Atlas Pipeline Benchmark Report¶

4,466 entities · 6,851 relationships · 22 min · ~$1.92

Full knowledge graph from a complete Atlas drug discovery program. Zero Atlas code modifications — graffold-ingest reads the filesystem.

Top 100 hubsFull extraction (363 nodes)363 nodes (Graphistry)

100 most-connected entities from 4,466 total · 345 edges · shows hub structure

All entities from a single pipeline run · 363 nodes, 221 edges · color-coded by type

363 entities · 221 relationships · GPU-accelerated layout · color by type

Source Documents (Atlas Phase Outputs)¶

These are the intermediate markdown files produced by each Atlas agent — the raw input that graffold-ingest parses into the knowledge graph above. The final two tabs show the output parquet files (the graph itself).

1: Disease Map1a: Anomaly1b: Tribunal2: Failure Analysis2b: Landscape3: Candidates3: Vulcan3b: Survey3c: Lit Sweep3d: Feasibility4: Kill Report4b: Board5: Coverageentities.parquetrelationships.parquet

phase-1-disease-map.md — 373 lines, 61,128 chars

# Phase 1 — Disease Map: Bovine Neonatal Cryptosporidiosis

**Program:** cargill-cryptosporidiosis-v11/v1
**Agent:** Pathfinder
**Date:** 2026-05-30
**Species:** Bovine (*Bos taurus*), neonatal calves 1–4 weeks of age
**Pathogen:** *Cryptosporidium parvum* (apicomplexan; intracellular-but-extracytoplasmic "epicellular" parasite)
**Prior versions:** v8 (2026-05-11), v9 (2026-05-28), v10 (2026-05-29). This map **extends and independently re-verifies** v10 (1 day prior). It is written as **disease biology + rate-limiting barrier + KE#1**; the three-tier portfolio and commercial framing are Anvil's job, not Pathfinder's.

---

## 0. Run context and what is new in v11

**Brief intent (authoritative).** Cargill will NOT license any existing asset — no existing small molecules, antibodies, vaccines, repurposed compounds, or feed additives (this explicitly retires the prior v6 P23-IgY recommendation). They want **novel drug targets Agteria can own end-to-end** (clean IP; novel chemistry OR novel biologic), organised downstream into three risk/reward tiers. Disease scope: *C. parvum* in calves ≤4 weeks, herd-level feedlot/dairy economics, slaughter-appropriate withdrawal (<30 d). A killed *molecule* does not invalidate the *target* behind it — novel Agteria chemistry on a validated target is the stated strategy (see §0.3).

### 0.1 The single most important biological fact (unchanged, [ESTABLISHED])
*C. parvum* occupies an **intracellular-but-extracytoplasmic ("epicellular") niche**: it invaginates under the enterocyte apical plasma membrane, forming a parasitophorous vacuole (PV) that is *inside* the host membrane but *outside* the host cytoplasm, connected to the host cell by a parasite-derived **feeder organelle**, a ring-like tight junction, an electron-dense band, and an actin pedestal. This geometry — re-confirmed in detail by the 2025 Striepen/Vinayak-lab life-cycle review (PMID 41539020, [DOI](https://doi.org/10.1016/j.mib.2025.102703)) — is unique among apicomplexans and makes pharmacology intuition imported from *Plasmodium*/*Toxoplasma* (both cytoplasmic) routinely wrong. It governs both pathology (the parasite destroys the absorptive brush border directly) and druggability (a drug must reach the gut lumen/apical surface; intracellular-acting chemistry largely misses it; **host-directed targets bypass the barrier entirely**).

### 0.2 v11 primary value-add #1 — citation-integrity health check: **CLEAN (third consecutive run)**
The substrate flags 127 unverified *C. parvum* PMIDs, and Pattern 11 warns that panel models (esp. o3-pro) fabricate *C. parvum* citations. The v10 map's primary risk was that its **ten "new-in-v10" PMIDs were verified only by dispatched subagents and never re-verified by Pathfinder** — including six high-numbered 2026 papers, exactly the fabrication-risk profile. **This run independently re-verified 23 load-bearing PMIDs against NCBI eutils/PubMed (title + journal + year + first author), prioritising those ten.** Result: **23/23 VERIFIED real and correctly attributed — including all six 2026 high-PMID papers (42118002, 41991009, 42003133, 42081373, 42162589, 41885307).** No fabrication, no mismatch. Full ledger in §13. Minor attribution nuances (not errors): the KRS paper (40818563) is primarily a *rodent-model* paper that uses a KRS inhibitor as a pharmacological tool (§4.8 reframed); first-author vs senior-author labels are consistent throughout.

### 0.3 v11 primary value-add #2 — the CpASP2 / KILL-083 decision-gate, substantially resolved (§11)
The brief explicitly asks whether KILL-083 (CpASP2 egress inhibitor) should be revived for Tier 3, hinging on whether the neonatal IFN-γ gap prevents clearance of *trapped* (parasitostatic) parasites. v10 framed this as a "$0 read of the Wallbank Methods to learn the mouse strain." **v11 finding:** the Wallbank 2025 abstract does **not** name the strain, and the full text is paywalled/not web-accessible (so the "$0 web check" actually needs institutional access). **But the question is resolvable on model-constraint grounds anyway** — and the answer **weakens KILL-083**. See §11.2.

### 0.4 Cross-program kill registry (respected)
Molecule-level SOFT_KILLs whose **target biology remains valid** for novel Agteria chemistry: KILL-090 Auranofin→**CpTrxR target valid** (§4.4); KILL-087 Lapaquistat→**host FDFT1/squalene axis valid** (§4.5, the v10 portfolio backbone); KILL-089 BEZ235→**PI3K-cMYB-BCL2A1 anti-apoptosis axis is real biology** (§6.6, 2026 in-vivo data — flagged for Reaper/Anvil); KILL-088 LY2090314/CpGSK3, KILL-084 Meloxicam, KILL-085 Vorinostat/CpHDAC, **KILL-086 Roc-A/CpeIF4A** (§4.12 — **new no-rebound-in-NCG-mouse evidence flagged for Reaper/Anvil**). Target/mechanism-level HARD_KILLs that must NOT be re-proposed without meeting resurrection conditions: KILL-081 Myb-M (no TF-activation precedent), KILL-082 NLRP6-IL18 (rBoIL-12 failure class), **KILL-083 CpASP2** (the brief authorises a fresh Tier-3 look — §11).

### 0.5 What is genuinely new in v11 (all PubMed-verified this run)
1. **CpeIF4A "fast-killing" with NO rebound in NCG (T/B/NK-null) mice** (§4.12) — decision-relevant for the parasiticidal thesis; relevant to KILL-086's resurrection condition.
2. **CpASP2 mouse-strain question resolved on model-constraint grounds** → KILL-083 rationale weakened (§11.2).
3. **INS6 microneme protease** — CRISPR-validated, reduces burden/improves survival in IFN-γ-KO mice (§4.13); companion **INS1** (macrogamont, M16 protease).
4. **Rhomboid protease family CpRom1/2/3** — invasion/egress + feeder-organelle proteostasis (§2.7).
5. **CpPyK pyruvate-kinase parasite-unique disulfide dimerization** — selective-target angle for the glycolysis thesis (§4.3).
6. **CpAOX genetic KO = non-essential** (target KILL) and **NDH2 = clofazimine bioactivator + field-resistance node** (caution flag) (§4.3).
...

phase-1a-anomaly-map.md — 170 lines, 28,889 chars

# Phase 1a: Anomaly Map — cargill-cryptosporidiosis-v11/v1

**Agent:** Anomaly | **Date:** 2026-05-30

> **What this document is.** Pathfinder mapped the disease as a taxonomy of pathogen lifecycle stages. This is a different lens: the **causal chain that produces the partner's problem** (dead/stunted calves + an endemic herd reservoir + economic loss), traced node-by-node from a healthy calf to problem manifestation, with an explicit Liebig's-barrel bottleneck analysis and — the unique value of this pass — a map of the nodes where **nothing currently addresses the vulnerability at all**.
>
> **The one-line thesis.** Bovine neonatal cryptosporidiosis is intrinsically self-limiting (immunocompetent calves self-cure in 1–2 weeks). It only becomes a *problem* because, during days ~5–14 of life, **parasite amplification wins a rate race against host clearance** — clearance is offline (the neonatal immune gap) and no drug supplies a clearance term. Every downstream stave of the barrel — epithelial destruction, dehydration, growth penalty, herd shedding — is a consequence of that one lost race. The bottleneck is therefore **not a biochemical node but a rate imbalance at Node 4**, and the single highest-value unknown gating it is whether host enterocyte turnover *alone* can finish the job once amplification is merely arrested.
>
> *Citations carried from Pathfinder's §13 verification ledger (23/23 PubMed-verified this run); no PMID is introduced here that Pathfinder did not verify. No compound, product, or modality is named — see the attestation canary at the foot of this file.*

---

## 1. The Causal Map

The chain runs: **healthy calf → exposure → establishment → amplification → (rate race: amplification vs clearance) → absorptive-surface destruction → pathophysiology → clinical/economic outcome + herd loop-back.** Seven nodes. Each is a *biological state*; sub-items are the specific failure modes that hold that state open.

### Node 1 — Exposure pressure (healthy calf, contaminated environment)
The calf is born into an environment whose oocyst load is, for practical purposes, unbounded and continuously replenished.
- **1a.** Maternal **periparturient shedding rise** seeds the calving pen; the calf is often infected within hours of birth. [MODERATE] ([Striepen/Vinayak-lab review (2025)](https://pubmed.ncbi.nlm.nih.gov/41539020/))
- **1b.** Thick-walled oocysts are immediately infective on shedding (no environmental maturation), survive months at 4–20 °C, and resist chlorine. [ESTABLISHED] ([Striepen/Vinayak-lab review (2025)](https://pubmed.ncbi.nlm.nih.gov/41539020/))
- **1c.** ID₅₀ is extremely low (experimental disease from ~17–132 oocysts) against shedding of 10⁸–10¹⁰ oocysts/calf/day — exposure is effectively unavoidable. [ESTABLISHED] ([Striepen/Vinayak-lab review (2025)](https://pubmed.ncbi.nlm.nih.gov/41539020/))
- **1d.** Sequential cohorts make the operation functionally endemic — the reservoir is self-sustaining independent of any one calf. [MODERATE]

### Node 2 — Establishment (excystation → invasion → epicellular niche forms)
A single sporozoite reaches the apical enterocyte surface and builds the shielded niche.
- **2a.** Bile salts + pancreatic proteases trigger excystation; attachment occurs within seconds, full host-membrane encapsulation by ~8 min. [ESTABLISHED] ([Striepen/Vinayak-lab review (2025)](https://pubmed.ncbi.nlm.nih.gov/41539020/))
- **2b.** The **adhesin system is highly redundant** (GP60→GP40/GP15, GP900, TRAP/TSP family, mucins, Gal/GalNAc lectin) — blocking any single adhesin leaves alternatives open. [MODERATE] ([Striepen/Vinayak-lab review (2025)](https://pubmed.ncbi.nlm.nih.gov/41539020/))
- **2c.** The parasite forms the **intracellular-but-extracytoplasmic ("epicellular") niche** under the apical membrane — inside the host membrane, outside the cytoplasm, fed by a parasite-derived feeder organelle. This single geometric fact governs both pathology and druggability: the parasite is shielded from antibody and from intracellular-acting agents, and reachable only from the lumen or from the host side. [ESTABLISHED] ([Striepen/Vinayak-lab review (2025)](https://pubmed.ncbi.nlm.nih.gov/41539020/))
- **2d.** Gliding motility (Cp23, genetically essential *in vivo*) is required to *reinfect neighbouring cells* — i.e. establishment of the *first* cell is not the leverage point; spread to the *next* cell is (this routes the leverage to Node 3, not Node 2). [ESTABLISHED] ([Watson et al. (2025)](https://pubmed.ncbi.nlm.nih.gov/40835841/))

### Node 3 — Within-host amplification (merogony + the autoinfection loop)
This is the **rate-of-increase node** — where one infected cell becomes a pathological burden.
- **3a.** Each sporozoite → 8-nucleated meront in ~36 h; merozoites egress, glide, reinfect → a single oocyst's 4 sporozoites yield thousands of merozoites in 72 h. The burden builds *exponentially*. [ESTABLISHED] ([Striepen/Vinayak-lab review (2025)](https://pubmed.ncbi.nlm.nih.gov/41539020/))
...

phase-1b-bottleneck-consensus.md — 260 lines, 37,702 chars

# Phase 1b — Tribunal: Bottleneck Consensus

**Program:** cargill-cryptosporidiosis-v11/v1
**Agent:** Tribunal (4 independent frames + Evaluator)
**Date:** 2026-05-30
**Inputs:** `phase-1-disease-map.md` (Pathfinder) + `external-input-phase-1.md` (validated 7-model panel). Anomaly's `phase-1a-anomaly-map.md` was **withheld** from the four frames and used by the Evaluator only, as an independent cross-check.

> **One-line determination.** Bovine neonatal cryptosporidiosis is intrinsically self-limiting; it only becomes the partner's problem because, during a fixed days-5–14 window, **parasite amplification wins a rate race against host clearance — and no drug supplies the missing clearance term.** The bottleneck is therefore the **clearance side of that rate race ("Node 4")**, and the single decision-gating unknown is **whether host enterocyte turnover (τ ≈ 3–5 d) can drain an amplification-arrested burden in the *calf* before villous destruction accumulates.** That one measurement reprioritises half the portfolio.

---

## 0. How the panel ran

Four agents analysed the disease from four frames, in parallel, blind to each other and to the Anomaly map:

- **A — Unframed Analyst** (no lens; follow the biology)
- **B — Pathogen Specialist** (the parasite's essential, non-redundant dependencies)
- **C — Host/Environment Analyst** (what host/herd variable decides outcome)
- **D — The Martian** (quantities only; no domain priors)

The Evaluator then mapped convergence, framed and resolved the central disagreement, and compared the result against Anomaly's independent conclusion.

---

## 1. Convergence Map

| Claim | A | B | C | D | Count |
|---|---|---|---|---|---|
| Disease is a **rate race** (∫ amp − clear) over a fixed window | ● | ● | ● | ● | **4/4** |
| Neonatal immune gap is a **FIXED, non-repairable boundary** — not a lever | ● | ● | ● | ● | **4/4** |
| **Calf turnover rate τ** (clears arrested parasites?) is the pivotal unknown | ● | ● | ● | ● | **4/4** |
| Host-directed **squalene/GSH–FDFT1** starvation is a top-tier lever | ● | ● | ● | ● | **4/4** |
| Delivery/PK barrier is real but **bypassed by host-directed targets** | ● | ● | ● | ● | **4/4** |
| Thin:thick **autoinfection fate switch** = highest leverage, un-targetable | ● | ● | ● | ● | **4/4** |
| **CONTINUE** the program (do not close) | ● | ● | ● | ● | **4/4** |
...

phase-2-failure-analysis.md — 405 lines, 70,056 chars

# Phase 2 — Failure Analysis: Bovine Neonatal Cryptosporidiosis

**Program:** cargill-cryptosporidiosis-v11/v1
**Agent:** Sapper (treatment archaeologist)
**Date:** 2026-05-30
**Reads from:** `phase-1-disease-map.md` (Pathfinder), `phase-1a-anomaly-map.md` (Anomaly), `phase-1b-bottleneck-consensus.md` (Tribunal), `external-input-phase-1.md` (validated 7-model panel), `brief.md`, and the v10 failure analysis (`../../cargill-cryptosporidiosis-v10/v1/phase-2-failure-analysis.md`).
**Method:** This run **extends and independently re-verifies** the v10 autopsy (mirroring Pathfinder's disease-map approach). Every load-bearing PMID below was checked against PubMed this run (title + first author + journal + year + abstract substance). The genuinely-new v11 entries were prioritised for verification because recent high-numbered *C. parvum* PMIDs are exactly the panel-fabrication risk profile (substrate flags 127 unverified PMIDs; Pattern 11).

---

## 0. What this document is — and the one sentence it proves

This is a forensic autopsy of **every treatment tried against *Cryptosporidium parvum* in calves (and the human/animal models that inform it), and exactly why each one failed.** It analyses **17 approaches across 7 classes**, tags each as a **TARGET FAILURE** (the biology defeats the approach) or a **COMPOUND FAILURE** (the biology is sound but the molecule/modality couldn't deliver), catalogues every **in-vitro → in-vivo translation gap**, ends with a gap map and the explicitly-named **rate-limiting barrier to cure**, and answers the failure-record half of the brief's **CpASP2 / KILL-083 Tier-3 question** (§11).

> **The single sentence the failure record proves:** *Not one intervention in 40 years is simultaneously **parasiticidal / immune-independent**, **able to reach the epicellular parasite (or act host-side to bypass the delivery barrier)**, **deliverable early enough to prevent the absorptive-surface lesion**, and **food-animal-safe with a clean, ownable IP path and a <30-day withdrawal**. Every agent fails at least one of those four filters; the large majority die on the first — they are parasitostatic and hand final clearance to a neonatal immune arm that is offline during the exact disease window (days 5–14).*

This is the same parasitostatic-vs-parasiticidal thesis the disease map and Tribunal converged on (the amplification-vs-clearance rate race at Node 4). **Sapper's job is to show it is written into the trial record of every drug, not merely inferred from biology — and to surface the two pieces of NEW v11 evidence that sharpen it.**

### 0.1 What is new in v11 (all PubMed-verified this run)

1. **Clofazimine is a DOUBLE failure, not a clean PK failure (§5.1).** The 2026 Buenconsejo/Striepen genetic-cross paper (PMID 42129495, *Nat Microbiol*) shows clofazimine is **bioactivated by the parasite's type-II NADH dehydrogenase (NDH2)**, and that a **conserved attenuated *ndh2* allele is carried widely across continents** in *C. parvum* AND *C. hominis* — so the field carries a **pre-existing pharmacogenomic resistance landscape**, on top of the disease-modified PK that sank the CRYPTOFAZ trial. Deng 2025 (NDH2 non-essential) explains *why* the attenuated allele is cost-free to carry. This converts the textbook PK-failure into "couldn't reach the parasite **and** wouldn't have been reliably bioactivated even if it had."
2. **The first real-world calf test of an arrest/attenuation mechanism fell short of cure (§4.4).** The CDPK5-KO genetically-attenuated strain in the **natural calf model** (Nava 2026, PMID 41885307) gave "significant improvement in overall clinical outcome" but **"diarrhea and oocyst shedding did not fully resolve."** Because this is a *genetic* attenuation, it removes the PK/delivery confound entirely — and still left residual disease in the actual bovine host. This is the single most decision-relevant new datapoint for the parasitostatic-vs-parasiticidal / KE#1 question (§11).
3. **A live, EMA-approved single-antigen vaccine now exists — and only partially works (§6.2).** Bovilis Cryptium (recombinant **GP40** maternal vaccine, EMA-authorised 29 Nov 2023) is licensed only to **reduce diarrhoea** (pivotal field study: duration 2.2 → 1.8 d, *p*=0.03), with **no prevention-of-infection or shedding-elimination claim** (~41% of vaccinates still infected). A modern deployed product is a *live* confirmation of the single-antigen + epicellular-niche + neonatal-timing target-biology limit — and it **resets the competitive baseline** for the partner conversation.
4. **A fast-killing, no-rebound counter-example was confirmed (§5.4).** Roc-A/CpeIF4A (PMID 40720562) cleared an arrested parasite with **no rebound in NCG (T/B/NK-null) mice** — corroborating that, *in mice*, clearance of a non-replicating parasite does not require adaptive/IFN-γ/NK immunity. The molecule (a rocaglate natural product) stays a lead, not a product — a clean COMPOUND failure on a valid cidal target.
5. **Three "validated target, molecule killed" host-/redox-directed cases added (§5.2–5.3):** lapaquistat (host FDFT1/squalene-GSH; Hy's-Law hepatotox in a human cholesterol program), BEZ235 (host PI3K→BCL2A1 anti-apoptosis; discontinued oncology dual inhibitor, dose-limiting tox), auranofin (parasite CpTrxR; gold heavy-metal liabilities). All three point to the same escape route: **novel Agteria chemistry on a validated target.**

**Regression-integrity result: the v10 backbone held — all spot-checked anchor PMIDs (halofuginone Jarvie/Trotz-Williams/Silverlås, NTZ Schnyder/Amadi, paromomycin Grinberg/Hewitt, KDU731 Manjunatha 2017/2024) re-verified with no number drift, no mis-attribution, no fabrication.** This is the third consecutive clean *C. parvum* citation run for the program.

---

## 1. CLASS A — Approved / licensed parasitostatics

### 1.1 Halofuginone lactate (Halocur) — the only EU-licensed calf anti-cryptosporidial

**What was tried.** Oral halofuginone lactate, **100–120 µg base/kg/day**, prophylactic for the first 7 days of life (label) or therapeutic from first diarrhoea.

**Result (with numbers, re-verified this run).**
- Jarvie 2005 (n=31, RCT, PMID 15829673, [DOI](https://doi.org/10.3168/jds.S0022-0302(05)72854-X)): odds of shedding **70% lower** vs placebo (positives 22.4% vs 42.5%); diarrhoea onset delayed ~3.1 d; **milk intake, body-weight gain and weaning age NOT significantly different.** The canonical parasite-kill-≠-weight-gain decoupling.
- Trotz-Williams 2011 (n=513, PMID 21546409, [DOI](https://doi.org/10.1136/vr.d1492)): shedding OR 0.6 (95% CI 0.4–0.9, *P*=0.009); **growth (P=0.05/0.10), mortality (P=0.07) and diarrhoea incidence all non-significant.**
...

phase-2b-competitive-landscape.md — 290 lines, 36,471 chars

# Phase 2b — Sentinel Run 1: Competitive Landscape (pre-Forge)

**Program:** cargill-cryptosporidiosis-v11/v1
**Agent:** Sentinel (Run 1 — competitive intelligence & landscape scout)
**Date:** 2026-05-30
**Disease:** bovine neonatal cryptosporidiosis (*Cryptosporidium parvum*), calves ≤4 weeks
**Brief constraint:** novel owned drug targets — novel chemistry or biologic, clean IP. EXCLUDED from the portfolio: existing small molecules, existing antibodies, existing vaccines, repurposed compounds, feed additives. (Competitor products in those classes still matter for landscape/market-sizing and are reported here in full.)

**Verification note.** Six parallel search streams were run (clinical trials, approved products/labels, patents/FTO, recent literature, stealth competitors, funding/regulatory). Every PMID cited as load-bearing was re-verified against NCBI eutils by Sentinel directly (titles confirmed, §Appendix A). NCT numbers were fetched from ClinicalTrials.gov. Product-label content is reported verbatim where the official source was retrieved; provenance caveats and a "could-not-verify" list are at §12. This is the **third consecutive clean *C. parvum* citation run** for the program (per Phase-2 failure analysis); no fabricated PMIDs or patent numbers surfaced in cross-check.

---

## 0. Headline for Forge (TL;DR)

1. **The dominant competitive threat is EDI048 (Novartis PI(4)K inhibitor), now in Phase 2.** It is a gut-restricted **parasiticidal** small molecule, **de-risked in the neonatal-calf model itself**, with clean human Phase 1 safety and a composition-of-matter patent to **~2042**. Phase 2 human controlled-infection readout (CRYPTONITE, NCT07249463) is due **~March 2027**. **Do not propose a PI(4)K target — that lane is owned and fortified.** But EDI048 is *good news for the thesis*: it proves a gut-restricted parasiticidal small molecule clears calf disease, and it makes "fecal oocyst shedding + diarrhoea resolution in the calf model" a regulator-credible efficacy endpoint.

2. **Occupied / crowded mechanisms to avoid as me-too plays:** PI(4)K (Novartis, +Merck filings); **CpCDPK1 / bumped-kinase inhibitors** (University of Washington IP, active to 2031/2034; ParaTheraTech + BKI-1708 now has direct calf efficacy); **lysyl-tRNA synthetase** (Dundee/Vermont, calf data, Sci Transl Med 2024); the MMV665917/SLU-2633 piperazine series (UVM + Saint Louis University IP).

3. **The cleanest novel-ownable target by IP and biology is CpTrxR (thioredoxin reductase):** no blocking patent found, now structurally enabled (first crystal structures 2025), parasite-selective C-terminal redox motif absent in humans, and *C. parvum* has **no glutathione-reductase backup** → parasiticidal in principle. Caveat: it is a **parasite-directed luminal** target and inherits the epicellular-niche delivery existential flag (Tribunal §4).

4. **Host-directed squalene/FDFT1 was strongly validated in 2025 (Cell), but the IP is contested:** the Francis Crick Institute has a **pending method-of-use patent** claiming FDFT1 inhibition for *Cryptosporidium*. The underlying chemistry (lapaquistat, zaragozic acids) is public/expired. Agteria can own novel squalene-synthase chemistry but must design around — or file ahead of — Crick's use claim. Flag to Anvil.

5. **The existing veterinary market is weak and exploitable.** Only Halocur (halofuginone — both prevention AND treatment, but parasitostatic, ~2× toxicity margin) and the 2024 Bovilis Cryptium maternal vaccine (reduce-diarrhoea-only; single significant endpoint a **0.4-day** diarrhoea reduction; protection demonstrated only to 14 days vs a disease peak at days 7–12). Paromomycin (Parofor Crypto) is licensed in UK/EU but only reduces shedding. **No approved agent is parasiticidal, immune-independent, food-animal-safe AND ownable** — the whitespace the brief targets is real.

6. **Regulatory tailwinds favour a calf-first novel agent:** FDA-CVM **expanded conditional approval** (cattle-eligible, unmet-need, actively granted in 2025) is the fastest US route; **withdrawal <30 days is not binding** (halofuginone's benchmark is 13 days; gut-restriction or a biologic shortens it further).

---

## 1. Clinical Pipeline

### Human trials — ACTIVE / forward-looking (the competitive set)

| NCT ID | Compound | Phase | Status | Sponsor |
|--------|----------|-------|--------|---------|
| NCT07249463 (CRYPTONITE) | EDI048 (PI4K inhib.) | Phase 2 | Recruiting | Novartis |
| NCT06600711 | Nitazoxanide vs placebo | Phase 3 | Not yet recruiting | Johns Hopkins / NIAID |
| NCT07388615 | Echinacea (Immulant) | Phase 2 | Recruiting | Al-Azhar Univ. |
| NCT05036668 | ABO809 (challenge strain) | Phase 1 | Completed 2022 | Novartis |

- **NCT07249463 (CRYPTONITE)** — EDI048 controlled-human-infection model (adults challenged with *C. parvum* ABO809, then randomised to EDI048 vs placebo). n≈96; start **2025-12-08**, primary completion **2027-03-02**; site Pharmaron Inc (US). **Threat level: HIGH (the only Phase 2+ novel-mechanism crypto drug).**
...

phase-3-candidates.md — 383 lines, 51,316 chars

# Phase 3 — Candidates: Bovine Neonatal Cryptosporidiosis

**Program:** cargill-cryptosporidiosis-v11/v1
**Agent:** Forge (literature-aware inventor stream)
**Date:** 2026-05-30
**Reads from:** `phase-1-disease-map.md` (Pathfinder), `phase-1a-anomaly-map.md` (Anomaly), `phase-1b-bottleneck-consensus.md` (Tribunal), `phase-2-failure-analysis.md` (Sapper), `phase-2b-competitive-landscape.md` (Sentinel), `external-input-phase-2.md` + validator report, `brief.md`, and the canonical `portfolio.json` / `kill-registry.json` (cross-program brain).

**What this document is.** A novel-ownable-target candidate set for a Cargill calf-crypto program, organised so Anvil can build three risk/reward tiers. It (1) acknowledges every prior cryptosporidiosis portfolio entry, (2) reframes the brief's strict "novel owned targets only" constraint into a validated-target-extraction exercise, (3) **decomposes the primary target** (the non-redundant redox economy + the clearance-supply problem) into all molecular intervention points, (4) proposes candidates across categories A–D covering **every disease stage**, (5) runs the activation-vs-inhibition and citation-discipline checks, and (6) appends five Forge-specific falsifiable predictions.

> **Forge's one-sentence thesis.** The upstream consensus is correct and I do not contest it: the bottleneck is the **Node-4 clearance race** and the single most coherent parasiticidal play is the **non-redundant redox economy attacked from both valves (host FDFT1 + parasite CpTrxR)** — so my job here is not to find a new bottleneck but to **convert every validated-target / killed-molecule case into a specific, ownable, novel-chemistry candidate with an honest druggability and IP read**, and to add the two creative modality pivots (host-turnover clearance via BCL2A1; autoinfection-loop break via the glideosome/Cp23 motor) and the one designed-in gap (the thin:thick fate regulator) that the literature-aware lens reaches.

---

## Prior Target Acknowledgment (MANDATORY — every cryptosporidiosis portfolio entry)

The shared brain holds ~40 cryptosporidiosis entries (portfolio T-### + kill-registry KILL-###). Disposition of each below. **No HARD_KILL is re-proposed without meeting its resurrection condition; the single authorised exception is CpASP2/KILL-083, which the brief explicitly opens for a Tier-3 look.**

### HARD_KILLs / mechanism-level kills — NOT re-proposed (except where brief authorises)

| Registry id | Target | Forge disposition |
|---|---|---|
| KILL-081 | Myb-M TF *activation* | **Not revisited.** No TF-activation precedent; resurrection bar unmet. |
| KILL-082 | NLRP6-IL18 / immunostim | **Not revisited.** rBoIL-12 failure class (immune-gap is FIXED). |
| KILL-083 | CpASP2 egress inhibitor | **Re-examined for Tier 3** (brief-authorised) — see C7; τ-contingent. |

### Molecule-level SOFT_KILLs — TARGET re-proposed as novel ownable chemistry

| Registry id | Killed molecule | Target re-proposed as |
|---|---|---|
| KILL-087 | Lapaquistat | host FDFT1/squalene-GSH → **C1** (gut-restricted) |
| KILL-090 | Auranofin (gold) | parasite CpTrxR → **C2** (non-gold binder) |
| KILL-089 | BEZ235 (PI3K/mTOR) | host BCL2A1 axis → **C3** (BH3-mimetic) |
| KILL-086 | Roc-A (rocaglate) | parasite CpeIF4A → **C4** (synthetic selective) |
| KILL-088 | LY2090314 / CpGSK3 | **Not re-proposed** — GSK3 host-conserved, repurposing-class; weaker than C1–C4. |
| KILL-084 | Meloxicam | **Out of scope** (existing drug; ORT adjunct only). |
...

phase-3-vulcan.md — 372 lines, 52,024 chars

# Phase 3 — Vulcan: First-Principles Vulnerability Analysis

**Program:** cargill-cryptosporidiosis-v11/v1
**Agent:** Vulcan (asymmetric partial quarantine)
**Date:** 2026-05-30
**Inputs consulted:** `phase-1-disease-map.md`, `constraints-ledger.md` ONLY. No Forge candidates, no Sapper failure analysis, no Tribunal bottleneck call, no partner context, no kill registry, no prior-program files. All mechanistic grounding below comes from primary structural/biochemical literature retrieved this run.

---

## 0. Method, framing, and the two lenses I reason through

This document does **not** ask "what has been tried?" It asks: *given the parasite's biology and the two factual constraints I'm allowed to see, where are the molecular points at which this organism breaks?* I decompose each parasite "machine" into every intervention point along its lifecycle, then run a mandatory host-dependency pass, then name the system-level fragilities.

Two factual lenses from the constraints ledger shape which intervention points are *pharmacologically* (not just biologically) viable:

**Lens 1 — the delivery barrier is the dominant filter, and the target patient changes it.** The disease map's central fact is the multi-membrane epicellular barrier (lumen → apical membrane → PVM → feeder organelle → parasite). Most chemistry dies before reaching the parasite. This makes **luminal** targets (the oocyst and the apical parasite surface, both bathed in gut contents) and **host-directed** targets (which bypass the barrier entirely) structurally advantaged over intracellular-acting parasite chemistry. I weight accordingly.

A factual physiology point the species constraint understates: **the target patient is a 1–4-week-old *pre-ruminant* calf.** The reticular/oesophageal-groove reflex shunts suckled milk/colostrum/milk-replacer directly to the abomasum, bypassing the undeveloped rumen. "Rumen degradation" is the dominant oral-delivery risk in the *adult* ruminant; in the *neonate fed in milk*, it is largely circumvented. This materially de-risks orally-delivered peptidomimetics and even some proteins **if dosed in the milk/colostrum vehicle** — and it is exactly the patient window the disease targets. I treat milk-vehicle oral delivery to the neonate as viable.

**Lens 2 — residue/withdrawal + dosing-frequency economics.** Food-animal residue limits and a <30-day withdrawal favour **gut-restricted / non-absorbed** drugs (no systemic residue) and disfavour systemically-distributed reactive chemotypes (e.g., redox-active heavy metals — a general tox/residue liability, not a target liability). "Repeated individual dosing is operationally weak" favours **single-dose or short-course oral-at-birth** regimens — which is operationally feasible precisely because each neonatal calf is already hand-fed colostrum. This strongly rewards a **prophylactic luminal block given once in the first feed.**

Net: the constraints push the centre of gravity toward (a) luminal / barrier-free targets, (b) host-directed targets, and (c) parasiticidal-or-loop-breaking mechanisms (the disease map's R₀ ≈ 5–15 means <90% reduction does not move herd outcomes). I flag every intervention point against these.

---

## 1. The core reframing — this pathogen has no toxin; it has an auxotrophy, a supply line, and a transmission gate

A bacterial-toxin decomposition (transcription → translation → modification → secretion → activation → receptor → effect) does not map onto *C. parvum*, because its pathology is not a secreted poison — it is the parasite itself consuming and destroying the brush border by massive replication. So the decomposable "machines" are different. From first principles, the parasite's pathology and persistence rest on **four machines**, each of which I decompose into all molecular intervention points:

- **Machine A — the redox economy** (the most fragile system in the organism: one reductase, no backup, import-dependent).
- **Machine B — the egress engine** (how merozoites/gametes get out to reinfect; parasitostatic-class).
- **Machine C — the entry/autoinfection gate** (excystation + the thin:thick fate switch — the within-host amplifier).
- **Machine D — the transmission gate** (the sexual cycle: fertilization + oocyst-wall hardening).

Plus a **fifth, systemic vulnerability**: the parasite is a near-empty metabolic bag — it lost de novo synthesis of fatty acids, sterols, purines, most amino acids, **and glutathione** — so it is utterly dependent on a single physical structure, the **feeder organelle**, to import everything. That dependency is both its greatest weakness *and*, I argue below, a delivery route we can hijack.

I derive 22 intervention points. Where my first-principles analysis happens to land on biology the disease map already discusses (CpTrxR, host FDFT1, CpASP2, Cp23), I add the *decomposition value* — the adjacent intervention points around it. The genuinely novel points (§10) — polyamine ADC, mitosomal Fe-S/ubiquinone, oocyst-wall cross-linking enzymes, HAP2 small-molecule fusion inhibition, the luminal excystation protease, the GSH Trojan-horse, and the T6PP toxic-intermediate strategy — do not appear in the disease map's target list and are where the quarantine earned its keep.

---

## 2. Machine A — The redox economy (decomposed): one pipe, attackable at six nodes

The disease map establishes the single most important druggable fact after the niche: *C. parvum* **deleted glutathione reductase entirely**, cannot synthesise glutathione (GSH), depends on **host-squalene-maintained reduction to import host GSH**, and recycles redox solely through **CpTrxR/Trx**. This is not a target — it is a *single non-redundant metabolic pipe*. A pipe with one valve and no bypass can be attacked at every node, and because there is no backup, blockade is **parasiticidal in principle** rather than merely growth-slowing. I decompose it into six intervention points.
...

phase-3b-survey-report.md — 1075 lines, 126,655 chars

# Phase 3b — Surveyor Survey Report: Bovine Neonatal Cryptosporidiosis

**Program:** cargill-cryptosporidiosis-v11/v1 | **Run date:** 2026-05-30 | **Agent:** Surveyor (computational biology + structural triage)
**Pathogen:** *Cryptosporidium parvum* (taxid 5807 / Iowa II 353152) | **Host:** *Bos taurus* (taxid 9913), calves ≤4 weeks
**Inputs read:** `phase-3-candidates.md` (Forge), `phase-3-vulcan.md` (Vulcan, quarantined), `phase-1-disease-map.md`, `external-input-phase-3.md` (6-model panel) + validator report, `brief.md`.

> **All findings tagged [COMPUTATIONAL] / [VERIFIED]. Computational evidence TRIAGES candidates for experimental validation — it does not promote or kill (Quality Standard 8).** Surveyor reports data; **Reaper** decides kills. No kill decisions are made here.

---

## Method

Every candidate from **Forge** (C1–C10, N2, N3), **Vulcan** (V1–V23, quarantined first-principles), the **Phase-3 6-model panel** (independent target proposals), and two **supplementary** Sentinel/Forge-Watch targets was assessed — **34 candidate workups, no skips** (known targets carry wrong annotations too). Each candidate was processed by an independent subagent: Step 0 identity resolution (UniProt/NCBI Entrez → accession + sequence) → conservation BLASTP (*C. parvum* vs *C. hominis*; the disease is anchored to *C. parvum* — essentially all neonatal-calf disease is *C. parvum*, so Cp↔Ch identity gauges **cross-species breadth**, and the near-identity of these conserved enzymes implies minimal intra-*C. parvum* IIa/IId variation; Quality Standard 16) → host-selectivity BLASTP vs bovine/human proteome (Quality Standard 14) → annotation verification → structure (AlphaFold DB + PDB harvest; AlphaFold-Server Tier-2 requests where a co-fold materially changes the druggability read) → druggability + **activation-vs-inhibition** reality check + precedent-compound verification (PubChem CID + PMID checks). Databases: UniProt REST, NCBI eutils, AlphaFold DB API, RCSB PDB, PubChem, Open Targets; local `blastp`/`makeblastdb`. Full per-candidate workups (commands, e-values, raw hits) persist in `bioinfo/results/`.

**Cross-references:** Sentinel (competitive landscape, `phase-2b`/`phase-3c`) and Apothecary (feasibility, `phase-3d`) outputs were available; their IP/competitive and feasibility flags are reflected in per-candidate **Key Flags** where they bear on a computational finding — notably the **Crick FDFT1 patent EP4598518A1** and the molecule-level kills (auranofin/gold for TrxR = KILL-090; lapaquistat for FDFT1 = KILL-087; BEZ235 for the BCL2A1 axis = KILL-089) whose **target biology Surveyor confirms remains intact** (the kills were molecule-level, not target-level).

---

## Verdict tally

| Verdict | Count | Meaning |
|---|---|---|
| **CONFIRMED** | 16 | Identity, conservation, and annotation match the proposer's claims |
| **CORRECTED** | 6 | Target real, but a claimed property was wrong; data corrects it (candidate may still be viable) |
| **FLAGGED** | 12 | A computational finding creates material risk (not a kill — Reaper decides) |
| **Total** | 34 | All Forge + Vulcan + Panel + Supplementary candidates assessed |

**Regression-core check (crypto = pipeline health check):** the stable core — **FDFT1 (C1), CpTrxR (C2), CpASP2 (C7)** plus the parasitostatic-vs-parasiticidal thesis — recurs and survives computational validation. No novel, unvetted lead appeared as a top target; the new Vulcan/panel entries are either well-grounded enzyme targets or are explicitly FLAGGED for their gaps. The regression signal is **intact**.

---

## Summary table A — identity & verdict

| ID | Target | Category | Conservation | Verdict |
|---|---|---|---|---|
| C1 / V4 | Host FDFT1 (squalene synthase, gut-restricted… | Host-dir. | host-dir (N/A) | CORRECTED |
| C2 / V3 | Parasite CpTrxR (thioredoxin reductase, non-g… | A/B | Cp↔Ch 99.6% | CONFIRMED |
| C3 | Host BCL2A1 (Bfl-1/A1, anti-apoptotic BCL-2 f… | Host-dir. | host-dir 77.7% h/b | CONFIRMED |
...

phase-3c-literature-sweep.md — 192 lines, 32,993 chars

# Phase 3c — Sentinel Run 2: Literature Gap Sweep (post-Forge/Vulcan)

**Program:** cargill-cryptosporidiosis-v11/v1
**Agent:** Sentinel (Run 2 — independent literature gap sweep + load-bearing citation cross-check)
**Date:** 2026-05-30
**Reads from:** `phase-3-candidates.md` (Forge: C1–C10 + N2/N3), `phase-3-vulcan.md` (Vulcan: 23 first-principles intervention points V1–V23), `phase-2b-competitive-landscape.md` (Sentinel Run 1), `brief.md`, `prediction-log.md`.

**Source attribution.** All citation verdicts below were obtained by querying **PubMed** directly (NCBI metadata + abstracts) this run. DOIs are linked per source-attribution requirements. Where a finding is from a preprint server it is flagged as such.

**Panel-input note.** No `external-input-phase-3.verified.md` exists for this run (the panel validator produced `external-input-phase-3.md` + a validator report but no struck-through verified file). Per the Run-2 mandate I worked from Forge's and Vulcan's candidate docs plus independent primary-literature retrieval — not the raw panel — so the missing `.verified.md` does not gate this sweep.

---

## 0. Headline for Reaper (TL;DR)

1. **Citation integrity is clean — the third+ consecutive clean *C. parvum* run holds.** Every one of the **~45 load-bearing PMIDs** across Forge's backbone, Forge's druggability sub-agent, and Vulcan's first-principles literature was verified against PubMed: **zero fabricated PMIDs, zero title mismatches, zero claim-not-found.** Three caveats only: one *species extrapolation* (Vulcan's oocyst-wall-enzyme cite is an *Eimeria* paper applied to *Cryptosporidium*, accurately disclosed by Vulcan), and two *preprints* among the negative-evidence set. Forge's own "no new unverified PMID is load-bearing" claim (§I) is upheld.

2. **The single most important new finding: C5 (parasite aaRS / MetRS) is materially wounded by a calf resistance result Forge did not surface.** Hasan 2021 ([PMID 33753338](https://doi.org/10.1128/AAC.00023-21)) — a dairy-calf trial of the MetRS inhibitor **2093** — saw shedding **resume in 2 of 3 calves by treatment day 5**, with CRISPR-confirmed **D243E (613×)** and **T246I (128×)** resistance mutations. This is "the first report of naturally emerging *C. parvum* drug resistance." Forge called MetRS "the cleanest open aaRS pick"; the most advanced MetRS inhibitor in the calf model has already demonstrated a **low resistance barrier**, and Shaw 2023 ([PMID 38147547](https://doi.org/10.1073/pnas.2313210120)) shows resistance alleles can spread by sexual recombination. **Reaper must add a resistance-barrier kill-test to C5.**

3. **Two genuinely missed glycolysis / metabolic targets.** (a) **CpLDH** (lactate dehydrogenase, [PMID 26562790](https://doi.org/10.1371/journal.ppat.1005250)) — arguably the *better-validated* essential glycolytic enzyme than Forge's CpPyK/C9 (the parasite relies solely on glycolysis for ATP), with published inhibitors (gossypol, FX11). Forge's glycolysis slot is CpPyK-only. (b) **CpAsnA** (type-A asparagine synthetase, [PMID 40483898](https://doi.org/10.1016/j.ijpddr.2025.100601)) — Forge has it as a low-priority Watch, but there is a *published novel-target druggability study* (5,000-compound screen → 3 selective cellular hits, clean target). Both deserve elevation/inclusion.

4. **Forge's C4 (CpeIF4A) survives a false alarm.** A search artifact suggested "no recent eIF4A data," but [PMID 40720562](https://doi.org/10.1371/journal.ppat.1012881) (PLoS Pathog 2025, EC50 1.77 nM, no rebound in NCG mice) is real and on-target. C4 stands.

5. **Vulcan's novel intervention points are citation-clean but mechanistically unvalidated *in C. parvum*** — exactly as Vulcan tiered them. Every primary cite (ADC polyamine, GST, HAP2, excystation EVs, mitosomal Fe-S, oocyst-wall enzymes) is real; the *universal* gap is C. parvum-specific genetic essentiality (most rest on localisation, biochemistry, or cross-species analogy). They are research-tier, honestly.

6. **The mitosome (Vulcan IP-V18/V19) is double-edged after Deng 2025.** [PMID 41369256](https://doi.org/10.1128/mbio.01120-25) (mBio) + [PMID 40875576](https://doi.org/10.1096/fj.202501254R) (FASEB J) **kill the AOX/NDH2 angle** (non-essential, not drug targets) — which *neither* Forge nor Vulcan proposes — while the *same* mBio paper supports Vulcan's framing that Fe-S + ubiquinone biosynthesis are retained and require membrane-potential-dependent import. Net: CpIscS/Fe-S survives only as a speculative research line; its essentiality is **not** KO-confirmed and the ΔΨ generator is now *unknown*.

---

## 1. Targets With New Evidence (Forge proposed, new data found)

| Target (Forge ID) | What Forge said | What this sweep found | PMID | Impact |
|---|---|---|---|---|
| C4 CpeIF4A | "no rebound NCG; EC50 1.77 nM" | Confirmed real & on-target (a "no recent data" search artifact was wrong) | 40720562 | Neutral→**confirms** C4 |
| C6 INS6/INS1 (M16) | INS6 + INS1 validated | Target *class* broader: INS6, INS1, INS-16, INS-19/20 all studied in C. parvum | 40938948, 33688009, 35886965, 39804945 | **Strengthens class; redundancy caveat** |
| C6 druggability | "M16/IDE precedent" | C. parvum-native INS-16 inhibitors (EC50 1.06 & 2.09 µM) already exist | 35886965 | **Strengthens** druggability |
| C2 CpTrxR | auranofin combo 93.5% | Confirmed (Run-1 PMID 42102990); structures confirmed | 40304242, 42162589 | Neutral (holds) |
| C3 BCL2A1 | covalent-peptide-only, Cys55, redox switch | All four Walensky BFL-1 papers confirmed incl. Cys55–Cys175 switch | 27617850, 29276033, 32661419, 32790056 | **Confirms** C3 chemistry read |
| C7 CpASP2 | egress essential, mouse cure | Confirmed; calf counterweight (41885307) already registry-verified | 40901734, 41885307 | Neutral (holds) |
...

phase-3d-feasibility-report.md — 344 lines, 53,373 chars

# Phase 3d — Feasibility Report: Bovine Neonatal Cryptosporidiosis

**Program:** cargill-cryptosporidiosis-v11/v1
**Agent:** Apothecary (translational-pharmacology & feasibility checkpoint)
**Date:** 2026-05-30
**Reads from:** `phase-1-disease-map.md` (Pathfinder), `phase-3-candidates.md` (Forge: C1–C10 + N2/N3), `phase-3-vulcan.md` (Vulcan: V1–V23), `phase-3c-literature-sweep.md` (Sentinel Run 2), `brief.md`, `constraints-ledger.md`. **Surveyor (`phase-3b-survey-report.md`) not yet present** — feasibility assessed from available evidence per the runner note; Surveyor's computational validation should be reconciled against this report when it lands.

**What this document is.** A pharmacology gate, not a biology gate. For every candidate I answer one question — *can this intervention physically reach the target, in a 1–4-week-old pre-ruminant calf, at a viable cost, through a registrable food-animal route?* — across four checks (species ADME, formulation, safety, regulatory) and assign **PROCEED / GATED / BLOCKED**. I do **not** make biological kill decisions (Reaper's job) or adjudicate the τ/parasitostatic question (KE#1). I kill only on delivery, formulation, safety, and regulatory grounds.

> **Apothecary's one-paragraph headline.** The portfolio is dominated by *delivery-class* differences, not target-quality differences. Three facts re-rate everything: **(1)** the patient is a **pre-ruminant calf fed in milk** — the oesophageal-groove reflex shunts the dose past the undeveloped rumen, so milk-vehicle oral delivery of small molecules is far more viable than naïve ruminant intuition allows (Vulcan is right on this — but the groove bypasses the *rumen*, **not** the abomasal pepsin / pancreatic-protease / bile gauntlet, which still degrades oral *proteins*). **(2)** The **five-membrane epicellular barrier** cleanly stratifies the set into three reachability classes — host-directed (bypasses it) > luminal/apical (barrier-free) > intracellular-parasite (existential delivery flag) — and the *single most portfolio-restructuring pharmacology experiment* is an **apical-to-PVM penetration screen in a purging-gut model**, which gates C2/C4/C5/C6/C7/C9 and most intracellular Vulcan IPs at once (the ADME analogue of KE#1). **(3)** **COGS is the under-priced killer**: a disease that costs $33–100/calf supports a drug course of perhaps **$1–5/calf**, which quietly disqualifies macrocyclic peptidomimetics (C7), stapled peptides (C3) and oral biologics (C10b) on economics long before the biology fails. The one intracellular-parasite candidate with *actual in-vivo oral efficacy evidence* (C4/CpeIF4A, via Roc-A) is the delivery proof-of-concept for that whole class; the host-directed backbone (C1) and the luminal class (Vulcan V11/V15/V16/V21) carry the cleanest delivery+residue profiles. **No candidate is BLOCKED on pharmacology alone** — but several are GATED on flags that, if unresolved, convert to kills.

---

## 0. The four invariants that drive every verdict (read once, applied throughout)

These are the species/setting facts the four checks reduce to. I state them once so each candidate assessment can reference them.

**INV-1 — The pre-ruminant milk-delivery window (a delivery *enabler*, partially).** At 1–4 weeks the calf is functionally monogastric: suckled milk/colostrum/milk-replacer triggers the **reticular (oesophageal) groove**, shunting the dose directly to the abomasum and bypassing the undeveloped rumen. So "rumen degradation" — the dominant oral-delivery killer in *adult* ruminants and the one the constraints-ledger names — is **largely circumvented for a drug dosed in the milk vehicle**. **Correction to Vulcan's Lens 1:** the groove solves the *rumen* problem, not the downstream one. After the groove the dose still meets **abomasal acid + pepsin** (neonatal abomasal pH is milk-buffered and higher than the adult ~2, but acidifies through the window), then **pancreatic proteases + bile** in the small intestine before reaching the ileal site of action. Net: oral **small molecules** in milk are well-positioned; oral **proteins/peptides** still face a proteolytic gauntlet and cannot be assumed deliverable just because the rumen is bypassed.

**INV-2 — The five-membrane epicellular barrier is the dominant ADME filter.** Site of action is the ileal enterocyte; the parasite sits under the apical membrane in a PV connected by a feeder organelle (lumen → apical membrane → PVM → feeder organelle → parasite). This stratifies reachability:
- **Host-directed** (target is host-side: FDFT1, BCL2A1, GCLC, actin-pedestal) → **bypasses the barrier entirely**, immune-status-independent, resistance-proof. Best ADME class.
- **Luminal / apical** (oocyst, excystation enzymes, surface rhomboids, Cp23, oocyst-wall enzymes) → **bathed in gut contents → barrier-free** for a non-absorbed agent. Best residue/withdrawal class.
- **Intracellular-parasite** (CpTrxR, CpeIF4A, aaRS, INS6, CpASP2, CpPyK, ADC, mitosome, GSH importer) → must cross all five membranes **and** survive CpMRP1 efflux. The **existential delivery flag**. Apical-to-parasite penetration is **unvalidated for nearly every member** and is the universal first ADME experiment.

**INV-3 — Withdrawal/residue rewards gut-restriction.** Slaughter-cattle withdrawal must be **<30 days** and a novel entity needs an MRL (Maximum Residue Limit) package. **Gut-restricted / non-absorbed** agents (gut-restricted C1; the entire luminal class) carry minimal systemic residue → short withdrawal, lighter residue package. **Systemically-absorbed** agents (C4, C5, any absorbed parasite-directed small molecule) need a full residue/MRL package (≈12–24 months, $5–10M) and a defensible withdrawal time — a cost+time tax, not a blocker.

**INV-4 — The COGS ceiling is brutally low.** Total disease cost is **$33–100/infected calf** (Pathfinder §7.4). A herd-scale prophylactic/therapeutic course must therefore land near **$1–5/calf** to be commercial. This is a hard, modality-level filter: **short-synthesis small molecules** (host-enzyme inhibitors, simple parasite-enzyme inhibitors) clear it; **macrocyclic peptidomimetics, stapled/cyclised peptides, complex natural-product scaffolds, and biologics** face a structural COGS problem regardless of biological merit. I price COGS explicitly per candidate because no upstream agent has.

---

## A. Forge candidate assessments (C1–C10 + N2/N3) — the four checks + verdict

### C1 — Host FDFT1 (squalene synthase), gut-restricted inhibitor
*Forge tier-lean: Tier-1. Class: host-directed, intracellular host enzyme reached luminally.*

**Check 1 — Species ADME: REACHABLE (efficacy-gated).** Host-directed → bypasses the epicellular barrier (INV-2), immune-independent, resistance-proof. Target engagement is **established at the enzyme level**: SQS chemistry is mature (zaragozic acids, quinuclidines, lapaquistat) and lapaquistat blocked *C. parvum* growth in vivo (mouse). Bovine FDFT1 is highly conserved with the human enzyme on which these scaffolds were built, so engagement of the *bovine* enterocyte enzyme is plausible. **The open ADME question is efficacy-of-restriction**, not engagement: does depleting the **enterocyte-local** squalene pool (with a deliberately non-absorbed inhibitor) starve the parasite, or does the parasite-relevant squalene require *systemic* depletion (the exposure that caused lapaquistat's tox)? Unproven for this target — this is prediction P-F3.

**Check 2 — Formulation: FEASIBLE.** Oral, in-milk, gut-restricted small molecule — the ideal fit for INV-1 and the herd setting. The gut-restriction design (minimal systemic absorption while retaining enterocyte mucosal exposure) is the formulation problem, and it has **clinical-grade precedent**: elobixibat, an approved minimally-absorbed ileal-transporter inhibitor (PMID 29959787). Short-synthesis small molecule → COGS-compatible (INV-4).

**Check 3 — Safety: YELLOW.** Two flags. **(a) Inherited hepatotoxic scaffold:** lapaquistat failed on Hy's-Law hepatotoxicity — but at the *systemic cholesterol-lowering* dose. Gut-restriction is the mitigation; the residual risk is incomplete restriction re-exposing the liver. **(b) On-target neonatal-gut tox (the one Forge under-weights):** the sterol/squalene pathway is essential to the *rapidly-dividing enterocytes themselves*. Locally starving sterol synthesis in a **regenerating neonatal gut** — the tissue whose recovery *is* the KPI — could impair epithelial renewal. Needs a neonatal-gut histology + plasma ALT/AST readout.

**Check 4 — Regulatory: NADA, tractable.** Novel small molecule → New Animal Drug pathway. Not an antimicrobial → not on the WHO/OIE critically-important antibacterial list (no AMR-class scrutiny). Gut-restriction → low residue → short withdrawal achievable (INV-3). FTO, not regulatory, is the watch-item: the **Crick FDFT1-in-Cryptosporidium use patent EP4598518A1** must be designed around / filed ahead (flag to Anvil + FTO counsel — outside my remit).
...

phase-4-kill-report.md — 288 lines, 48,595 chars

# Phase 4 — Kill Report: Bovine Neonatal Cryptosporidiosis

**Program:** cargill-cryptosporidiosis-v11/v1
**Agent:** Reaper (red team)
**Date:** 2026-05-30
**Reads from:** `phase-3-candidates.md` (Forge C1–C10 + N2/N3), `phase-3-vulcan.md` (V1–V23), `phase-3b-survey-report.md` (Surveyor — 34 computational workups), `phase-3d-feasibility-report.md` (Apothecary), `phase-1-disease-map.md`, `phase-2-failure-analysis.md` (Sapper), `brief.md`, and the canonical `portfolio.json` / `kill-registry.json`.

> **Reaper's one-paragraph verdict.** I tried to destroy every candidate. **I succeeded on exactly two — and both are Vulcan mechanisms (V15 oocyst-wall peroxidase, V16 oocyst-wall transglutaminase) whose claimed enzymology does not exist in *C. parvum* (it is extrapolated from *Eimeria*).** No Forge candidate fell to a HARD STOP — not because they are unkillable, but because Forge pre-killed the dead material (every brief-excluded molecule) and tiered honestly. The remaining attack surface is wounds and unresolved experiments, not fatal flaws. Three findings change the tiering: **(1)** C2/CpTrxR's "Tier-1" lean is delivery-optimistic — a parasite-side intracellular target with *unproven apical penetration* and an *unvalidated non-gold scaffold* cannot sit in a SAFE tier; **(2)** C5/MetRS carries the single hardest durability wound in the set — **128–613× resistance in 2 of 3 calves within 5 days** of dosing, in the target species, by a single-point mechanism that spreads sexually (PMID 33753338) — so MetRS monotherapy is dead-on-arrival and only PheRS/dual-site survives; **(3)** C3/BCL2A1 is the closest Forge candidate to a kill — there is **no deliverable oral format** for the one undruggable Bcl-2 family member, stacked on neonatal-gut apoptosis safety and over-ceiling COGS. The backbone (C1 host-FDFT1 + C2 parasite-CpTrxR) survives as the only parasiticidal, immune-independent node-pair, but C1's wounds are an **IP/FTO overlap (Crick EP4598518A1)** that directly threatens the brief's "own it outright" requirement, and C2's wound is the existential apical-penetration gate. The entire intracellular-parasite class (C2, C4, C5, C6, C7, C9, and the intracellular Vulcan IPs) shares **one** gate — apical-to-PVM penetration in a purging gut (~$15–30K, C4/Roc-A as positive control) — and the arrest class (C6, C7, C10, V9) shares a second — the **KE#1 calf-τ readout**. Run both before any target-specific spend. **Citation integrity is clean** (see §0): all load-bearing crypto PMIDs verified against PubMed this run; the historic PMID-33753338 "BKI" miscitation is corrected; the Watson-2025 anti-Cp23-mAb negative is now properly reflected.

---

## 0. Kill Test 6 — Citation verification (MANDATORY, mechanical, done this run)

I do **not** inherit upstream verification. Every load-bearing *C. parvum* PMID was pulled fresh from PubMed (title + first author + journal + abstract substance) this run. **Result: clean — no fabrication, no mis-attribution, no cherry-picked-positive among the load-bearing set.**

| PMID | Confirmed as claimed | Bears on |
|---|---|---|
| 40706591 | Marzook, *Cell* — host CRISPR; squalene→GSH | C1 |
| 40304242 | Gabriele, *Biochemistry* — CpTrxR; GR absent | C2 |
| 42162589 | Gabriele, *IJBM* — TrxR essential; gold cmpds | C2 |
| 40720562 | Li, *PLoS Pathog* — CpeIF4A fast-kill, Roc-A | C4 |
| 41991009 | Zhou, *Microb Pathog* — BEZ235→BCL2A1 axis | C3 |
| 40901734 | Wallbank, *MBoC* — ASP2 egress (mero+male) | C7 |
| 41885307 | Nava, *J Infect Dis* — CDPK5-KO calf model | C7 (τ) |
| 40835841 | Watson, *Nat Commun* — essential-gene CRISPR | C10 |
| 40938948 | He, *PLoS NTD* — INS6 proliferation/pathol. | C6 |
| 41765216 | Hayden, *BBA Proteins* — CpPyK SS-dimer | C9 |
| 42118002 | Luo, *FASEB J* — HNF4α↓ microvillus damage | N3 |
| 33753338 | Hasan, *AAC* — **MetRS-2093 resistance, calf** | C5 |

**Two prior-run citation failures are now corrected — I verified both:**

1. **PMID 33753338 is cited correctly this run.** Overwatch v3 and Jarvis v6 both cited 33753338 as "BKI resistance in calves." It is **not** — it is the spontaneous selection of resistance to the **MetRS inhibitor 2093** in a dairy-calf trial. Apothecary and Surveyor both attribute it correctly (MetRS/2093). The historic miscitation did not recur. **This PMID is the load-bearing evidence for C5's resistance wound (§C5) — and it is real.**
...

phase-4b-board-decision.md — 216 lines, 30,115 chars

# Phase 4b — Board Decision: Bovine Neonatal Cryptosporidiosis

**Program:** cargill-cryptosporidiosis-v11/v1
**Agent:** Board (external review + strategic force-ranking)
**Date:** 2026-05-30
**Reads from:** `phase-1-disease-map.md`, `phase-1b-bottleneck-consensus.md` (Tribunal I+II), `phase-2-failure-analysis.md`, `phase-3-candidates.md` (Forge C1–C10 + N2/N3), `phase-3-vulcan.md` (V1–V23), `phase-3b-survey-report.md` (Surveyor), `phase-3d-feasibility-report.md` (Apothecary), `phase-4-kill-report.md` (Reaper), `external-input-phase-4.md` (7-model adversarial panel, validator-cleaned), `validator-report-external-input-phase-4.md`, `brief.md`.

---

> **Board's one-paragraph verdict.** Reaper killed only two targets (V15/V16 — refuted *Eimeria* mechanisms) and the 7-model adversarial panel agrees those kills are correct. But the panel converges hard — **6 of 6 substantive models** — on a verdict Reaper would not pull the trigger on, and I adopt it: **C3/BCL2A1 should be KILLED, not carried as a Tier-3 "asymmetric optionality" bet.** It is undruggable (the one Bcl-2 member every clinical BH3-mimetic program has failed), its target-validation is a downstream-node artifact of a PI3K/mTOR inhibitor, and its mechanism (forced apoptosis of merozoite-containing enterocytes) plausibly *amplifies* autoinfection in a calf with no clearance arm. The panel also forces three tiering corrections that **restructure the top of the portfolio**: **(1)** the Crick method-of-use patent on FDFT1×*Cryptosporidium* (EP4598518A1 — I independently confirmed it: Francis Crick Institute, priority 2022-10-04, names lapaquistat) is a **HARD STOP on the brief's "own it outright" non-negotiable**, not a "wound" — so C1, Reaper's proposed Tier-1 backbone, **cannot be the clean SAFE lead it was sold as**; **(2)** the aaRS class (incl. PheRS), not just MetRS, is dead as a monotherapy product — carry it only as a combination component gated on a *prospective resistance-barrier assay*; **(3)** CpASP2/C7's default outcome is **KILL** (panel puts P[KE#1 favours revival] <30%) — genetic ablation of an egress gene already failed to cure the calf (Nava, verified), and pharmacological inhibition is always less complete than a knockout. The net effect: the brief's central irony surfaces — **the biologically strongest target (C1/FDFT1) has the worst IP for this brief, while the cleanest-IP target (C2/CpTrxR) has the worst delivery uncertainty.** My force-ranking puts **C2 (CpTrxR) first on brief-fit (clean ownable IP + parasiticidal + immune-independent), C1 (FDFT1) second on biology-strength-but-IP-gated, C4 (CpeIF4A) third as the mechanistically-independent cidal with the only in-vivo oral-delivery evidence.** Two cheap, decorrelated master-gates resolve half the portfolio before any target-specific medchem: a **$0 Crick FTO opinion + an in-vitro AP-1 apical-penetration screen + a $5–10K/target bovine-ortholog selectivity panel** must run *before* the **~$70–90K three-arm calf study (KE#1: τ + delivery + modality-class survival)**. Citation integrity is clean: the validator verified all 13 panel PMIDs + the patent + the residue claims against primary sources this run (0 fabrications), and I independently re-confirmed the two most verdict-changing items (Nava PMID 41885307; Marzook PMID 40706591) and the patent.

---

## Step 1 — External Review Panel (synthesis, not re-run)

The Atlas coordinator ran the Board adversarial panel (`cross-check --tier full`, 7 models) and wrote `external-input-phase-4.md`; the panel validator wrote the cleaned version and `validator-report-external-input-phase-4.md`. I did **not** re-run the panel (that would fire a second billed panel and clobber the coordinator's file). I synthesize it below.

**Panel roster & quality.** qwen3.7-max, gpt-5.5, gemini-3.1-pro, ring-2.6-1t, mimo-v2.5-pro, sonar-deep-research, glm-5.1. **gemini-3.1-pro produced only repetitive chain-of-thought with no findings** (its known failure mode) — disregarded, contributes to no convergence count. So the effective panel is **6 substantive models**.

**Headline:** This is a high-consensus, high-integrity panel. It found **zero wrong kills** (all 6 substantive models endorse the V15/V16 kills) and instead argues Reaper was **too generous on survivals**. The signal is unusually corroborated — the convergence is not noise.

After the API panel, this output should also go to the **web-based models** (Claude Web, GPT-5.4 Web, Gemini Extended Thinking) for the human-in-the-loop step before the Cargill conversation — **@Daniel**, that is the one manual step the Board cannot do for you.

---

## Step 1b — Panel Citation Verification Gate (Sentinel-B)

Panel-cited specifics are the known fabrication vector. The panel validator performed a full mechanical PubMed/Google-Patents verification of every load-bearing specific **this run** (functionally a Sentinel-B pass) and found **0 fabricated specifics / ~20 distinct verifiable claims = 0.00**, well under the 0.25 escalation threshold. I treat that as the verified-citations-registry equivalent for citations checked <90 days ago, per protocol, and **independently re-confirmed the three highest-stakes, verdict-changing items** myself (PubMed metadata + Google Patents). No re-fetch of the already-verified remainder.

**Sentinel-B table (consolidated; = independently re-confirmed by Board this run):**

| PMID / ref | Claim it bears on | Verdict |
|---|---|---|
| 40706591 Marzook *Cell* 2025 | C1 FDFT1/squalene; patent source | VERIFIED |
| EP4598518A1 (Crick patent) | C1 IP HARD-STOP | VERIFIED REAL |
| 41885307 Nava *J Infect Dis* 2026 | C7 kill / calf-τ counterweight | VERIFIED |
| 33753338 Hasan *AAC* 2021 | C5 MetRS in-calf resistance | VERIFIED |
| 40901734 Wallbank *MBoC* 2025 | C7 CpASP2 egress / mouse cure | VERIFIED |
| 40720562 Li *PLoS Pathog* 2025 | C4 Roc-A; SI 56k/1.4k; D165/V192 | VERIFIED |
...

phase-5-coverage-map.md — 143 lines, 16,920 chars

# Phase 5 — Coverage Map: Bovine Neonatal Cryptosporidiosis

**Program:** cargill-cryptosporidiosis-v11/v1
**Agent:** Anvil (portfolio architect)
**Date:** 2026-05-30
**Reads from:** `phase-1-disease-map.md` (Pathfinder), `phase-2-failure-analysis.md` (Sapper), `phase-3-candidates.md` (Forge), `phase-3-vulcan.md` (Vulcan), `phase-3b-survey-report.md` (Surveyor), `phase-4-kill-report.md` (Reaper), `phase-4b-board-decision.md` (Board), `brief.md`.

---

## 0. The portfolio-restructuring experiments (KE#1 + AP-1) — NOT YET RUN

> **CRITICAL — read before the coverage numbers.** This program has **two** cheap, decorrelated master-gate experiments that restructure the portfolio, and **neither has been run.** Every coverage estimate and tier assignment below is conditional on them.
>
> 1. **KE#1 — the three-arm challenged-calf study (calf enterocyte turnover τ; ~$70–90K, 8–10 wk).** Decides whether the *arrest class* (CpASP2/C7, INS6/C6, Cp23-MyoA/C10, CpPKG/V9) is investable at all. The one real-host datapoint that exists (Nava CDPK5-KO calf, PMID 41885307) currently leans **against** arrest-only cure. Until KE#1 returns, the arrest class is **do-not-spend**.
> 2. **AP-1 — the in-vitro apical-to-PVM penetration screen (~$15–30K, C4/Roc-A as positive control).** Gates the **entire parasite-directed class** (C2 CpTrxR, C4 CpeIF4A, CpNMT, CpAsnA, CpLDH, C6, C9, C7) simultaneously. It is the single most portfolio-restructuring pharmacology experiment in the program.
>
> A third $0 desk experiment — the **Crick FTO opinion on patent EP4598518A1** — decides whether the strongest biology (C1/FDFT1) is ownable under the brief at all.
>
> **Recommendation: run Tranche 0 (Crick FTO $0 + AP-1 $15–30K + bovine-ortholog selectivity panel $5–10K/target) and Tranche 1 (KE#1 calf-τ ~$70–90K) BEFORE committing any target-specific medicinal-chemistry spend.** Together they can kill or de-risk roughly half the candidate set for under ~$110K combined, and they write the C1/C2/C4/C7 tier structure from decisive readouts rather than inference (Board Step 5). Spending on CpTrxR non-gold chemistry, CpASP2 macrocycles, or M16 inhibitors before these readouts is spending into an unmeasured existential gate.

---

## 1. Threshold lock and the brief's coverage-test override

**Threshold applied (stated up front, applied consistently — no mid-run change):** Cryptosporidiosis is an **emerging/understudied disease** (fundamental biology — the thin:thick fate switch, the feeder-organelle transporters, the neonatal effector gap — is still being resolved). The emerging-disease raw-coverage threshold is therefore **≥60%**, not 70%. There is no `program-context.md` locking a different value; 60% is the disease-class-appropriate bar and is held fixed below.

**Brief coverage-test override (policy reframe #4 — load-bearing).** The brief restricts the portfolio to **novel drug targets Agteria can own end-to-end** — no existing small molecules, antibodies, vaccines, repurposed compounds, or feed additives. The brief states explicitly: *"The standard Anvil 70% tractable-pathology test does NOT apply. Restricting to novel owned drug targets makes full coverage of cryptosporidiosis pathology intentionally unachievable; Anvil documents the coverage gap as a designed-in consequence of the brief, not a portfolio failure."*

I therefore **run the test (it is mandatory and cannot be skipped)** and report all three scores honestly, **then apply the override** to the two stages that have no ownable handle. The override does not waive the test — it reframes a sub-threshold result on the highest-leverage, no-ownable-target nodes as **designed-in**, not as a portfolio failure.

---

## 2. Disease-stage tractability classification

Stages are from Pathfinder §12 (disease-stage → druggable-biology summary). "Tractable" here is the **brief-restricted** definition: *addressable by a novel drug target (small molecule or biologic) Agteria can own outright.* A stage that is biologically real but has **no identified ownable molecular target** is classified NON-TRACTABLE-AS-OWNABLE and reported for completeness — it does not count toward or against the threshold.

| Disease stage | Pathology role | Tractable as ownable target? |
|---|---|---|
...

Output: entities.parquet — 4,466 rows (preview of first 15)

id	name	type	source	method
uniprot:P31483	Nucleolysin TIA-1 isoform p40	Protein	olink_explore_3072	llm
panel:cardiometabolic	Cardiometabolic	Panel	olink_explore_3072	llm
uniprot:P21964	Catechol O-methyltransferase	Protein	olink_explore_3072	llm
uniprot:Q9NRD8	Dual oxidase 2	Protein	olink_explore_3072	llm
uniprot:P16860	Natriuretic peptides B	Protein	olink_explore_3072	llm
uniprot:O60635	Tetraspanin-1	Protein	olink_explore_3072	llm
uniprot:O96017	Ser/Thr-protein kinase Chk2	Protein	olink_explore_3072	llm
uniprot:Q9UKL0	REST corepressor 1	Protein	olink_explore_3072	llm
uniprot:Q8NHS0	DnaJ homolog subfamily B member 8	Protein	olink_explore_3072	llm
uniprot:P58546	Myotrophin	Protein	olink_explore_3072	llm
uniprot:O43854	EGF-like repeat discoidin I-like...	Protein	olink_explore_3072	llm
uniprot:P40225	Thrombopoietin	Protein	olink_explore_3072	llm
uniprot:Q99549	M-phase phosphoprotein 8	Protein	olink_explore_3072	llm
uniprot:P08319	All-trans-retinol dehydrogenase	Protein	olink_explore_3072	llm
uniprot:P25815	Protein S100-P	Protein	olink_explore_3072	llm
...	4,451 more rows

Schema: id (UniProt/panel/compound ID), name, type (Protein, Panel, Disease, Target, Compound, etc.), source_doc_id, extraction_method (llm/regex), community_id, ingested_at

Output: relationships.parquet — 6,851 rows (preview of first 15)

source_id	target_id	type	weight
uniprot:P31483	panel:cardiometabolic	MEASURED_ON	1.0
uniprot:P21964	panel:cardiometabolic	MEASURED_ON	1.0
uniprot:Q9NRD8	panel:cardiometabolic	MEASURED_ON	1.0
uniprot:P16860	panel:cardiometabolic	MEASURED_ON	1.0
uniprot:O60635	panel:cardiometabolic	MEASURED_ON	1.0
uniprot:O96017	panel:cardiometabolic	MEASURED_ON	1.0
uniprot:Q9UKL0	panel:cardiometabolic	MEASURED_ON	1.0
uniprot:Q8NHS0	panel:cardiometabolic	MEASURED_ON	1.0
uniprot:P58546	panel:cardiometabolic	MEASURED_ON	1.0
uniprot:O43854	panel:cardiometabolic	MEASURED_ON	1.0
uniprot:P40225	panel:cardiometabolic	MEASURED_ON	1.0
uniprot:Q99549	panel:cardiometabolic	MEASURED_ON	1.0
uniprot:P08319	panel:cardiometabolic	MEASURED_ON	1.0
uniprot:P25815	panel:cardiometabolic	MEASURED_ON	1.0
uniprot:Q8TE57	panel:cardiometabolic	MEASURED_ON	1.0
...	6,836 more rows

Schema: source_id, target_id, type (MEASURED_ON, INHIBITS, VALIDATED_BY, KILLED_BECAUSE, etc.), weight, description, source_doc_id, ingested_at

Program: Bovine Neonatal Cryptosporidiosis (cargill-crypto-v11) Date: 2026-08-24 Graph: 4,466 entities, 6,851 relationships Runtime: 22 minutes | Cost: ~$1.92 | Success rate: 93%

Pipeline Architecture¶

How Atlas research markdown flows through graffold-ingest into a queryable knowledge graph:

flowchart LR
    subgraph Atlas["Atlas 22-Agent Pipeline"]
    direction TB
    A1[Sentinel] -->|markdown| MD[Intermediate Reports]
    A2[Surveyor] -->|markdown| MD
    A3[Forge] -->|markdown| MD
    A4[Reaper] -->|kill reports| MD
    end

    subgraph Graffold["graffold-ingest"]
    direction TB
    C[Chunk] --> E[LLM Extract]
    E --> R[Entity Resolver]
    R --> P[Publish Parquet]
    end

    subgraph Storage["Knowledge Graph"]
    direction TB
    MG[(Memgraph)]
    DK[(DuckDB)]
    end

    subgraph Query["graffold-api"]
    direction TB
    QA[Query Agent] --> UI[app.graffold.com]
    end

    MD -->|fetch & parse| C
    P --> MG
    P --> DK
    MG --> QA
    DK --> QA

Atlas agents produce intermediate markdown reports (Sentinel surveys, Surveyor analyses, Kill reports, Board decisions). Graffold-ingest parses these into entities and relationships, resolves duplicates, and publishes to Parquet → graph database. The query agent then answers natural language questions over the assembled knowledge graph.

Detailed Pipeline¶

graph TD
    subgraph Atlas["Atlas Pipeline (untouched)"]
    B[brief.md] --> P1[Pathfinder]
    P1 --> P2[Anomaly/Tribunal]
    P2 --> P3[Forge/Vulcan]
    P3 --> P4[Reaper/Board]
    P4 --> P5[Anvil]
    P5 --> MD["phase-*.md files 701K chars / 16 files"]
    end

    subgraph Graffold["Graffold-Ingest (background)"]
    MD -->|filesystem watch| AC[AgteriaConnector]
    AC --> RX[Regex Extract 270 entities, 0.007s]
    AC --> CH[Chunk 192 × 4K chars]
    CH --> LLM["Claude Haiku 4.5 5 concurrent domain-tuned prompt"]
    LLM --> ER[Entity Resolution UniProt / MONDO / PubChem]
    RX --> MG[Merge + Dedup]
    ER --> MG
    MG --> PQ["Parquet Store 4,466 entities 6,851 relationships"]
    end

    subgraph Value["Value Layer"]
    PQ --> QE[Query Engine]
    PQ --> AU[Graph Audit]
    PQ --> VZ[Visualization]
    QE --> PK["prior-knowledge.md"]
    AU --> AR["audit-report.md"]
    PK -->|"feeds back"| B
    end

Extraction Pipeline Detail¶

sequenceDiagram
    participant FS as Atlas Filesystem
    participant AC as AgteriaConnector
    participant CK as Chunker
    participant CL as Claude Haiku 4.5
    participant RS as Resolver
    participant PQ as Parquet Store

    FS->>AC: Read 16 phase-*.md files (701K chars)
    AC->>AC: Regex: 270 entities, 228 rels (0.007s)
    AC->>CK: Split into 192 chunks (4K each)

    loop 192 chunks (5 concurrent)
    CK->>CL: Domain-tuned extraction prompt
    CL-->>CK: JSON {nodes, edges}
    end

    Note over CK,CL: 180/192 succeeded (93%) 6,036 entities, 5,585 rels 22 min total | CK->>RS: Merge regex + LLM results | RS->>RS: Name-based dedup (6,306 → 4,466) | RS->>PQ: Write entities.parquet + relationships.parquet

Results Summary¶

Metric	Value
Source files	16 phase markdown files
Source size	701,788 characters
Chunks processed	192 (4,000 chars each)
Concurrent requests	5
Success rate	93% (180/192)
Total entities	4,466
Total relationships	6,851
Dedup ratio	29% removed (6,306 → 4,466)
Runtime	22 minutes
API cost	~$1.92 (Claude Haiku 4.5)
Output size	146 KB (Parquet)

Graph Connectivity¶

How connected is each entity? The "degree" of a node = how many relationships it has. A degree-1 target has one connection (e.g. "CpTrxR → targets cryptosporidiosis"). A degree-59 disease node is the hub that all targets connect to.

All 363 nodes — connections per entity

157

176

15

012356781230424959

x-axis: number of connections per node · y-axis: how many nodes have that many connections

Target (protein) nodes only — how connected are the drug targets?

27

125

12

2

0 connections1 connection2 connections3 connections

Of 166 drug targets: 125 (75%) connect to exactly one other entity (their disease indication). 27 are isolated (mentioned but not yet linked). Only 14 targets have 2+ connections — these are the well-characterized targets with multiple evidence links.

Node type	Count	Connections	Role in graph
Target	166	1–3 each	Drug targets — leaf nodes connected to disease hubs
Disease	22	up to 59	Hub nodes — aggregate all targets for an indication
Candidate	57	1–3 each	Novel chemistry proposals linked to targets
Evidence	31	0	Publication references (isolated in this subgraph)
Mechanism	22	0	Pathway annotations
Other	65	0–1	Organism, Decision, Compound, etc.

Relationship Type Distribution¶

Relationship	Count	Description
PART_OF	1,765	Target → mechanism/pathway membership
TARGETS_DISEASE	1,638	Target → disease indication
VALIDATED_BY	1,087	Target → publication evidence
INHIBITS	740	Compound → target inhibition
KILLED_BECAUSE	655	Target → kill decision with reason
PROPOSED_FOR	385	Candidate → disease indication
SELECTIVE_OVER	222	Target selectivity over host orthologue
ACTIVATES	90	Direct activation relationships
CAUSES	60	Mechanism → pathology causation
RE_PROPOSED_AS	21	Killed target revived as new candidate

Extraction Prompt (Domain-Tuned)¶

drug-discovery-specific extraction prompt that produces typed entities with properties:

Entity types: Target, Disease, Compound, Mechanism, Organism, Evidence, Decision
Properties: organism, function, essentiality, PMID, DOI, selectivity
Relationships: INHIBITS, VALIDATED_BY, KILLED_BECAUSE, SELECTIVE_OVER, PART_OF

Example extraction (single chunk → Claude Haiku):¶

Input: 4,000 chars of Atlas phase-1 disease biology Output:

{
    "nodes": [
    {"id": "cptrxr", "label": "Target", "name": "CpTrxR",
    "properties": {"organism": "C. parvum", "function": "redox-economy", "essentiality": "high"}},
    {"id": "auranofin", "label": "Compound", "name": "Auranofin",
    "properties": {"mechanism": "thioredoxin reductase inhibitor", "status": "killed"}},
    {"id": "wallbank2025", "label": "Evidence", "name": "Wallbank et al. 2025",
    "properties": {"pmid": "40901734", "journal": "Mol Biol Cell"}}
    ],
    "edges": [
    {"source": "auranofin", "target": "cptrxr", "type": "INHIBITS"},
    {"source": "cptrxr", "target": "wallbank2025", "type": "VALIDATED_BY"},
    {"source": "cptrxr", "target": "cptrxr", "type": "SELECTIVE_OVER",
    "properties": {"note": "no host orthologue"}}
    ]
}

Graph Audit (5 domains)¶

graph LR
    subgraph Audit["Graph-Powered Audit"]
    A1[KILL_CONSISTENCY 10 findings]
    A2[EVIDENCE_COVERAGE 15 findings]
    A3[OMISSION mechanism clusters]
    A4[VERSION_DRIFT cross-run delta]
    A5[NOVEL_CONNECTIONS shared-neighbor prediction]
    end

    PQ[(Parquet 4,466 entities)] --> A1
    PQ --> A2
    PQ --> A3
    PQ --> A4
    PQ --> A5

    A1 --> |"P1: Lapaquistat proposed but KILL-087 active"| R[audit-report.md]
    A2 --> |"P2: CpTrxR lacks VALIDATED_BY link"| R
    A5 --> |"P3: CpRom2 shares 3 neighbors with INS6"| R

Audit Results¶

Domain	Findings	Severity	Example
KILL_CONSISTENCY	10	P1 (major)	Auranofin proposed but KILL-090 active
EVIDENCE_COVERAGE	15	P2 (minor)	CpASP2 lacks validated evidence link
VERSION_DRIFT	1	P3 (info)	301 graph targets not in this run
Total	26	—	Verdict: CERTIFIED

Feedback Loop: Graph → Atlas¶

graph LR
    subgraph Run_N["Atlas Run N"]
    B1[brief.md] --> ATLAS1[Atlas Pipeline]
    ATLAS1 --> OUT1[phase outputs]
    end

    subgraph Graffold["Graffold (accumulates)"]
    OUT1 --> INGEST[Ingest]
    INGEST --> KG[(Knowledge Graph grows each run)]
    end

    subgraph Run_N1["Atlas Run N+1"]
    KG --> |"graffold-ingest query"| PK[prior-knowledge.md]
    PK --> B2[brief.md + prior context]
    B2 --> ATLAS2[Atlas Pipeline]
    ATLAS2 --> OUT2[phase outputs]
    OUT2 --> INGEST
    end

What Atlas receives at startup (prior-knowledge.md):¶

301 explored targets with status (active/killed/wounded)
655 kill decisions with documented reasons (“don’t re-propose without X”)
1,087 evidence links connecting targets to verified PMIDs
222 selectivity assertions (target has no host orthologue)
Mechanism clusters grouping related targets

Net effect: Atlas Run N+1 starts with the full accumulated intelligence of all prior runs. No more rediscovering known landscape.

Target Deep-Dive: CpTrxR¶

Metric	Value
Mentions in graph	92
Relationships	42
Kill status	Active (KILL-090 killed the molecule Auranofin, not the
target)
Evidence	Gabriele 2025 (PMID 40304242), validated sole redox pathway
Selectivity	No host thioredoxin reductase orthologue
Mechanism cluster	Parasiticidal redox economy (with FDFT1)

Cost Analysis¶

Approach	Entities	Rels	Time	Cost	Quality
Regex only	270	228	0.007s	$0	Tables only
Claude Haiku (20 chunks)	728	810	3.5 min	$0.20	Good
Claude Haiku (192 chunks)	4,466	6,851	22 min	$1.92	Full
Claude Opus 5 (192 chunks)	est. ~5,000	~7,500	est. ~3 hr	~$30	Best

CLI Commands¶

# Ingest (one-shot or daemon)
graffold-ingest watch ~/atlas/programs/ --poll 60
graffold-ingest pipeline --source agteria --path dir> --service anthropic

# Query
graffold-ingest query cryptosporidiosis -o prior-knowledge.md

# Audit
graffold-ingest audit program-dir> -o audit-report.md

# Target history
graffold-ingest trajectory CpTrxR

Output Files¶

~/.graffold/parquet/atlas-full/
├── entities.parquet 133 KB (4,466 entities)
├── relationships.parquet 13 KB (6,851 relationships)
├── prior-knowledge.md 4 KB (Atlas startup context)
├── audit-report.md 6 KB (26 findings)
└── extraction-stats.json 155 B (run metadata)

Next Steps¶

Entity resolution at scale — resolve all 1,130 targets against UniProt/MONDO
Community detection — Leiden clustering for global search over mechanism groups
Multi-program accumulation — ingest all 46 Atlas program directories
Parquet-native DRIFT search — multi-hop reasoning without Neo4j
PubMed continuous enrichment — watch for new papers on tracked targets
GNN link prediction — predict novel target-disease associations from graph structure