipping snapshot is stored in `metadata.capability_vectors_2026_snapshot`.
- **Three-scenario framing.** Every occupation has low/mid/high bands for 2026 and for 2030/2035/2040/2041. Mid is the moderate, displayed default. Low is conservative; high is accelerated.

---

## Coverage

| Record type | v3.2 | v5.0 |
|---|---:|---:|
| Occupations | 388 | **480** |
| Labor records (territory × year) | 608 | **635** |
| Technology timeline events | 57 | **77** |
| Territory × year replaceability entries | 327 | **936** (585 historical 1970–2014 + 351 modern 2015–2041) |
| Historical macro groupings | 3 | **3** |
| Modern territories | 13 | **13** |
| Inline task decompositions | — | **4,818 tasks across 480 occupations** |
| Displayed occupations (curated) | 104 | **104** (13 territories × 8) |

Task decompositions are a v5-only addition — v3.2 had none. The inlined 4,818 tasks (mean 10.0 per occupation, range 2–14 after Audit 3 removals) each carry a capability vector, difficulty threshold, time weight, and source-id citations. Phase 4 originally produced 5,655 tasks; Phase 11 audits removed 825 scrape artifacts / wrong-occupation imports across four intervention passes. The 423-record `metadata.source_count` is the unique-by-registry-key count; the inline `sources[]` array has 6,686 task-level and labor-level references for per-entry provenance.

### Territory counts

| Territory | v3.2 | v5.0 | Δ |
|---|---:|---:|---:|
| Making Things | 45 | 75 | +30 |
| Money & Data | 43 | 57 | +14 |
| Care & Health | 40 | 56 | +16 |
| Land & Sea | 31 | 40 | +9 |
| Making Meaning | 34 | 37 | +3 |
| Building Things | 25 | 34 | +9 |
| Maintaining & Fixing | 30 | 34 | +4 |
| Governing & Protecting | 26 | 31 | +5 |
| Moving Things | 32 | 31 | −1 |
| Feeding & Hosting | 24 | 26 | +2 |
| Buying & Selling | 24 | 22 | −2 |
| Learning & Teaching | 23 | 22 | −1 |
| Thinking & Leading | 11 | 16 | +5 |

Thinking & Leading was redefined at Checkpoint 2 ("Organization-wide leadership, cross-functional knowledge work, and the administrative support surrounding them — deciding what the organization does and coordinating how it runs"). General office clerks (4110) and secretaries (4120) moved into Money & Data at Checkpoint 3; the HR / executive-support sub-cluster stayed in T&L.

---

## Methodological changes

### New fields on every occupation

- `task_decomposition[]` — the ordered list of tasks with per-task `vector`, `difficulty_threshold`, `time_weight_pct`, and `source_ids`.
- `replaceability_2026_low` and `replaceability_2026_high` — scorer-disagreement bands preserved from Phase 5 reconciliation (mean spread 38.8 pp, median 34 pp; wider on creative / unstructured-physical work where honest uncertainty is larger).
- `replaceability_2030_moderate` and `replaceability_2040_moderate` for the projection anchors (the full dataset carries 2030/2035/2040/2041 values with low/mid/high).
- `editorial_justification` (mandatory) — short free-text tying each occupation-level decision to the editorial principle it applies.
- `confidence` and `confidence_reason` — qualitative tags per occupation.
- `projection_note_2030` — narrative note supporting the 2030 anchor where non-trivial.

### Retained from v3.2 schema

`isco_code`, `occupation_name`, `display_name`, `territory_id`, `common_titles`, `primary_barrier`, `barrier_types`, `benchmark_anchor_2026`, `key_barrier_note`, `notes`, `employment_weight_pct`, `global_estimated_employment`, and the `displacement_{2026,2030,2035,2040}_moderate` series.

### Schema amendments applied during Phase 10 assembly

Five patch rounds were required to bring the release candidate to zero validation failures. The substantive schema adjustments: `display_name` maxLength raised from 25 to 35 (to accommodate titles like "Early Childhood Educators"); `barrier_types` and `primary_barrier` enums now include `NONE`; `technology_events` entries allow `access_date`, `citation`, `source_url`, `v32_status`, and `verifier_note`; occupations optionally carry an `isco_verification_note`.

---

## Scoring changes

### Aggregate shifts vs v3.2

Across all 293 occupations whose ISCO codes match in both datasets:

- **Mean replaceability_2026:** v3.2 58.63 → v5 44.34 (**−14.29 pp** on matched subset).
- **Mean displacement_2040 (moderate):** v3.2 37.70 → v5 27.03 (**−10.67 pp** on matched subset).

Across the full v5 population, mean r26 is **45.01** (v3.2's full-population mean was 58.24) and mean displacement_2040 (mid) is **27.27** (v3.2 was 37.97). Two effects drive the r26 drop: (a) the Phase 11 recalibration lowered 2026 capability values on every vector except Phi_U, and (b) four Phase 11 audit rounds removed scrape artifacts and reclassified physical/embodied tasks that had been mis-coded as cognitive in Phase 4.

### Territory-level r26 means (v5, mid)

| Territory | r26 mean | r40 mean (mod) |
|---|---:|---:|
| Making Things | 68.7 | 99.9 |
| Money & Data | 68.5 | 100.0 |
| Buying & Selling | 63.7 | 100.0 |
| Learning & Teaching | 54.5 | 99.7 |
| Governing & Protecting | 44.5 | 94.8 |
| Thinking & Leading | 40.8 | 100.0 |
| Making Meaning | 39.0 | 95.5 |
| Feeding & Hosting | 37.7 | 99.5 |
| Moving Things | 35.9 | 98.2 |
| Care & Health | 35.3 | 93.3 |
| Maintaining & Fixing | 23.3 | 98.9 |
| Building Things | 23.0 | 99.4 |
| Land & Sea | 21.9 | 99.2 |

The r40-mid surface is aggressive: 88.1% of occupations reach replaceability 100 by 2041 under the moderate scenario; only 3.1% remain below 80. This is honest task-math: once capability trajectories saturate past Phase 4 per-task difficulty thresholds, per-task replaceability saturates too. The variance at 2041 is carried by displacement, not by r40. Mid-scenario displacement_2040 ranges from 9.3% to 47.9% with a mean of 27.27%.

### Top 20 individual movers (matched occupations, |Δ r26| largest)

| ISCO | Occupation | v3.2 r26 | v5 r26 | Δ |
|---|---|---:|---:|---:|
| 9112 | Cleaners and helpers | 81 | 0 | −81 |
| 7211 | Metal moulders and coremakers | 85 | 8 | −77 |
| 9611 | Garbage and recycling collectors | 72 | 3 | −69 |
| 2653 | Dancers and choreographers | 69 | 0 | −69 |
| 6210 | Forestry workers | 66 | 0 | −66 |
| 9122 | Vehicle cleaners | 67 | 1 | −66 |
| 8311 | Locomotive engine drivers | 74 | 10 | −64 |
| 9213 | Mixed crop and livestock labourers | 63 | 0 | −63 |
| 6111 | Field crop and vegetable growers | 63 | 1 | −62 |
| 7119 | Building-frame and trades workers | 62 | 0 | −62 |
| 3131 | Power production plant operators | 17 | 78 | +61 |
| 7312 | Musical instrument makers | 68 | 8 | −60 |
| 6221 | Aquaculture workers | 59 | 0 | −59 |
| 1223 | R&D managers | 75 | 16 | −59 |
| 6121 | Livestock and dairy producers | 59 | 0 | −59 |
| 6224 | Hunters and trappers | 58 | 0 | −58 |
| 5241 | Fashion and other models | 58 | 0 | −58 |
| 1113 | Traditional chiefs | 58 | 0 | −58 |
| 7321 | Pre-press technicians | 37 | 94 | +57 |
| 3255 | Physiotherapy techs and assistants | 77 | 21 | −56 |

The direction of the largest drops reflects two Phase 11 patterns. First, embodied-craft, performance, and unstructured-field work that v3.2 scored as cognitive-replaceable was re-vectored to Phi_U (0.15) once audits caught the mis-coding (dancers, models, metal moulders, musical-instrument makers, foresters, farm labourers). Second, physical occupations carrying the `HUMANOID_DEPENDENT` barrier now consistently reflect the deployment absence — in 2026, there is no scaled humanoid for cleaners, vehicle cleaners, farm workers, hunters, or chiefs. The largest upward movers are structured-industrial operators (3131 power plants, 7321 pre-press) and teachers at the cognitive-codifiable end (2353 language teachers), where Phase 4 decomposition better captures the automatable core.

---

## Barrier taxonomy changes

### Enum

`NONE` is a v5 addition to both `primary_barrier` and `barrier_types`. v3.2 had only 2 occupations with no explicit barrier; v5 has 42 — cases where no structural barrier blocks deployment and only adoption-diffusion time gates it.

### Distribution

| Barrier | v3.2 | v5.0 |
|---|---:|---:|
| HUMAN_PREFERENCE | 170 | 93 |
| HUMANOID_DEPENDENT | 99 | 123 |
| ECONOMIC | 65 | 154 |
| REGULATORY | 52 | 69 |
| NONE | 2 | 42 |

The sizeable v3.2 → v5 shifts are driven by the corrected Replaceability ≠ Replacement framing (see the editorial process record). Many v3.2 HUMAN_PREFERENCE assignments were really ECONOMIC (unit economics break the swap today), and NONE cases that v3.2 buried under HUMAN_PREFERENCE defaults are now surfaced honestly.

---

## Projection structure changes

- **Three-scenario framing** replaces v3.2's single central projection. Low = conservative; mid = `min(A_mid, B_mid)` across two Phase 7 forecasters (the tempered read, the default displayed); high = the accelerated read anchored on the more aggressive forecaster's mid.
- **2030 anchors for every occupation,** computed directly from task-decomposition × capability-trajectory × barrier, then reconciled.
- **Terminal year 2041** (= 2026 + 15). v3.2 terminated at 2040. v5's 2041 anchor is canonical and the 2040 values are retained for continuity with v3.2 readers.
- **Capability trajectories published.** `metadata.capability_trajectories_summary` shows the year-by-year mid-band trajectory for each of the six vectors (2026, 2030, 2035, 2041). This summary preserves the Phase 7 mid trajectory values as projection inputs; the 2026 snapshot in `capability_vectors_2026_snapshot` records the post-recalibration starting values used for scoring.

---

## Formula changes (replacement lag)

v3.2 → v5 parameter diff:

| Parameter | v3.2 | v5.0 |
|---|:---|:---|
| Conversion rate | 0.333 | **0.30** |
| Lag schedule 2026 / 2030 / 2035 / 2040 / 2041 | 5 / 28 / 48 / 58 / 60 % | **7 / 18 / 30 / 42 / 44 %** |
| `already_absorbed` | uniform 0.15 | **era-stratified:** pre-2015 0.30 / 2015–2022 0.15 / post-2022 LLM 0.05 |
| NONE barrier multiplier | (implicit 1.0) | **1.20 ×** |
| ECONOMIC | (implicit 1.0) | **1.00 ×** |
| HUMAN_PREFERENCE | (implicit 1.0) | **0.562 ×** (Phase 11 strengthened from Phase 6's 0.70) |
| REGULATORY | (implicit 1.0) | **0.405 ×** (Phase 11 strengthened from Phase 6's 0.50) |
| HUMANOID_DEPENDENT | (implicit 1.0) | **0.15 ×** |

Barrier-type stratification was validated empirically at Phase 6: cross-section R² jumps from 0.61 (plain conversion × lag) to 0.85+ when barrier type is included as a categorical predictor. The Phase 11 recalibration strengthened HUMAN_PREFERENCE and REGULATORY proportionally to preserve empirical fit against the Anthropic Economic Index observed-coverage anchor after Phase 5 r26 scores were rescored to pure task-math. Historical-case anchors (telephone operators 1920–2023, bank tellers, travel agents, cashiers, translators, CSRs, bookkeeping clerks) constrain the lag schedule.

---

## Display curation

104 occupations carry `display_in_territory = true` (13 modern territories × 8 each), matching v3.2's display density. The curation was locked at Phase 11 (`decisions.json`) after reviewer-directed swaps to improve editorial representation.

### Display name overrides applied

- 7126 "Plumbers and pipe fitters" → **Plumbers** (Building Things)
- 7311 "Precision-instrument makers and repairers" → **Precision Instrument Makers** (Making Things)
- 0310 "Armed forces occupations, other ranks" → **Soldiers** (Governing & Protecting)

### Notable swaps from v3.2 display selection

- **Making Things:** Mechanical Engineers dropped in favour of Robotics Engineers (`2149-robotics`). CNC Operators and Potters (7314) added.
- **Thinking & Leading:** Sociologists dropped (too similar to Humanities Academics). Kept Mathematicians, HR Specialists, Chief Executives, Policy Professionals, Humanities Academics, Economists, Physicists.
- **Governing & Protecting:** Military Officers (0110) swapped for Soldiers (0310, ~44% of G&P employment vs ~13% for officers; more iconic). Tax Inspectors retained over Customs Officers per reviewer: "automated borders are already a thing; AI auditing tax returns is the scary one."
- **Care & Health:** Child Care Workers added (previously no representation for the ~13% employed in child care).
- **Land & Sea:** Farm Workers and Horticulturalists replaced by Vegetable Growers (26.80% of L&S employment), Offshore Oil Workers (rigs/offshore signal), and Earth Scientists (scientific-layer representation).
- **Feeding & Hosting:** Waiters dropped (redundant with Kitchen Helpers + Cooks). Added Catering Managers and Bookmakers / Croupiers (gaming/non-restaurant).
- **Making Meaning:** Authors and Screenwriters added — the canonical "is AI replacing writers?" role, distinct from Journalists.
- **Learning & Teaching:** Driving Instructors added (out-of-classroom teaching context).
- **Maintaining & Fixing:** Watchmakers and Clock Repairers added (artisan/craft signal).
- **Money & Data:** Securities Traders dropped (weak outlier; four other finance roles cover the space).

---

## What's preserved from v3.2

v3.2's hand-curated editorial substrate was the most valuable inheritance. v5 preserves:

- **Occupation meaning as the preservation anchor.** Where v3.2 had attached a wrong ISCO code to a correctly identified occupation, Phase 2a reassigned the code while keeping v3.2's display name, territory assignment, employment weight baseline, and the spirit of the barrier note. 112 of 112 `PRESERVE_V32` cases held their territory across Phase 2d reconciliation.
- **Territory framework.** Thirteen modern territories plus three historical macro groupings. v5 refined T&L's definition and moved a handful of occupations between territories but did not add or remove any territory.
- **Display density convention.** Eight displays per modern territory, 3 historical groupings × 8 each for 1800–1950 curation.
- **Timeline events structure.** v5 preserved 38 v3.2 events verbatim, edited 16 (year or description corrections), removed 3 unverifiable entries, and added 23 new 2024–2026 events (AlphaFold 3, Claude 3.5 Sonnet, Claude Computer Use, Devin AI, EU AI Act, Anthropic Economic Index, etc.).
- **Barrier notes where they were still strong.** 125 v3.2 notes preserved unchanged at Phase 8; 158 light-edited; 189 got targeted rewrites with mandatory evidence URLs after a template-slop detection pass.

---

## Known limitations

- **Editorial overrides queued but not applied.** Phase 11 surfaced two evidence-backed concerns that are documented but not baked into v5.0:
  - *ISCO 5245 Service Station Staff* (r26 = 44, `primary_barrier = HUMANOID_DEPENDENT`). Self-service pumping is already universal across most of the world; the low r26 appears to be a Phase 4 task-decomposition artifact (tasks coded as Phi_U when they should be P_A or Phi_S). Proposed override: r26 → ~88, primary_barrier → REGULATORY.
  - *REGULATORY-barrier occupations with r41 ≥ 95* (Lawyers at r41=100, Senior Government Officials at 99, Patent Attorneys, Legislators, Litigation Attorneys). Projections imply near-full replaceability despite regulatory floors (bar exam, named-human mandates). The multiplicative barrier model probably needs a floor component for these cases.

  Both are deferred to post-launch editorial work or v5.1.

- **High r40-mid shape.** Under the moderate scenario, mean replaceability at 2040 is 84.6 across 480 occupations and 25% score ≥ 90. The logistic crossover (k=8) ceiling is asymptotic, so no occupation reaches exactly 100 — the maximum is around 98. Most occupations sit in the 80–95 band by 2040; this still pushes editorial weight onto the displacement column where the variance lives. Readers should interpret 2040 workforce implications via displacement (mean 24%, range 9–47%) rather than via r40 alone — replaceability above 90 means "most of the task profile crosses the capability threshold"; displacement at 24% is what tells you what the workforce actually looks like.

- **Residual scrape artifacts below the audit bar.** The Phase 11 audit bar was consistently 2-model agreement (Audits 1 and 3 both used adversarial QC). Below-threshold contamination likely remains. A v6 re-run of Phase 4 with tightened anti-scrape guardrails is recommended.

- **Benchmark-vs-deployment calibration evolves.** The post-recalibration 2026 capability values trim Phase 5's benchmark-anchored set toward observed deployment reliability. This is our best honest read as of April 2026; production reliability data for humanoid robotics, agentic coding, and autonomous vehicles is still developing and values may need further adjustment in v5.1 or v6.

- **Within-vector heterogeneity.** C_G (specialist QA to novel reasoning), Phi_S (robotaxis to heavy equipment), and S_E (coding agents to incident command) each cover wide internal distributions. Single-point values average over heterogeneity; the bounds carry some of the uncertainty but a v6 split of these vectors would add precision.

- **Phase 6 barrier-model form.** Multiplicative barrier multipliers work well across most of the distribution but produce counterintuitive results at high r26 × REGULATORY (lawyers at r41 = 100). A floor-adjusted barrier model for regulatory frictions is a likely v6 change.

- **Task-classification ceiling persists.** Phase 4's three-way reconciliation achieved 84.2% agreement; the 15.8% split cases were defaulted to the Codex adversarial pass and flagged. The Phase 11 audits re-examined a large subset (notably high-r26 physical occupations via Audit 2 and every task in every territory via Audit 3) but exhaustive per-split human adjudication was out of scope.

- **Historical backfill uses current task decompositions + weights.** Phase 12 produces honest historical capability values but applies them against today's occupation task lists and within-territory employment weights. The number reads as "what share of today's labor would have been technically replaceable at that year's capability" — not "what share of 1985's actual labor force." Reconstructing historical occupation composition at the 481-occupation granularity is deferred to a future version.

---

## Post-launch editorial polish (2026-04-21)

Two parallel review instances surfaced quality issues, applied after reviewer sign-off:

- **Phase 14 legibility sweep** — 33 display renames covering truncated labels, ambiguous bare-adjective displays, and three display-vs-ISCO mismatches. Examples: ISCO 2230 "Traditional & Complementary" → "Traditional & Complementary Medicine"; ISCO 4413 "Coding" → "Coding & Proofing Clerks" (it had been reading as software coding); ISCO 7234 "Aircraft Engine Mechanics" → "Bicycle Repairers" (the official ISCO 7234 scope; aircraft mechanics already exist separately at 7232); ISCO 2635 "Religious Professionals" → "Social Workers & Counsellors" (caught during Phase 15 review — 2635 is officially Social work and counselling; 2636 is the actual religious-professionals occupation).
- **Phase 15 common-titles enrichment** — 11 occupations updated to fill search gaps. Crypto/blockchain titles added to ISCO 3311 (cryptocurrency trader, digital asset trader, crypto market maker), 2512 (blockchain engineer, smart contract developer, solidity developer), and 2413 (on-chain analyst, crypto analyst). Yoga / Pilates / barre / wellness coaching titles added to ISCO 3423. Modern AI roles (prompt engineer, AI researcher, MLOps engineer, LLM engineer, AI safety researcher) replacing generic titles on ISCO 2512-ml. Three previously-broken arrays (2431-social, 3153-uas, 2512-ml) had been carrying contamination from unrelated occupations and are now correct.
- **MLM/waiter contamination dedup** — the `direct sales coach`, `independent beauty consultant`, `independent distributor`, `independent sales associate/representative`, `door-to-door sales trainer` cluster was contaminating four occupations where it didn't belong (ISCO 5211 Market Salespersons, 5212 Food Truck Operators, 9520 Street Vendors, 5221 Shopkeepers). Stripped from all four; preserved on ISCO 5243 (Door-to-door Salespersons) where the cluster is correct. ISCO 5212 also had five waiter-titles (food server, kitchen runner, etc.) belonging to ISCO 5131 — also stripped. Affected occupations now have thin titles flagged for v5.1 web-verified enrichment at `metadata.deferred_to_v5_1`.
- **ISCO 1219 territory move + display rename.** "Hospitality Directors" → "Admin Services Directors" + reassigned from feeding_hosting to thinking_leading. Rationale: ISCO 1219 official scope is "Business services and administration managers not elsewhere classified" — corporate admin/services directors, not hospitality industry executives, which already exist separately at ISCO 1411 (Hotel Managers) and 1412 (Restaurant Managers). The v5 functional-manager principle says functional managers go with their function. Within-territory employment weights renormalized in both territories using `global_estimated_employment` as ground truth — both feeding_hosting and thinking_leading still sum to 100.000% post-move. 1219's weight in thinking_leading: 18.19% (substantial chunk of admin-services managers globally; previously diluted at 2.99% inside larger feeding_hosting).
- **ISCO 3422-personal dropped.** Off-by-one specialized split: personal training officially sits in ISCO 3423, not 3422. The split had zero employment_weight (didn't drive territory aggregates) so removal had no aggregate impact. ISCO 3423 is now the canonical home for personal trainers, yoga, Pilates, group fitness, and wellness coaches. **Occupation count: 481 → 480.**
- **End-to-end recompute** of replaceability + displacement + territory aggregates against the modified inputs. Phase 10 validator: PASS, 0 critical failures.

Deferred to v5.1 (recorded at `metadata.deferred_to_v5_1`):
- ISCO 5211 / 5212 / 9520 thin-titles after MLM/waiter dedup (need Phase 15.x web-verified enrichment)
- Proposed specialized split ISCO 2422-compliance for AML/KYC/financial-crime officers (≥50K US LinkedIn postings, distinct task profile)

---

## Late-cycle fixes (2026-04-20)

After the initial 2026-04-19 release, a spot-check surfaced a set of small-but-real issues that were fixed on the dev branch before production deploy:

- **Stale capability trajectories summary in dataset metadata.** `metadata.capability_trajectories_summary` was carrying pre-parallel-shift Phase 7 values (C_R 2041 0.98 rather than 0.87, etc.). The block was regenerated from the post-shift `phase7/reconciler/capability_trajectories_2026_2041.json` and a traceability note was added.
- **Methodology doc cited the same stale 2041 anchors.** The v5_methodology.md paragraphs for 2041 mid and high anchors were patched to cite post-shift values, with a note pointing to the `*_phase7_original` fields in the trajectory file for traceability.
- **`metadata.scenarios_shipped` flag added.** v5.0 ships the moderate (mid) scenario as the editorially-reviewed display; low and high bands are structural only (produced deterministically from Phase 7 low/high trajectories; not reviewed occupation-by-occupation). The flag encodes this disclosure for downstream consumers.
- **Common-titles contamination cleanup.** A cross-occupation contamination pattern (the strings "caddymaster", "community life director", "aquatics supervisor", "bar and restaurant manager", etc.) was present in 7 ISCO 5xxx service-sector occupations. Titles stripped and 6 occupations augmented with domain-correct replacements (ISCO 3423, 5113, 5141, 5142, 5164, 5311, 5322).
- **Business-development coverage.** ISCO 3322 Commercial Sales Representatives common_titles were updated to include modern BD-track titles (account executive, business development representative, business development associate, sales development representative).
- **UI schema-compat shim.** `index.html` was updated with a one-line alias (`V2.replaceability ?? V2.territory_replaceability`) so the v5.0 rename of the territory-level replaceability array doesn't break the canvas. Both v3.2 and v5 schemas now render with the same front-end code.

---

*Large Labor Model v5.0 is the work of a reviewer-directed build pipeline that treated every prior version as sanity-check only. The editorial principles, phase-by-phase process, and Phase 11 emergency interventions are documented in the companion files [v5_methodology.md](v5_methodology.md) (public methodology page) and [v5_editorial_process.md](v5_editorial_process.md) (full post-mortem of the build journey).*

---

## Post-release revisions

### June 8, 2026 — display-data polish (Phases 16–17)
- `task_short` display labels added for all 4,811 tasks (Phase 16).
- Phase 17 Mirror-surfaced polish: four occupation score corrections (5321, 8142, 9623, 9629-municipal), duplicate-task cleanup on ten occupations, 22 territory-year points adjusted by ±0.1. No change to the model's method.

### August 15, 2026 — seam correction, currency patch, count fixes
- **Data/projection seam corrected** (methodology §6.3): each capability band's history previously reached the 2026 honest-deployment anchor early and sat flat (mid/low bands: 2025 duplicated the anchor; high band: C_R flat from 2024, P_A from 2023). Flat runs re-laid as strictly increasing approaches — single-year: `v2025 = max(midpoint(v2024, v2026), v2026 − (v2027 − v2026))`, clamped strictly inside `(v2024, v2026)`; multi-year: linear fill. The 13 territory-level 2025 replaceability records were recomputed (all lower, −1.0 to −7.2 points). Invariants verified by assertion: run-base years, all 2026+ values, and every published 2026 figure unchanged; all 18 vector×band series strictly increasing 2014 → 2028. Scripts: `v5-build/phase12/reconciler/seam_fix_aug2026.py`, `seam_fix_downstream.py`.
- **Timeline currency patch:** `technology_events` 77 → 94 (May–August 2026 frontier wave; embodiment and regulatory milestones; two 2026 mathematics results), one source entry each. No scores or capability values changed. Display remains one dot per year with one intentional second 2026 slot (year key 2026.5, floored for display).
- **Count fixes:** `metadata.source_count` 423 → actual entries (6,704); `metadata.occupation_count` 481 → 480.
- **Methodology page:** §4.1 conflated the workforce-weighted mean (33.7 for 2026) with the unweighted occupation mean (41.2); corrected.

### August 15, 2026 — 2025 anchored (Phase 12b)
- The interpolated 2025 capability values were replaced by a **three-instance research anchoring**: Instance A (Claude family, live retrieval, 44 URL-cited milestones), Instance B (GPT family, internal-knowledge sourcing, disclosed), Instance C (GPT-5.6 Sol, owner-run, de-biased ordinal-first protocol without the tempered-tiebreak instruction, 32 URL-cited milestones). Reconciliation: per-band minimum. The de-biased instance was lowest on the cognitive vectors (tempering is evidence-driven) and highest on the physical vectors (bracket narrowness bounds the disagreement to ≤0.014). No bracket stress reported by any instance.
- Territory 2025 records recomputed: −0.4 (Land & Sea) to −4.3 (Learning & Teaching) versus the interpolation. 2026 anchors, forward values, and all published 2026 figures unchanged (assert-verified; Nurse r2026 = 29.49).
- Standing evidence ledger for future recalibrations added: `docs/evidence_ledger_2026H2.md` (12 families, 52 live-verified readings, per-family tripwires).
- Protocol record: `v5-build/phase12b_2025_anchor/` (briefs, instance files, evidence, reconciliation and integration scripts).

### August 17, 2026 — 2041 checkpoint completion
- The replacement formula always specified a 2041 lag entry (`lag_schedule["2041"] = 0.44`) and the methodology anchors displacement interpretation at 2041, but per-occupation emission stopped at 2040. Added `replaceability_2041_moderate/low/high` and `displacement_2041_moderate` for all 480 occupations via the documented pipeline (logistic k=8 task-math, post-shift forward trajectory, Phase 6 formula).
- **Reproduction gate:** before emitting 2041, the extension script reproduced every stored 2040 field (480/480 within rounding). Zero 2040→2041 monotonicity violations.
- Results: mean 2041 displacement 24.2 (range 9.6–48.7); employment-weighted mean 2041 replaceability 84.2; replaceability-vs-displacement gap at the horizon: 60.0 points. Methodology §5 range corrected 9–47 → 10–49; horizon rationale (2041 = now + 15) documented in §6.
- Script: `v5-build/assembler/add_2041_checkpoint.py`.

### August 28, 2026 — job-request intake, first alias corrections
- The Mirror now takes requests for titles its search cannot place (`/api/request-job`; requests land publicly on the repo's issue #2). First triage from three days of live requests: **`delivery lead` and `delivery manager` added to IT Managers (1330) `common_titles`** — the nearest prior suggestion (Delivery Driver) was semantically wrong. Search metadata only; no scores changed (bundle rebuild verified Nurse r2026 = 29.5, all 4,811 task labels intact).
- `Survey Scripter` / `Survey Programmer` logged as **new-occupation candidates for v6** rather than aliased: no existing task table honestly represents survey-instrument development, and a wrong table is worse than no match.