# Adversarial Audit — Report
**Roman Economic & Production Centers / SEDZ project** · audit run 2026-07-07 · 10 domain-specialist seats + verification
Panel: Roman economic history · Roman law/juristic · mining archaeology · marble/quarry · Greek/Hellenistic economy · comparative-SEZ economics · epigraphy/papyri data · quant/DH methods · GIS/cartography · completeness/selection-bias. ~50 raw findings → deduped + verified below. Brief: `AUDIT-BRIEF.md`.

## Disposition summary
- **CLEAN (independently confirmed sound):** (a) `comparanda-matrix.json` matches the entries exactly; the genus rule (all 3 convergence ≥1) holds for all 23 with zero violations (quant seat). (b) **P5's citation fixes HOLD** — the epigraphy seat independently re-verified Dokimeion=ILS 1477, lex Hadriana 25943/26416, Souk-el-Khmis 10570=14464 (Kehoe "14454"=typo), Possessor CIL II 1180=ILS 1403=AE 1965 237, Noricum 4809, Alburnus pp.948-949; **no fabricated corrections**. (c) Map coordinates spot-check correct; non-spatial cases correctly dropped (GIS seat).
- **FALSE POSITIVE (flagged but wrong — no change):** the Roman-law seat suspected the Vipasca "customs-like duty on ore-concentrates entering the fines" was fabricated; it is **attested** — Vipasca I ch. 7 (scriptura scaurariorum et testariorum): 1 denarius/100 lb on *rutramina* brought *in fines metallorum* from outside (Domergue 1983). Keep; strengthen the citation to name Vip. I.7.
- **CONFIRMED — actionable** (below).

The audit's verdict in one line: **a sound data layer (clean matrix, verified citations, correct geodata) wrapped in over-claimed inference.** The corrections mostly *soften claims and disclose limits*; none touches the verified evidentiary base.

---

## TIER 1 — Major, multi-seat convergent (highest priority)

**A1. The temporal-arc "ice-core confirmation" is an overclaim, and likely backwards.** *(seats: econ-history, mining-archaeology, quant/DH, +law/coverage — 3–4 seats)*
- The claim (D2 §6; P4 doc; finding-page caption) that mine-attestation activity "independently reproduces the Greenland ice-core lead record" is false as stated: McConnell et al. 2018 (PNAS) put the lead peak in the **1st c. AD** with a **plunge at the Antonine plague (~AD 165)**, whereas the project's curve peaks **AD 200**. The two diverge exactly where "confirmation" is claimed.
- The curve is a **statistical artifact**: only 230/1,399 mines are doubly-dated (16.4%); median open→close span **335 yr**; **94% of dated mines are "active" at AD 200** — the hump is interval-overlap geometry, not synchronic activity. Dates are coarse *conquest-block period-floors* (−330, −30 Augustan, 106 Trajanic). It also duplicates the **epigraphic-habit** curve (MacMullen peak ~150–225, which the project's *own* epigraphic-habit finding records) — so it may measure documentation, not production. "Independent" is further undercut by shared source chronologies (Domergue's Iberian dates drive both). Attribution error: "Domergue's ice-core curve" — the record is Hong 1994 (Science) / McConnell 2018.
- **Fix:** delete "independently reproduces"; reframe the arc as attestation-density (epigraphic habit), state the 230/1,399 denominator + 94%-simultaneity + conquest-floor caveats, or drop it as evidence. **[apply now — D2, P4, finding page]**

**A2. Selection bias — the 23 cases are a curated gradient with no hard case.** *(seats: econ-history, coverage, marble, mining-arch — 4 seats)*
- Every genus member is a tightly-administered imperial mine/quarry (10/11) known a priori to be dirigiste; every failure is a private export industry known a priori to be market-run. Finding (i) ("every Roman member is dirigiste/state-fiscal") is baked into the sampling frame (marble set anchored entirely on Hirt 2010 = a book about *imperial* quarries), not discovered.
- The decisive omission: a **Republican *societas publicanorum* mining district** — e.g. **Sisapo** cinnabar/minium (Pliny NH 33.118–122: state-owned, run by the *societas Sisaponensis*, *vena signata*, price fixed HS 70/lb): bounded + allocated + fiscal on the rubric (a genus member) yet organized around *publican contractor profit* — a market-polarity Roman member that would break the "single dirigiste species."
- **Fix:** soften finding (i) to "the **high-imperial procuratorial** metallum is dirigiste"; score ≥1 Republican-publicani district and ≥1 civic/private quarry (Aphrodisias/Pentelicon) as controls; state the marble frame is an imperial-administration sample. **[soften now; new cases = queued]**

**A3. The scoring rubric is uncalibrated soft-coding presented as measurement.** *(seats: quant/DH, marble)*
- Magnitudes were assigned by single (AI) raters from "the evidence actually read," with **no codebook thresholds, no second rater, no inter-rater reliability** (the P1-promised blind second-scoring never happened). The 12-vs-11 genus split hinges on fragile 0/1 calls: **Luna** qualifies only via `fiscal_monopoly=1` (its note says "approaches 0"); **Proconnesus** fails only via the same axis at 0; **Luna's `bounded_jurisdiction=1` directly contradicts its own note** ("no differential lex, no jurisdictional perimeter") — the consistent score is 0, which flips Luna to genus-fail (12→**11/23**).
- **Fix:** disclose the rubric is expert-judgment coding, not measurement; publish an anchored codebook; flag every membership that flips under ±1 (luna, proconnesus, karystos, noricum); correct Luna to BJ=0/genus-false. Reframe "each axis independently necessary" (n=1 per axis) as "illustrated." **[Luna fix + disclosure now; codebook/2nd-rater = queued]**

**A4. The "regime-direction" axis is too blunt; "dirigiste" vs "liberalizing" overstate both poles.** *(seats: econ-history, comparative-SEZ)*
- The metallum works through *farmed/auctioned/concession* instruments (centesima auction-tax, service-monopoly auctions, occupatio) — delegated, market-mediated revenue extraction, not central planning; "dirigiste" overstates state direction. Conversely the modern SEZ is a **state-led industrial-policy** instrument (dirigiste in construction/planning), not laissez-faire; coding it pure "liberalizing" caricatures it. The "polar inversion" holds only on *specific margins* (fiscal/labor/customs), not as a total mirror.
- **Fix:** reframe regime_direction as *which margin is relaxed vs intensified*; qualify the inversion to those margins; note the metallum's auction/concession market-internals. **[D2 revision — apply]**

**A5. The modern side is n=1, and Shenzhen is the worst exemplar + mis-scored.** *(seats: comparative-SEZ, coverage)*
- The headline "polar inversion" rests on a single modern case; Shenzhen is the most *state-directed* SEZ (weakest possible "liberalizing/footloose-capital" exemplar). Its `bounded_jurisdiction=2` is a **scoring error** — the rubric's 2 requires *civic displacement* (extra civitatem, no ordo), but a modern SEZ is a regulatory layer *over* intact municipal government → should be **1** (still qualifies genus at 1/1/1).
- **Fix:** rescore Shenzhen BJ→1 [now]; add ≥1 classic export-processing zone + a free port before the type-level "the modern SEZ" claim [queued]; until then phrase as "Shenzhen, taken as an SEZ exemplar." Cite **Farole 2011 / FIAS 2008** — their independent structural SEZ definition (delimited area + single administration + separate customs area) matches the 3 convergence axes and *defuses the circularity objection*. **[D2]**

---

## TIER 2 — Major, single-seat (verify/fix)

**B1. Penal labor is over-read at the type specimens.** *(law, marble, Greek-econ)*
- **Vipasca**: attests *servile concession-discipline* (flogging for breach) + slaves-banned-from-mines; *damnati ad metalla* at Vipasca specifically is imported from the general Digest 48.19 category — yet the abstract/matrix make "mixed+penal" a defining feature (the same caution the project rightly applies at Mons Claudianus is not applied here). **Simitthus**: `labor=2 (penal)` — the district's signature — rests on the *ergastulum* reading, which **Mackensen** (Simitthus III, 2005/07: "Militärlager oder Marmorwerkstätten") contests (possibly marble workshops); no epigraphic *damnati* from Simitthus. **Laurion**: coded `convict_coerced_penal` but the entry's own note says "chattel slavery, not judicial damnatio" — a category error (chattel = market-supplied capital, Nikias' 1,000 slaves rented for profit; ≠ penal sentence).
- **Fix:** Vipasca — qualify penal as general/inferred not site-attested; Simitthus — downgrade to 1/conditional, cite Mackensen; Laurion — relabel the pole "chattel-slave (coerced, non-penal, market-supplied)". **[entries — apply]**

**B2. The "closest cousin" claim for Laurion is inflated.** *(Greek-econ)*
- Stripping the chattel≠penal conflation, Laurion shares only **two** clean polarities with Vipasca (dirigiste, state-fiscal) — the *same two the Ptolemaic monopoly shares, which the paper says fails*. Structurally Laurion is a competitive *private market-auction* lease with disputes in ordinary *dikai metallikai* (BJ=1). "Closer than the SEZ" is true only on a 4-equal-weight polarity count. **Fix:** soften §3(iii) to "shares regulatory *sign* on 2–3 axes but diverges structurally." **[D2]**

**B3. "The enclave is rare" is asserted, not measured.** *(coverage, mining-arch, quant)*
- §6/Abstract claim rarity, but the 1,399-mine denominator is never scored on the rubric. And that corpus is skewed (gold 828 = 59%; Iberian 72%; lead 60% dated vs copper 2.2%). **Fix:** score a stratified subsample to get a base rate, or delete "rare/most… ordinary" and restate as a definitional point. **[queued/D2]**

**B4. "The genus is not mining-specific" rests on one case.** *(coverage, SEZ, marble)*
- Only Henchir Mettich (1/2/1) is a non-extraction genus member; the genus is empirically 91% mining. **Fix:** demote finding (iv) to "appears at least once in agriculture; manufacture/processing untested," or add ≥1 more non-extraction candidate (Tiber-valley imperial figlinae, a salina, a temple estate). **[D2 + queued]**

**B5. Map attribution/licensing is missing on the deployed page.** *(GIS)*
- `magalia-map-kit.js` hardcodes `attribution:''`, so the deployed finding shows no Esri/DARE credit on the control — breaching Esri ToS + DARE CC BY 4.0 (only the standalone synthesis.html attributes). DARE credit should read "Johan Åhlfeldt (Univ. of Gothenburg), CC BY 4.0" (Pelagios is a network, not the provider). Esri Physical/Terrain are **modern** coastlines/relief, disclosed only as "borderless." **Fix:** add `attribution` to the map-kit BASES + caption; add the modern-geography caveat; fix the DARE credit. **[apply — map-kit + page]**

**B6. The epigraphic production index is method-flawed and unused.** *(quant, epigraphy)*
- 86% generic (`officina` 63% + `figlin` 23%); substring matching false-positives (`salin`→*Salinator* cognomen, `purpur`→*purpureus*, `metall`→*Metellus*/damnatio/Christian, `argentari`→banker not silver); and it is cited by *no entry and no line of D2*. **Fix:** relabel as an *instrumentum-survival* index (not production-intensity), filter to production genres (tituli fabricationis), and mark exploratory/unused. **[dataset note — apply]**

---

## TIER 3 — Minor / concrete data fixes (apply)
- **C1. Simitthus misattribution:** "20 Numidian columns to Smyrna" cited to "Pausanias 1.18.9" (that passage = Hadrian's *Athens* pantheon/gymnasium) → replace with **IGRR IV 1431** (the same inscription already cited in the Dokimeion entry for the 72 columns). *(marble)*
- **C2. Vipasca `fiscal_regime`** = `conductor_lease` contradicts `domergue_regime=regie_indirecte`; Vipasca is *the* régie-indirecte exemplar (occupatio); conductor_lease is Noricum → relabel to a régie-indirecte-consistent value; keep "conductor" for the *service*-monopoly farming. *(mining-arch)*
- **C3. Mons Claudianus** `primary_sector=building_stone` vs Mons Porphyrites `marble_decorative_stone` — same district/regime/prestige role; granito del foro is a prestige stone → retag to marble_decorative_stone (or add monumental_hardstone). Amend evidence[2] so I.Pan 42 is dedicated by **Epaphroditos** (Chresimus-in-this-text is disputed, per P5). Fix place-name **Tricomia → Troknada/Trocnades (Kaymaz)** in dokimeion.json. *(marble)*
- **C4. portorium-asiae** "P. Terentius (1 BC)" → **c. 48 BC** (Cic. Att. 11.10 = 48 BC; Terentius Hispo is late-Republican). *(law)*
- **C5. Noricum `allocated_exploitation=2`** contradicts its own note (fisc farms the vectigal; occupatores self-occupy) → downgrade to 1 (→ 1/1/1, still genus, more marginal). *(mining-arch)*
- **C6. les-martys** `primary_sector=gold` — it's Domergue's Roman *iron* ferrier; the gold-first priority mis-ranks multi-metal sites (inflating gold=828) → re-derive/flag primary_sector for multi-metal mines. Laurion carries a Roman-province "Achaea" stamp + null dates (it's pre-Roman) → footnote. *(mining-arch)*
- **C7. Ptolemaic monopoly** "purely aspatial" overstates (per-nome tax-farm + worker confinement); keep BJ=0 on the correct ground (uniform ordinary royal admin, no *differential*-law enclave). **Laurion** lease-terms: *ergasimon* IS in Ath.Pol. 47.2; only *kainotomia*/*anasaximon* are stelai-specific. *(Greek-econ)*
- **C8. Aljustrel** appears twice (mines_attested undated/surveyed/silver-copper-iron vs vipasca.json excavated/silver/Hadrianic) → note the dataset carries raw OXREP fields only. *(mining-arch)*

## TIER 4 — Representation / framing (D2 wording)
- **D1.** "Enclave economy" is equivocated: the dependency-theory term means export-extractive + weak linkages + foreign capital, not a bounded differential-law jurisdiction → engage it properly or coin a distinct term. *(econ-history)*
- **D2.** The field is relegated to the appendix — Hopkins "Taxes and Trade" (JRS 1980), Wilson (OXREP metal-production output, the actual scholar linking mining to the ice core), Finley/Bang — belong in the body. *(econ-history)*
- **D3.** The paper silently adopts Hirt's contested *fiscus=aerarium=state* equation the entries flag as disputed → add a hedging sentence (scoring unaffected — both readings sit on the state pole). *(law)*
- **D4.** §3(iv) "extra civitatem" for Henchir Mettich overstates what the entry scores as partial (saltus extra-civic status is debated) → soften to match the entry. *(law)*
- **D5.** The cumulative §5 concessions (function inverts + geology-not-policy + Development bracketed) hollow the thesis toward a thin residuum → state the surviving claim's thinness and defend that production+fiscal-instrumentation is non-trivial; treat "Development" as a substantive limitation. *(SEZ)*

---

## Remediation plan

### ✅ APPLIED 2026-07-07 (verified)
- **A1** — ice-core "confirmation" WITHDRAWN in D2-PAPER §6, P4 doc (§2 heading + body + banner), and the deployed finding page caption (EN+中文); reframed as an epigraphic-habit artifact (94% simultaneity, 16% subsample, ice-core actually peaks 1st c. AD / falls ~AD 165). Finding page browser-verified: "independently reproduces" gone, caveat present.
- **A3 (Luna)** — `bounded_jurisdiction` 1→0, `genus_qualifies`→false (matches its own note); **genus count 12→11/23**; matrix + finding page (KPI=11, Luna row 0/1/1 ✗) + D2 matrix row corrected. Rule-consistency re-verified.
- **A5 (Shenzhen)** — BJ 2→1 (still genus 1/1/1); entry + matrix + finding page + D2 row corrected.
- **B5 (map attribution)** — added `attribution` to `magalia-map-kit.js` BASES (Esri + DARE CC BY 4.0); finding-page caption gets the Esri modern-geography caveat + full DARE credit (Åhlfeldt / Univ. Gothenburg / CC BY 4.0). Verified: control now shows "Tiles © Esri".
- **C1** Simitthus Pausanias 1.18.9 → IGRR IV 1431. **C2** Vipasca `fiscal_regime` conductor_lease→direct_fisc. **C3** Mons Claudianus primary_sector→marble_decorative_stone; Dokimeion Tricomia→Troknada/Trocnades (Kaymaz). **C4** portorium P. Terentius (1 BC)→Hispo c. 48 BC.
- All 23 entries re-validated (0 invalid); D2 abstract "demonstrated"→"illustrated"; audit banner added to D2-PAPER.md.

### ✅ QUEUE WORKED THROUGH 2026-07-07 (round 2)
- **B1/B2/C5/C6/C7/C8** all applied: Simitthus penal 2→1 (Mackensen); Vipasca penal qualified; Laurion labor penal→**mixed** (chattel≠penal) + "closer cousin" softened; Noricum AE 2→1; Les Martys gold→iron; Ptolemaic per-nome note; Aljustrel dataset note.
- **B6** — epigraphic index relabelled an *instrumentum-survival* index (`datasets/epigraphic_production_index.README.md`): 86%-generic disclosed, keyword false-positives named, genre-filtered diagnostic counts given.
- **A3** — `schema/SCORING-CODEBOOK.md`: anchored 0/1/2 decision rules per axis + fragile-boundary flags. (Blind second-rater κ still outstanding — noted in the codebook.)
- **B3** — computed: all 1,399 mines are `surveyed` with zero administrative attestation, so "rare" is a statement about attestation, not a measured rate. Folded into D2.
- **A2 (the killer test) — DONE + material result:** **Sisapo** (Republican *societas Sisaponensis* cinnabar; Pliny NH 33.118–122) scored **1/2/2, genus TRUE but on the MARKET pole** (beneficiary = fisc+publican; capital = *censoria locatio* auction). → **finding (i) rewritten: the genus is a spectrum, not a single dirigiste species-cluster.** **Aphrodisias** civic quarry **0/1/0 FAIL** (proves the marble set's hit-rate was a Hirt-2010 imperial-sampling artifact).
- **B4 — DONE:** **Tiber figlinae** (imperial brick, max-ownership) **0/2/1 FAIL** — imperial ownership is proprietary not jurisdictional; sharpens finding (iv) (genus turns on the *extra-civitatem* apparatus, not ownership). Henchir Mettich stays the lone non-extraction member.
- **A5 — DONE:** **Mauritius EPZ 1/1/1 TRUE** (clean World-Bank exemplar) + **free port 1/0/1 FAIL** — the free port fails on production exactly like the *portorium* (opposite polarity), isolating `allocated_exploitation` on BOTH poles → strengthens the thesis. Modern liberalizing pole now n≥3.
- **D1–D5 — DONE in D2:** enclave-economy sense distinguished; Hopkins/Wilson/Finley/Bang engaged in the body; **Farole 2011 / FIAS 2008** cited (defuses circularity); fiscus hedge; thesis-thinness stated. Matrix +5 rows; abstract 23→28 candidates; §4 mirror robustified (EPZ) + spectrum note.
- **Sample now 28 candidates / 13 genus.** Finding page (browser-verified) + comparanda-matrix.json + all docs regenerated. All 28 entries validate.

### ✅ A3 IRR — DONE 2026-07-07 (`_audit/IRR-REPORT.md`)
Blind independent second rater scored all 28 from the codebook alone: **linear-weighted Cohen's κ = 0.87 / 0.84 / 0.82** on the three convergence axes, **genus κ = 0.85**, within-1 = 100% ("almost perfect"). The rubric is now a **reproducible criterion-explicit instrument**, not single-rater coding. The only 2 genus disagreements are **Sisapo** and **Laurion**, both at the `bounded_jurisdiction` 0/1 line (state regime inside a functioning city) — flagged **calibration-fragile** in D2 + a codebook tie-breaker added. Schema gaps also closed: `fiscal_district` center_type (portorium + free port retagged), `market_lease` + `chattel_slave_market` polarity values (Sisapo/Laurion retagged).

### ⏳ Still open (genuinely future work)
- A **3rd rater** for a Krippendorff's-α (the 2-rater κ is solid; two further blind raters stalled on environment instability). Adjudicate the Sisapo/Laurion BJ boundary with a domain panel.
- Full sentence-level polish of every downstream D2 clause; deploy (`./push.sh`, user's call).
**Queue (substantial new scholarship):** A2 (add Sisapo publicani + a civic quarry control), A5 (add 2nd/3rd modern SEZ), A3 (anchored codebook + blind second-rater + κ), B3 (score a mine subsample for a base rate), B4 (add non-extraction genus candidates), D1–D5 (D2 body revisions: engage Hopkins/Wilson/Finley, cite Farole/FIAS, enclave-economy term, fiscus hedge, thesis-thinness).

**Net:** the audit found **no fabrications and no broken evidentiary base** — the citations and matrix survived. What it found is a pattern of **inferential over-reach** (the temporal arc, the ice-core "confirmation," an uncalibrated rubric sold as measurement, a curated sample generalized to "the Roman metallum," and an n=1 modern mirror). Every Tier-1 issue is fixable by softening the claim to what the evidence supports; doing so makes the work more defensible, not less.