Method
How the lab works
AI agents read and argue under rules published in advance. Humans do the digging.
26Drafted
26Reviewed
100Target
- 01ScoutFinds candidate anomalies — famous and forgotten — with two real sources each.
- 02ResearcherDrafts one record from the literature; every source checked.
- 03SkepticAttacks the draft: citations, mundane explanations, hindsight, overconfidence.
- 04AnalystTests the pre-registered hypotheses on reviewed records only.
- 05Field partnerA credentialed archaeologist tests the best leads: new dates, re-examination, fieldwork.
Principles
- Evenhanded. Anomalies are checked as hard as orthodoxy.
- Mundane first. Contamination, dating error, mixing, misidentification, hoax and recording error are ruled out first.
- Falsifiable. Every record and finding says what would disprove it; hypotheses are registered before the data.
- No fabrication. Every claim cites a real source that was actually checked.
- Honest uncertainty. Disputed cases stay disputed; small samples get small claims.
- Heritage ethics. Country-level locations only; no coordinates; fieldwork only by permitted professionals.
Checks on every record
- Schema. Required fields, allowed values and hard length limits; the build fails on any violation.
- Crossref. Every DOI is matched by script against its title, year and authors.
- Source labels. ● checked opened and confirmed · ◐ exists metadata only · ○ unverified.
- Skeptic review. An independent agent tries to break each draft; issues go back to the researcher.
- Hindsight guard. Evidence is coded as it stood at the time of the claim; a blind re-coding test is planned (M1).
- Starter briefs corrected. Each researcher starts from a brief written from general knowledge. So far 76 corrections across 25 of 26 records (about 2.9 per case) — why every claim is re-verified.
Agents
Each role has its own model, effort and tools. The analyst has no web access, so it can only use the checked catalog.
| Role | Job | Model · effort | Tools |
|---|---|---|---|
| Scout | Finds candidate cases to expand the Stratum catalog toward 100+, from review articles, histories of paleoanthropology, and books on frauds and disputed sites. | sonnet · medium | Read, Write, Edit, Bash, Grep, Glob, WebSearch, WebFetch |
| Researcher | Researches ONE Stratum catalog case from published sources and writes (or revises) catalog/cases/<id>.json. | opus · high | Read, Write, Edit, Bash, Grep, Glob, WebSearch, WebFetch |
| Skeptic | Adversarial reviewer for ONE Stratum draft record. | opus · high | Read, Write, Bash, Grep, Glob, WebSearch, WebFetch |
| Analyst | Tests the pre-registered hypotheses and looks for patterns across the verified Stratum catalog (counts, cross-tabs, exact tests; regression only once n is large enough). | opus · high | Read, Write, Edit, Bash, Grep, Glob |
Progress
Records drafted26 / 24
Skeptic-reviewed26 / 24
Toward 100 cases26 / 100
- Phase 1 · now. Anomaly fates: the catalog, the lessons, the leads.
- Phase 2. Radiocarbon residue: dates that contradict their site once mundane causes are removed.
- Phase 3. Grey literature: unpublished excavation reports, read at scale.
- Phase 4. Landscape: open LiDAR and imagery; findings go privately to heritage authorities.
- Continuous. Partnership: ranked leads to a field archaeologist or dating lab.
Costs approximate tokens per batch
| Date | Batch | Agents | Tokens | Notes |
|---|---|---|---|---|
| 2026-09-23 | Wave 1 · 10 researcher drafts | 10 × researcher (Opus, session effort) | ≈ 2.96M (270–340k each; 69–105 tool calls; ~21–26 min) | Before the lookup tool: raw API JSON and whole PDFs read into context. Budget stop sent mid-run. Monte Verde got a targeted exception for its 2026 re-contest. |
| 2026-09-23 | Wave 1 · 10 skeptic reviews | 10 × skeptic (Opus) | ≈ 1.50M (139–175k each; 27–37 tool calls; ~11–16 min) | With lookup.py and a 35-call budget. Verdicts: 1 pass, 9 revise, 0 fail; no fabricated sources. Majors were hindsight codings, misattributions, status calibration. |
| 2026-09-23 | Wave 2 · 14 researcher drafts (trial) | 14 × researcher (Sonnet), with lookup.py | ≈ 2.4M (148–203k each; 46–94 tool calls) | Skeptic major issues averaged ≈ 6.6 per record vs ≈ 2.7 for Opus drafts, incl. misreadings and false corrections → researchers return to Opus (D17). |
| 2026-09-23 | Wave 2 · 14 skeptic reviews | 14 × skeptic (Opus) | ≈ 2.1M (135–159k each; 26–38 tool calls) | 0 pass, 14 revise, 0 fail; no fabricated sources; one invented detail (Hueyatlaco). |
| 2026-09-23 | Revisions, waves 1–2 | 14 × Sonnet + 9 × Opus revisers | ≈ 3.4M (105–176k each; 11–35 tool calls) | Opus used where drafts had misread sources. All 24 starter records now reviewed. |
| 2026-09-23 | Discovery track R1 + megalith audit | 1 × Opus scan, 2 × Sonnet auditors | ≈ 0.50M | R1 null; audit under Opus review. |
| 2026-09-23 | Hancock cases (Sphinx, Gunung Padang) | 2 × researcher (Sonnet), 2 × skeptic (Opus), 2 × reviser (Opus) | ≈ 0.98M | Sphinx open (enclosure never directly dated); Gunung Padang refuted (retracted 2024). |
| 2026-09-23 | Discovery track R2 + triage; megalith audit review | 2 × Opus (scan, triage), 1 × Opus (audit review) | ≈ 0.70M | R2: no genuine leads after triage. H7 consistent (35 of 40 megaliths dated only by association). |
| 2026-09-23 | **Session total (agents)** | ≈ 100 agent runs | **≈ 15M** | Excludes the orchestrating main session. |
| 2026-09-30 | Scout for batch 3 (resolved cases) | 1 × scout (Sonnet · medium) | ≈ 0.05M (11 tool calls) | 15 candidates, all chronology, none from North America: 7 vindicated, 5 refuted, 2 unsupported, 1 open. All 30 DOIs resolve to the cited works (4 page errors fixed). Main session added 2 refuted candidates from the owner's list (Old Crow flesher, Del Mar/Sunnyvale). |
Open data
Everything on this site is generated from these files.