# NOTES_2026-08-14.md — The Journeyman Age

The standing reference. Read this before touching the paper again.

---

## 1. Where the project stands

| Deliverable | State |
|---|---|
| `journeyman_age_v4_core.pdf` / `_full.pdf` | **Current draft** (v4, post-trial, 14 Aug 2026) |
| `journeyman_age_v3_core.pdf` / `_full.pdf` | Superseded; the draft the trial convicted; retained |
| `The Journeyman Age (v2).pdf` | Prior lineage (pre-pipeline, AI-reviewed); retained |
| `journeyman-age.zip` | v1 lineage (markdown paper, notes, CSVs, old fig script); retained |
| `data.py` | 62 entries, every number with source + status (V 22 / V2 37 / U 3); extended `audit()` fails on V-tags with tertiary-only sources |
| `verification_bottleneck.py` → `verification_bottleneck_result.json` | The identity, its sensitivity grid, the figure_params block, and the REFUSED institutional-compression field |
| `generate_figures.py` → `figures/` | Five figures, all reading data.py / the JSON (verified post-trial) |
| `ATTACK_NOTES.md` / `DEFENSE_NOTES.md` / `VERDICT.md` | The v3 trial, complete; 16 charges, ruling, 24 orders |
| `build_numeral_audit_v4.txt` | Every numeric token in v4's prose with context (order 7 sweep) |
| `research_notes.md` | Phase 0 header + the 14 Aug 2026 re-verification record |

Rebuild (two commands; the model runs first only if the JSON is absent):

```
python3 generate_figures.py
python3 build_paper_v4.py
```

## 2. The thesis, stated precisely (post-verdict form)

Two parts, different standings, never merged:

1. **Mechanism claim (task level — demonstrated for its study vintage):**
   print collapsed the cost of copying knowledge; AI collapses the cost of
   the competent first draft — the candidate application of knowledge — and,
   at the task level, shifts the binding constraint from generating work to
   standing behind it.
2. **Market claim (staked prediction, not a finding):** the relocation will
   surface in wages, hiring, and institutions as adoption deepens. First
   readings: Dallas Fed experience-premium result (supporting sign), NY Fed
   postings parallel-movement (contrary). Disconfirmer: no verification
   premium as the intensive margin deepens past 0.5–3.5% of work hours.

Plus the naming claim: "Journeyman Age" is accurate at the machine (day-rate
competence without accountable standing-behind) and at the human (issued
middle rung, dissolving apprenticeships, mastery scarce). The name carries a
**death condition**: the day a model can stand behind work — carry
liability, be deposed — the certification gap closes and the name dies.

## 3. What is established, and how well

| Claim | Strength |
|---|---|
| Novice-tilted compression gradient (Noy&Zhang; BLR QJE; Dell'Acqua; Cui) | Demonstrated for 2023–24 vintage; frontier replication open (§4.5(a)) |
| Expert reversal (METR 19% slower; 2026 follow-up weak) | Suggestive only |
| Aggregate null (Humlum & Vestergaard, effects >2% ruled out) | Best estimate **for the shallow-adoption period it measures** |
| Atrophy is a design choice (Bastani guardrailed-tutor arm) | One clean RCT |
| Entry-level decline (Canaries ~16%, firm-time FE) | Plausible and unproven; EIG band-sensitivity standing; check in 2028 |
| Print history (all headline dates/numbers) | Re-verified 14 Aug 2026, corrections logged below |
| Verification-bottleneck identity | An identity; parameters assumed; evidence for nothing; locates g/v and m as the empirical question |

## 4. What it will be attacked on, and the answers

The full attack surface IS `ATTACK_NOTES.md` (16 charges) — and the answers
are `DEFENSE_NOTES.md` adjudicated by `VERDICT.md`. The residual soft spots
after v4, for the next hostile reader:

- **The identity still assumes its gradient.** Answer: stated in §5 text
  now; g/v and m are offered as the measurement target. A critic who wants
  more is right — someone should measure g/v.
- **The market claim rests on two between-occupation proxies.** Answer: the
  paper says exactly that, and stakes the within-occupation test.
- **The arc still does rhetorical work as a diagnostic.** Answer: §3.4's
  demotion + fig scale warning + candidate-analogue labels; the residue is
  the price of using history at all.
- **In-house AI review ≠ external human review.** Conceded in the
  disclosure, in print. VERDICT order 23 (external human review before
  circulation) is OPEN.
- **Canaries band-sensitivity.** Standing objection, acknowledged in text.

## 5. What changed across drafts

- **v1 → v2** (pre-pipeline AI review): thesis re-cut from "cost of applying
  knowledge" to "cost of the competent first draft / relocation to
  verification"; the "260 years → 25" numerology withdrawn; timeline demoted
  to diagnostic; naming verdict darkened (workshop reading demoted from
  description to program).
- **v2 → v3** (pipeline port, 14 Aug 2026): every headline number
  independently re-verified (3 agents); ~20 corrections/updates (log below);
  data.py + model JSON + figure scripts created; new §5 (the identity);
  editions core/full from one script.
- **v3 → v4** (the trial): thesis split into mechanism/market with separate
  standings; falsification conditions made individually damaging and
  operationalized; §5 demoted to instrument + first market readings added
  (one supporting, one contrary); V-tags re-graded and audit() hardened;
  Ogilvie claim hedged and re-sourced; Community Notes corrected; internet
  base-rate contradiction fixed (n=3 + pending); METR time-horizon
  confronted in §11 + name's death condition stated; subtitle changed
  ("...Making of a Second Renaissance" → "...the Price of Standing Behind
  Work"); figures made to read their data, relabeled, rescaled-with-warning,
  renumbered; disclosure rewritten (AI review, commissioned, in-house; no
  external human review yet); status-of-claims appendix added; numeral
  audit added to the build. Nothing was retracted silently; v3 stands in
  the folder as convicted.

## 6. ERROR LOG (cumulative — so mistakes cannot silently revert)

Corrections from the 14 Aug 2026 re-verification sweep (entered v3):

| # | Was (v1/v2) | Is (source) |
|---|---|---|
| E1 | 15–20M incunabula attributed to Buringh & van Zanden | Febvre & Martin (1976); B&vZ supply only the ~11M manuscripts |
| E2 | "prices −75% by 1530, stabilizing at one-third" | −2/3 by 1500 → ~1/3 (Dittmar); −75% by 1530 → ~1/4 (Clark) — two statements, two bases |
| E3 | 3,600 pages/day implied Dittmar | Wolf (1974) via Wikipedia |
| E4 | print cities "+20pp" as THE estimate | "at least ~20pp" (OLS); matched ≥35pp; IV 60–80pp |
| E5 | "Luther = 20% of German printing 1518–27" | 20% of German PAMPHLET editions 1500–1530 (Köhler via Edwards) |
| E6 | 95 Theses → London in 17 days as fact | Palmer's illustrative claim only; primary undocumented; status U |
| E7 | Malleus "condemned by Cologne 1490" flat | "reportedly condemned"; episode contested (Mackay) |
| E8 | Trithemius printed 1492 | Composed 1492, printed 1494 (Mainz, von Friedberg) |
| E9 | BLR novices +34% | ~+30% in the published QJE version (34% was the 2023 WP) |
| E10 | Goh: "JAMA", LLM ~90% | JAMA Network Open; LLM median 92% |
| E11 | Adoption 39.4% (static) | +updates: 54.6% (2025 wave), 61.8% / 45.2% at work (May 2026) |
| E12 | ChatGPT 800–900M weekly "late 2025" | 900M weekly + 50M paying verified (OpenAI, 27 Feb 2026) |
| E13 | Capex "~$400B 2025, guided higher" | ~$410B 2025; ~$725B 2026 guidance (FT, Apr 2026) |
| E14 | Physical Intelligence "raised ~$1.6B" | ~$2.1B lifetime; $11.2B valuation (Mar 2026) |
| E15 | 1X NEO "shipments planned"/"shipping" | Production started 30 Apr 2026; NO verified deliveries as of Aug 2026 |
| E16 | SynthID ">10B watermarked" | >100B images/videos + ~60k years audio (May 2026) |
| E17 | EU AI Act "high-risk 2026–28" | Digital Omnibus (May 2026) defers Annex III to Dec 2027, Annex I to Aug 2028 |
| E18 | "retractions >10k/yr" as ongoing rate | Record >10,000 in 2023 (Hindawi-driven); annual counts fell after |

Corrections and process failures from the v3 trial (entered v4 — VERDICT order 22, retained permanently):

| # | Failure | Correction |
|---|---|---|
| E19 | Frontmatter claimed "every number re-verified 14 Aug 2026" while data.py carried "v2 basis" entries — **the flagship claim was falsified by the file it pointed to** | Claim corrected to "every headline number…; carried-over entries tagged"; audit() now fails on V-tags with tertiary-only sources (and immediately caught SCRIBES, re-graded V2) |
| E20 | **U-tag leak (process failure, not typo):** Ogilvie "permanent journeyman proletariat" unhedged in §11, unattributed as fact in the Conclusion | Re-sourced (V2, direction-claim only: Encyclopedia of European Social History; Truant); §11 hedged; Conclusion attributed + hedged. The strong phrasing is reported, not certified |
| E21 | "The Community Note is eleven years old" — **wrong by ~2×** (Birdwatch: Jan 2021), and the number lived outside data.py | Corrected to five and a half; COMMUNITY_NOTES entry added; the correction is acknowledged in the paper's own text; numeral-audit sweep added to the build |
| E22 | Out-of-pipeline numbers: literacy 1500, Republic of Letters, electricity ~40 yrs, Erasmus/Gessner | Entries added (LITERACY_1500 and REPUBLIC_OF_LETTERS as U/hedged; ELECTRIFICATION_LAG V2 via David 1990; OVERLOAD_QUOTES V2) |
| E23 | Figure scripts hardcoded 40 / 55.8; fig2 hand-set x-positions; fig5 g/v values absent from the JSON under a "Data:" caption | All figures now read data.py / the JSON's figure_params; x-positions computed from survey dates |
| E24 | fig1 axis label false for sign-normalized bars; fig4 merged Dittmar+Clark into one line (the exact merger data.py condemns); fig3 title "255 years" vs JSON "~260" | Relabeled + marked; two labeled series; 260 harmonized |
| E25 | Internet counted both as stalled failure (§3.4) and completed success (§7 base rate) | Base rate recomputed: 3 completed + internet pending |
| E26 | §4.5 conditions conjunctive ("and"), (d) unoperationalized, 2030 mis-scoped | Rewritten: individually damaging; (a)+(d) jointly fatal; proxies named; 2030 scope stated |
| E27 | "External adversarial review" (v3 frontmatter) — the review was AI, commissioned by the author | Disclosure rewritten in plain terms; "not a substitute for external human review" added |

## 7. Standing rules (inherited + trial-added)

1. No number in text/tables/figures unless it lives in data.py with source
   + status; the numeral audit file is the enforcement surface.
2. Never pool the study effect sizes; the claim is the gradient.
3. The identity is an instrument, never evidence; "predicts" is reserved
   for the staked §4.5 indicators.
4. The institutional-compression ratio stays REFUSED in the JSON.
5. Old build scripts are never edited; a new draft is a new copy.
6. U-tagged entries appear only with hedge attached, never in headline
   claims — enforced by shame of E20.

## 8. Open items

- VERDICT order 23 (RECOMMENDED): external human review before circulation
  beyond the project. **Not yet done; disclosure says so.**
- The 2028 check on Canaries (does the entry-level gap close?).
- Someone should actually measure g/v and m — the paper's own stated
  research target.
- Next trial (v4 → v5) only when a reader or the record produces new
  charges; the standing attack surface is §4 above.