Pharmakon

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

2,000 years of Greek plant medicine, measured.
An uncertainty-aware map from Dioscorides' 1st-century Greek to modern species, chemistry and untested research leads.

[!IMPORTANT] Research hypotheses only, not medical advice. Pharmakon generates research hypotheses from historical texts and open databases. Nothing here is medical advice, a recommendation to use any plant, or evidence of safety or efficacy.

Watch the film

**[▶ Watch the film on YouTube](https://youtu.be/d-kbvFd2K04)** (9 min) The problem, the data, the method, a live demo of the site and of the MCP server inside Claude, and every result, including the one that did not work. **[Pitch deck](https://drive.google.com/file/d/1cX2WHtz4g6Yu8RxI-wfN-3ifHq9VsYnv/view?usp=sharing)** (Google Drive) **[Slides, PDF](docs/slides/Pharmakon_EthnoHack2026_slides.pdf)** (10 slides, built from the pipeline outputs) **[Live site](https://pharmakon.adoranto737.workers.dev)** · **[MCP endpoint](https://pharmakon.adoranto737.workers.dev/mcp)** · **[Latest release](https://github.com/LSUDOKO/Pharmakon/releases/latest)**

In one minute

Around 60 AD, the Greek physician Dioscorides wrote De materia medica, about 600 medicinal plants and what each one treats. Doctors copied it for 1,500 years. Three problems stop anyone turning it into testable science:

1. Nobody is sure which plant each ancient name means. Pharmakon gives a probability for every candidate species, with an explicit "unidentified" option. It agrees with expert botanical historians 62% of the time; the standard Greek dictionary manages 31%.
2. Nobody has measured, at scale, which uses survived. Pharmakon follows every use through Galen (2nd c.), John the Physician (13th c.) and modern Greek villages. 49% of uses recur today, against 40.8% expected by chance (p = 0.001).
3. Herb–disease predictions are rarely tested honestly. Pharmakon runs a strict time-travel test across all body systems and reports a null result: history did not predict which pairs science studied next beyond research popularity.

The output is 1781 untested plant–condition leads, each weighted by how sure we are of the plant, with toxicity and conservation flags. Ignoring that uncertainty would change 5 of the top 20.

Use it three ways: the website, the MCP server for AI assistants, or the JSON API. All three are below.

Project status

part	status
Pipeline: 13 CLI stages, Parquet with provenance on every row	done
Identification model, persistence test, temporal backtest, leads	done, results below
Website with explorer, charts and leads gallery	live on Cloudflare
MCP server (6 read-only tools) and JSON API	live on Cloudflare
Streamlit research dashboard, 27 tests, CI on every push	done
Calibration (ECE 0.21 vs target ≤ 0.10)	missed, cause identified below
Human audits of extraction, conditions, merges, labels	pending: blinded sheets generated in data/gold/audits/
LLM claim extraction, LLM lexicon extension, plant-morphology feature, LLM-judge baseline	done (Claude session + Groq), every span re-verified; kept out of the backtest
All body systems (beyond the skin/wound pilot)	done: PubMed labels and chemistry re-run for every use group
Slide deck (10 slides, PDF)	done: docs/slides/Pharmakon_EthnoHack2026_slides.pdf, built from the pipeline outputs by docs/slides/build_slides.py
Pitch film (9 min) and pitch deck	done: YouTube · deck

A worked example

What the pipeline produces for one name, wormwood:

ἀψίνθιον (apsinthion), De materia medica 3.23 · urn:cts:greekLit:tlg0656.tlg001.1st1K-grc1:3.23.1

ἀψίνθιον· γνώριμος ἡ πόα. ἐστι δὲ αὐτοῦ βέλτιον τὸ ἐν Πόντῳ καὶ Καππαδοκίᾳ γεννώμενον ἐν ὄρει τῳ καλουμένῳ Ταύρῳ. δύναμιν δὲ ἔχει στυπτικήν, θερμαντικήν, ἀποκαθαρτικὴν τῶν ἐνηρεικότων στομάχῳ καὶ κοιλίᾳ χολωδῶν. ἐστι δὲ καὶ οὐρητικὸν καὶ ἀκραίπαλον προπινόμενον, καὶ πρὸς ἐμπνευματώσεις δὲ ἁρμόζει καὶ κοιλίας καὶ στομάχου ἀλγήματα πινόμενον μετὰ σεσέλεως ἢ νάρδου Κελτικῆς, καὶ ἀνορεξίας καὶ ἰκτερικούς θεραπεύει τὸ ἀπόβρεγμα αὐτοῦ ἢ τὸ ἀφέψημα καθʼ ἑκάστην ἡμέραν λαμβανόμενον εἰς πλῆθος κυάθων τριῶν.

The model's answer:

candidate	family	P(species)	90% interval
Artemisia absinthium	Asteraceae	0.75	0.62–0.88
(unidentified)		0.23	0.10–0.36
Artemisia monosperma	Asteraceae	0.02	0.01–0.05

The uses through time:

use	Dioscorides	Galen	John the Physician	modern Greece (expected)
Blood, spleen	attested	absent	—	0.00
Cardiovascular	attested	absent	—	0.00
Skin & wounds	attested	attested	—	0.77
Fever	attested	absent	—	0.00
Digestive	attested	absent	—	0.00
Gynaecology	attested	absent	—	0.00
Liver & gall	attested	absent	—	0.00
Metabolic	attested	absent	—	0.00
Eyes	attested	attested	—	0.00
Ears	attested	absent	—	0.00
Nose & throat	attested	absent	—	0.00
Urinary	attested	absent	—	0.00

The highlighted evidence spans, photos and the Galen passage for this name are shown in the explorer:

Still image

Results

1. Which plant is it?

Each of the 61 names identified by Lardos et al. (2024) is held out in turn; the model ranks GBIF-verified candidates using geography, name continuity, scholarly agreement and genus consensus, with no pharmacology at all.

model	top-1 [95% CI]	top-3	mean reciprocal rank	Brier ↓	ECE ↓
Pharmakon (calibrated)	0.62 [0.51, 0.75]	0.92	0.78	0.50	0.21
Pharmakon, before temperature scaling	0.62 [0.51, 0.75]	0.92	0.78	0.50	0.14
Transfer model (no Lardos counts; used for all names)	0.56 [0.43, 0.69]	0.92	0.74	0.55	0.15
Pharmakon + f1 morphology (LLM traits)	0.62 [0.51, 0.74]	0.93	0.78	0.50	0.14
Transfer model + f1 morphology	0.59 [0.48, 0.72]	0.89	0.74	0.54	0.16
…without the ethnobotanical-reference count	0.62 [0.51, 0.74]	0.90	0.77	0.51	0.18
…without geography (f2)	0.66 [0.54, 0.79]	0.92	0.79	0.54	0.24
…without name continuity (f3)	0.62 [0.51, 0.75]	0.92	0.78	0.49	0.18
…without scholarly agreement (f4)	0.54 [0.41, 0.67]	0.90	0.72	0.55	0.14
…without genus consensus (f6)	0.66 [0.54, 0.79]	0.92	0.80	0.49	0.20
LSJ dictionary baseline	0.31 [0.20, 0.44]	0.80	0.57	1.34	0.69
Genus-frequency prior	0.38 [0.26, 0.51]	0.87	0.62	0.59	0.09
Uniform guess	0.39 [0.28, 0.52]	0.87	0.63	0.63	0.12
LLM judge (Claude session, Greek text + candidate list; comparison only, may have memorised)	0.59 [0.46, 0.71]	0.89	0.75	0.55	0.10
Oracle: + expert-coded traits (ablation only)	0.84 [0.74, 0.92]	0.98	0.91	0.22	0.09

• The model beats the dictionary baseline on top-1 and on Brier score.
• Removing the one feature that overlaps the experts' own tie-break (the ethnobotanical-reference count) leaves top-1 unchanged, so the headline does not rest on it. The transfer model uses no Lardos-derived counts and is the cleanest number.

What the model learned (conditional-logit weights; the "unidentified" option has its own utility):

feature	what it measures	weight (full)	weight (transfer)
f2_geo_log_occ	log(1 + GBIF occurrences in GR+CY+TR)	+0.22	+0.38
f2_present_gr	≥1 GBIF occurrence in Greece	+0.69	+0.68
f3_vernacular_sim	max similarity of ancient name to modern Greek vernacular names (Wikidata P1843@el)	+0.03	+0.02
f3_epithet_sim	max similarity of Latinised ancient name to genus / epithet	+0.08	+0.23
f4_lit_dmm	log(1 + # botanical-history references proposing the taxon) [Lardos file 5]	+0.47
f4_lit_ethno	log(1 + # ethnobotanical references linking the name to the taxon) [Lardos file 5]	+0.35
f4_lsj	LSJ gives the binomial (1 asserted, 0.5 hedged)	-0.29	-0.31
f6_genus_consensus	share of the unit's candidates in the same genus	+0.18	+0.15
f1_morph	morphological agreement: Dioscorides' description vs the species' Wikipedia description (LLM-extracted traits, spans verified)	+0.18	+0.25

2. Did the uses survive?

• Galen re-attests 37% of Dioscorides' uses for the names that could be linked to his text, John the Physician 34%.
• Galen is not copying: mean word 4-gram containment against Dioscorides is close to zero.

Survival by body system (Dioscorides uses whose plant could be identified):

use group	persisted to modern Greece	to Galen only	to John the Physician only	modified	lost
Skin & wounds	74	51	0	13	77
Digestive	34	9	1	19	60
Gynaecology	16	12	0	42	52
Urinary	25	11	2	20	40
Muscles & joints	15	7	0	20	41
Respiratory	22	7	2	10	33
Fever	7	3	0	19	36
Eyes	5	3	1	25	25
Liver & gall	12	4	0	11	18
Mouth & teeth	6	5	2	16	13
Pregnancy & birth	5	6	1	10	19
Blood, spleen	4	6	0	10	18
Cardiovascular	6	0	0	7	19
Neurological	6	2	0	7	15
Other	5	1	0	10	14
Ears	4	2	0	13	8
Nose & throat	1	1	0	4	6
Infections	0	0	0	4	4
Metabolic	0	0	0	1	1

3. The time machine

Stand in 2000, 2005, 2010 or 2015, forget everything published after, rank the plant–condition pairs nobody had studied yet, then check PubMed up to 2025. Features are gated by date: compound reference years, ChEMBL document years and modern-study publication years must all precede the cutoff, and no LLM output is used.

cutoff	test pairs	gained support by 2025	Random	Popularity	History	History + popularity	Everything	Δ history [95% CI]
2000	1404	122	0.08	0.38 [0.31, 0.47]	0.13 [0.09, 0.18]	0.36 [0.29, 0.46]	0.35 [0.28, 0.44]	-0.08 to +0.03
2005	1384	102	0.08	0.34 [0.28, 0.45]	0.08 [0.06, 0.12]	0.27 [0.22, 0.36]	0.22 [0.16, 0.30]	-0.14 to -0.02
2010	1361	79	0.06	0.31 [0.24, 0.42]	0.07 [0.05, 0.10]	0.29 [0.22, 0.42]	0.30 [0.22, 0.42]	-0.08 to +0.06
2015	1332	50	0.04	0.27 [0.18, 0.42]	0.06 [0.03, 0.11]	0.28 [0.19, 0.40]	0.28 [0.19, 0.40]	-0.05 to +0.06

The same test on the skin/wound pilot and on all body systems (more positives per cutoff):

scope	cutoff	test pairs	gained support	AUPRC popularity	AUPRC history + popularity	Δ history [95% CI]
skin/wound pilot	2000	392	46	0.17	0.16	-0.07 to +0.04
skin/wound pilot	2005	385	39	0.38	0.34	-0.14 to +0.07
skin/wound pilot	2010	373	27	0.56	0.55	-0.06 to +0.06
skin/wound pilot	2015	366	20	0.50	0.46	-0.13 to +0.05
all body systems	2000	1404	122	0.38	0.36	-0.08 to +0.03
all body systems	2005	1384	102	0.34	0.27	-0.14 to -0.02
all body systems	2010	1361	79	0.31	0.29	-0.08 to +0.06
all body systems	2015	1332	50	0.27	0.28	-0.05 to +0.06

Finding: a null. Historical features alone score close to random, and adding them to research popularity never moves AUPRC beyond its confidence interval. The history coefficients are mostly negative, which points to selection: a pair is eligible only while unstudied, so well-attested remedies that science studied early drop out of the pool. Labels are a PubMed co-mention proxy awaiting a human audit.

4. What the LLM layer added

Two LLMs were used, and neither is ever the final authority. A Claude session (claude-opus-5-5) read all Greek passages twice (once for therapeutic claims, once for the plant's described form), turned most modern Wikipedia descriptions into the same trait vocabulary, and played the LLM-judge baseline without access to the expert answers. Groq's openai/gpt-oss-120b (temperature 0, cached API calls) extracted traits for the first 121 species until its daily token budget ran out. Every LLM output carries a verbatim span that the pipeline re-checks against the source; claims reach MeSH only through the verified lexicon; and nothing LLM-derived enters the temporal backtest, which keeps the deterministic claims and the frozen non-LLM identification.

LLM step	result
Passages read by the LLM (Claude session)	1,158
Claims the LLM extracted	5,075
…with verbatim Greek spans that passed the substring check	100%
Rule-extracted uses the LLM also found	94.9%
'Ambiguous' rule claims the LLM confirmed as uses	468 of 614
LLM claims outside the lexicon (kept apart, proposals for the lexicon)	1,616
Usable (ok) claims before → after	2561 → 3,550
Dioscorides chapters with an LLM trait record (spans verified)	324
Candidate species with a Wikipedia description / with traits	382 / 350 of 485
Candidate rows with ≥1 comparable trait (mean traits compared)	495 of 839 (2.51)

The morphology feature (f1) compares what Dioscorides says a plant looks like with what a modern description says, trait by trait (habit, height, leaf form, flower colour, fruit, root, smell, taste, habitat …). Its effect on identification is in the first table above (rows marked f1).

5. Leads for the lab

Plant–condition pairs used historically, with no matching PubMed pharmacology paper yet, ranked by forecast score × P(species). The top 20:

[!WARNING] Hypotheses to test, never advice to use. Rows marked ⚠ contain a compound with a PubChem GHS hazard record.

#1 Vicia faba Eye Diseases · P = 0.66 Photo: Wikimedia Commons contributor, Public domain	#3 Foeniculum vulgare Kidney Diseases · P = 0.58 Photo: Ryan Hodnett, CC BY-SA 4.0	#4 Artemisia absinthium ⚠ Eye Diseases · P = 0.75 Photo: David Monniaux, CC BY-SA 3.0	#5 Cupressus sempervirens Wounds and Injuries · P = 0.66 Photo: Taoujouti Abdelmouhaïmen, CC BY-SA 4.0
#7 Ruta graveolens ⚠ Eye Diseases · P = 0.52 Photo: Franz Eugen Köhler, Köhler's Medizinal-Pflanzen, Public domain	#8 Punica granatum Hernia · P = 0.73 Photo: H. Zell, CC BY-SA 3.0	#9 Myrtus communis Eye Diseases · P = 0.46 Photo: Wikimedia Commons contributor, Public domain	#10 Isatis tinctoria Wounds and Injuries · P = 0.65 Photo: Wikimedia Commons contributor, CC BY-SA 3.0
#11 Hedera helix ⚠ Pain · P = 0.65 Photo: AnRo0002, CC0	#12 Raphanus raphanistrum Wounds and Injuries · P = 0.59 Photo: Didier Descouens, CC BY-SA 4.0	#13 Quercus ilex Inflammation · P = 0.67 Photo: Fritz Geller-Grimm, CC BY-SA 3.0	#14 Vitex agnus-castus Inflammation · P = 0.64 Photo: peganum from Henfield, England, CC BY-SA 2.0

_{One photo per species, best-ranked lead first: Wikimedia Commons image (Wikidata P18; each file under its own licence, click for the file page). ⚠ = known toxic constituent.}

#	species	condition (MeSH)	P(species)	lead score	toxicity	IUCN	used in modern Greece	rank if certainty ignored
1	Vicia faba	Eye Diseases	0.66	0.387				1
2	Vicia faba	Hemorrhage	0.66	0.361				2
3	Foeniculum vulgare	Kidney Diseases	0.58	0.290			yes	4
4	Artemisia absinthium	Eye Diseases	0.75	0.276	⚠ known toxic constituent (PubChem)			11
5	Cupressus sempervirens	Wounds and Injuries	0.66	0.244		least concern	yes	12
6	Artemisia absinthium	Pain	0.75	0.242	⚠ known toxic constituent (PubChem)			18
7	Ruta graveolens	Eye Diseases	0.52	0.228	⚠ known toxic constituent (PubChem)			7
8	Punica granatum	Hernia	0.73	0.226		least concern		21
9	Myrtus communis	Eye Diseases	0.46	0.224		least concern		5
10	Isatis tinctoria	Wounds and Injuries	0.65	0.218				16
11	Hedera helix	Pain	0.65	0.217	⚠ known toxic constituent (PubChem)			17
12	Raphanus raphanistrum	Wounds and Injuries	0.59	0.211				13
13	Quercus ilex	Inflammation	0.67	0.209		least concern	yes	20
14	Vitex agnus-castus	Inflammation	0.64	0.203		Data Deficient		19
15	Senecio vulgaris	Wounds and Injuries	0.69	0.192	⚠ known toxic constituent (PubChem)			30
16	Vicia faba	Diarrhea	0.66	0.188				28
17	Lupinus albus	Stomach Diseases	0.54	0.183	⚠ known toxic constituent (PubChem)	least concern		15
18	Ruta graveolens	Hemorrhage	0.52	0.179	⚠ known toxic constituent (PubChem)			14
19	Lolium perenne	Liver Diseases	0.31	0.178				1127
20	Cupressus sempervirens	Inflammation	0.66	0.173		least concern	yes	34

• 2 of the top 20 rest on a plant identification below 0.5.
• 7 of the top 20 carry a known-toxicity flag.
• Ranking by the single most likely plant instead of the full probability changes 5 of the top 20 (Kendall τ = 0.50).

Leads by condition:

condition	leads	best lead	its P(species)	flagged toxic
Menstruation Disturbances	93	Daucus carota (#81)	0.54	18
Skin Ulcer	80	Daucus carota (#42)	0.54	12
Bites and Stings	66	Vicia faba (#50)	0.66	7
Edema	60	Daucus carota (#47)	0.54	6
Uterine Diseases	55	Hedera helix (#33)	0.65	12
Dysuria	55	Vitis vinifera (#614)	0.43	3
Inflammation	42	Quercus ilex (#13)	0.67	9
Jaundice	42	Cicer arietinum (#167)	0.69	3
Eye Diseases	41	Vicia faba (#1)	0.66	14
Colic	41	Daucus carota (#312)	0.45	1
Cough	40	Vicia faba (#64)	0.66	6
Splenic Diseases	40	Lupinus albus (#186)	0.54	10

Use it

From an AI assistant (MCP)

The MCP server runs on the same Cloudflare Worker as the site. It is read-only, needs no key, and every reply carries the research-only notice and toxicity flags.

Claude Code

claude mcp add --transport http pharmakon https://pharmakon.adoranto737.workers.dev/mcp

Claude Desktop / claude.ai: Settings → Connectors → Add custom connector → https://pharmakon.adoranto737.workers.dev/mcp

Cursor, Windsurf, VS Code and other clients (mcp.json):

{
  "mcpServers": {
    "pharmakon": { "url": "https://pharmakon.adoranto737.workers.dev/mcp" }
  }
}

Clients that only speak stdio can bridge with npx mcp-remote https://pharmakon.adoranto737.workers.dev/mcp.

tool	what it returns
search_ancient_names	Search the ~650 plant and drug names in Dioscorides' De materia medica (1st c.) by Greek (accents optional), transliteration, or modern Latin species. Returns ids for get_ancient_name.
get_ancient_name	Full record for one ancient name: Greek passages with CTS URNs and verbatim evidence spans, extracted uses, P(species
list_leads	Ranked research leads: plant–condition pairs used historically, with no matching PubMed pharmacology paper yet, weighted by identification certainty. Always check known_toxicity.
get_lead	One lead by rank with its compound → human target → disease evidence paths (LOTUS, ChEMBL, Open Targets) and the exact PubMed query behind its 'unstudied' label.
list_conditions	Conditions that have leads, with counts. Use the names as list_leads condition filters.
get_results	Evaluation results with confidence intervals: overview, identification (vs expert gold), persistence (vs permutation null), forecast (rolling-origin backtest), leads, labels, extraction, texts, chemistry, galen, graph.

Things to ask:

• "Which ancient Greek plants were used for burns, and how sure are we of each identification?"
• "Show the evidence for lead #2 and the PubMed query behind it."
• "Did Galen agree with Dioscorides about wormwood?"
• "How well did the backtest do in 2010?"

From code (JSON API)

Plain HTTPS GET, no key, CORS open. Same data and disclaimer as the MCP tools.

endpoint	returns
GET /api/names?q=&filter=&limit=	ancient names matching Greek, transliteration or Latin species
GET /api/names/{id}	one name: passages, uses, identification, survival, Galen, modern uses
GET /api/leads?condition=&species=&min_p=&hide_toxic=&limit=&offset=	ranked leads
GET /api/leads/{rank}	one lead with compound → target → disease paths
GET /api/conditions	conditions with lead counts
GET /api/results?section=	evaluation results (overview, identification, persistence, forecast, …)

curl "https://pharmakon.adoranto737.workers.dev/api/leads?condition=burn&hide_toxic=true&limit=3"

import requests
r = requests.get("https://pharmakon.adoranto737.workers.dev/api/names", params={"q": "Artemisia"}).json()
for n in r["results"]:
    print(n["lemma"], n["top_species"], n["p_top"])

On the website

The problem in one word. One Greek name, several candidate plants, each with a probability.	How it works. Read, extract, identify, trace, test, rank.
What the evidence says. Both tests, reported the same way whether they worked or not.	Open data. Every source checked live, licences recorded, limits stated.

Run it yourself

git clone https://github.com/LSUDOKO/Pharmakon && cd Pharmakon
make setup      # uv venv + install (Python 3.12)
make test       # 27 tests: parser, extraction, model, metrics, provenance guard, app pages
make app        # Streamlit research dashboard on the committed tables (no network)
make data verify   # download raw sources, check every API live
make reproduce     # the full pipeline from scratch
make readme        # regenerate this README's charts and tables

Or pull the container (CLI and committed tables, no network needed for the dashboard):

docker run --rm ghcr.io/lsudoko/pharmakon:latest --help
docker run --rm -p 8501:8501 --entrypoint streamlit ghcr.io/lsudoko/pharmakon:latest run src/pharmakon/app/Home.py --server.address=0.0.0.0

From a fresh clone, PubMed takes about an hour without an NCBI key. Optional keys go in .env: ANTHROPIC_API_KEY (LLM extraction), NCBI_API_KEY, IUCN_TOKEN.

Rebuild and deploy the site, API and MCP server (from web/): npm install && npm run data && npm run deploy.

Method

flowchart LR
  T[Dioscorides + Galen TEI] --> S[passages + CTS URNs]
  S --> X[claims with verified Greek spans]
  X --> M[MeSH / ICPC-2 / use groups]
  S --> N[name clusters + John the Physician crosswalk]
  N --> C[candidates: Lardos, LSJ, GBIF-verified]
  C --> I["P(species | name): conditional logit + unidentified"]
  I --> L[LOTUS, ChEMBL, Open Targets paths]
  I --> P[persistence: Dioscorides, Galen, JC, modern]
  M --> E[PubMed labels per year]
  P --> F[rolling-origin backtest]
  L --> F
  E --> F
  F --> R["leads × P(species)"]
  R --> W[website + API + MCP]

stage	command	output
Read texts	python -m pharmakon ingest-texts	1,500 passages, 835 chapters, headwords parsed 97.6%
Extract uses	python -m pharmakon extract-claims	3,696 claims, 0 span mismatches, extractor rule_extract_v1
Map conditions	python -m pharmakon map-conditions	158 Greek conditions, 134 with verified MeSH, 24 with no modern equivalent
Expert gold	python -m pharmakon gold	61 Dioscorides names, 50 John the Physician lemmas
Identify plants	python -m pharmakon identify	839 candidates, 537 of 683 names verified in GBIF
Read Galen	python -m pharmakon ingest-galen	476 drug entries, 1033 claims, copy containment 0.002
Link chemistry	python -m pharmakon link-chemistry	10,252 compounds, 455 human targets, 4,880 closed paths
Trace persistence	python -m pharmakon persistence	1,533 name × use rows, 1,226 modern records
Label evidence	python -m pharmakon evidence-labels	1999 plant–condition pairs, 218 with PubMed support by 2025
Backtest	python -m pharmakon forecast	cutoffs 2000, 2005, 2010, 2015
Rank leads	python -m pharmakon leads	1781 untested pairs, Kendall τ 0.50 vs top-1 identity
Knowledge graph	python -m pharmakon build-kg	2,585 nodes, 9,146 edges

Rules the pipeline enforces:

• Verbatim evidence. Every extracted claim stores the exact Greek substring and its CTS URN; a Unicode-normalised substring check archives any mismatch instead of keeping it.
• Provenance everywhere. Every row carries source_id, retrieved_at and method; a test fails the build if a table lacks them.
• Generate, then verify. Species names are verified against GBIF, diseases against MeSH, compounds by InChIKey. Unverifiable candidates are dropped and counted.
• No leakage. Backtest features use only information dated on or before the cutoff; LLM outputs are never used as a score.

Data and licences

data	source	licence
Dioscorides, De materia medica (Wellmann), TEI	First1KGreek / Open Greek and Latin	CC BY-SA 4.0
Galen, De simpl. med. temp. ac fac. books 6–8 (Kühn), TEI	First1KGreek	CC BY-SA 4.0
Liddell–Scott–Jones Greek lexicon	Perseus Digital Library	CC BY-SA 4.0
Expert identifications, John the Physician crosswalk	Lardos et al. 2024, figshare	CC BY 4.0
Mount Pelion ethnobotany	Tsioutsiou et al. 2023	CC BY
North Aegean ethnobotany	Axiotis et al. 2018	CC BY
Taxonomy and Greek/Cypriot/Turkish occurrences	GBIF backbone + occurrence API	CC0 / CC BY
Vernacular names, conservation status, plant photos	Wikidata / Wikimedia Commons	CC0 / per file
Plant → compound occurrences	LOTUS via Wikidata	CC0
Compound → target activity	ChEMBL 37	CC BY-SA 3.0
Target → disease association	Open Targets Platform	CC0
Disease vocabulary	NLM MeSH	public domain
Evidence labels (co-mention)	PubMed E-utilities	NLM terms
Species descriptions for the morphology feature	English Wikipedia (MediaWiki API)	CC BY-SA 4.0
Toxicity citations (GHS)	PubChem	public domain

Every source was checked live before use; the checks are recorded in docs/source_verification.md. Modern translations under copyright (Beck, Osbaldeston) are not used.

Research roadmap

What would make the results stronger, in order of value per effort:

next step	what it fixes	needs
Fill the four blinded audit sheets (now also covering LLM claims and the LLM lexicon extension)	turns proxy precision into audited numbers	expert reviewer time
Richer modern descriptions (Flora of Greece, POWO with permission)	f1 compares only ~3 traits per candidate today	data access
Historical Open Targets snapshots	removes the last time-leak in the chemistry features	archived releases
Paul of Aegina (7th c.), Oribasius, Aetius	completes the Byzantine chain	no open TEI exists yet; OCR of the CMG editions
Arabic and Latin transmission (Ibn al-Bayṭār, Latin Dioscorides)	follows the remedies out of the Greek world	new corpora and lexica
Zenodo release with DOI	citable, versioned dataset	release step
Screening partnership for top non-toxic leads	the only real test of a lead	a pharmacognosy lab

Limitations

• Expert gold, not ground truth. The gold standard is the judgement of Lardos et al.; intervals are wide with 61 names.
• Proxy labels. "Supported" means a PubMed co-mention with a pharmacology filter, not a validated finding. Its precision is not yet audited.
• LLM outputs. Claims and Greek trait records come from a Claude session and are committed rather than regenerated; species traits come from Groq. All spans are verified, but no human has audited them yet, and the LLM-proposed lexicon stems await the condition-mapping audit.
• Scope. Results cover all body systems; the original skin/wound pilot is kept for comparison. The Mount Pelion study covers skin ailments only, which favours skin uses in the persistence test.
• Text chain. Paul of Aegina is missing; John the Physician is reached only through the 50 plants Lardos et al. analysed; 137 of 476 Galen entries are unmapped.
• Chemistry. LOTUS via Wikidata, capped at 15 compounds per taxon and 120 taxa; human single-protein targets only; Open Targets associations are current, so a sensitivity model without them is reported.
• Conservation status comes from Wikidata P141, not the IUCN API.

Responsible research

• No dosing is displayed; quantities appear only inside quoted historical text.
• Toxic genera carry a PubChem GHS citation resolved live; every API and MCP reply repeats the disclaimer.
• Modern ethnobotanical data are used only at the aggregate species and use level from published tables.
• Folk knowledge belongs to the documented communities: access-and-benefit-sharing under the Nagoya Protocol applies to any commercial follow-up.

Cite and acknowledge

Pharmakon (2026). An uncertainty-aware, 2,000-year map of Greek and Byzantine pharmacology. https://github.com/LSUDOKO/Pharmakon. Citation metadata: CITATION.cff (GitHub's "Cite this repository" button).

Built by Arpit (B.Tech Computer Science, NIT Hamirpur).

Built for EthnoHack 2026, Track 1 (Traditional knowledge). Thanks to Open Greek and Latin (First1KGreek), the Perseus Digital Library, Lardos et al. (2024), Tsioutsiou et al. (2023), Axiotis et al. (2018), GBIF, Wikidata and LOTUS, NLM, EMBL-EBI ChEMBL, Open Targets and PubChem. Manuscript images: Vienna Dioscorides (Codex Vindobonensis med. gr. 1), public domain.

Code: MIT. Data: under each source's licence, listed above.

Built With

• apache-parquet
• chembl
• claude
• cloudflare-workers
• docker
• duckdb
• gbif
• github-actions
• groq
• javascript
• model-context-protocol-(mcp)
• networkx
• pandas
• plotly
• pubmed
• python
• react
• remotion
• sarvam-ai
• scikit-learn
• scipy
• streamlit
• typescript
• vite
• wikidata