Research HQ
EarthTalk
Hypothesis status, strategy map, and the live punch-list — pulled from earthtalk/wiki/.
Strategy map
Pulled from earthtalk/wiki/strategy-map.md
Strategy Map
Last updated: 2026-05-10 | Maintained by Alfred. Review weekly (Sunday brief) or when a major decision shifts the picture.
1. North Star
A continent-scale, then worldwide, real-time soil biology dashboard — the Weather Network for soil. Cheap electrodes in the ground, AI decoding what the signals mean, open data standard (FEDS) as the platform layer.
BHAG framing: Every farmer, agronomist, and carbon market verifier has continuous soil health readouts the way every pilot has weather data.
2. Current Wedge
AMF colonization detection for cannabis/MJ growers.
Problem: Grower plants 10 acres with AMF inoculant. After 6 months, 50% of plants failed to colonize — unknown which ones until maturity. Yield loss is significant and invisible until too late.
Solution: In the first 4 weeks after planting, identify which plants didn't inoculate — so they can be swapped before yield is lost.
Sensor geometry: Soil-only electrodes. Near-root vs. far-from-root differential measurement. No plant contact. AMF hyphae extend ~1-5cm into the rhizosphere and change local electrical properties — the gradient between near and far captures their presence.
Second commercial angle: Fusarium detection (same platform, different target signal). TBD on timing.
3. Experiments
Active
- 4-block P. ostreatus recording (2 inoculated, 2 uninoculated), 12 days, 24-bit ADS1220
Planned (near-term)
- Paper 2 experiment (2026-05-11 setup): 16-bit vs 24-bit ADC comparison with light stimuli. Pre-specified analysis required before data collection.
- P. ostreatus presence-detection pilot: heat-pulse or minimal stimulus on 4 existing blocks. Pre-specified. Decision gate for scale-up.
- AMF + plant minimal (3+3): First direct test of near-root vs. far-root differential in colonized vs. uncolonized plants. 14 days, 24-bit.
- Hybrid active + passive: Software-controlled current injection on ESP32 alongside continuous passive recording. Direct comparison to Cseresnyés lineage.
Missing / not yet planned
- Fusarium detection experiment (no design yet)
- Soil heterogeneity baseline study (to characterize confound magnitude)
- n=6+6 AMF replication (needed for statistical power beyond feasibility)
4. Data Strategy
| Layer | What | Why |
|---|---|---|
| Raw | Time-series voltage from differential electrode pairs, 1 Hz, 24-bit | The primary signal; everything else derived from this |
| Environmental | DHT22 temp + humidity per block | Confound characterization and conditioning |
| Stimulus log | What stimulus, when, duration, amplitude | Required for stimulus-response analysis |
| Ground truth | Inoculated / uninoculated status; destructive staining at end | Training labels for detection models |
| Metadata | Substrate type, soil moisture, electrode geometry, plant species, AMF species | Required for generalizing across conditions |
Moat formation: The dataset is the moat. Detection accuracy improves with scale and diversity of training data. No competitor can replicate years of labeled fungal electrophysiology data quickly. Accuracy advantage compounds over time.
FEDS (feds-standard.org) is the platform bet: if the open standard takes hold, all future data in the field uses our schema — giving EarthTalk structural access to external datasets.
5. Partnerships
| Partner | Status | Role |
|---|---|---|
| Andrew Adamatzky (UWE Bristol) | Active — email engagement | Scientific credibility, journal access, potential affiliation |
| Moshe's PhD friend | Informal | Paywalled paper access |
| Fungal Ecology (journal) | Target | Paper 1 submission venue (Adamatzky's recommendation) |
Needed:
- Lab with AMF inoculation capability for controlled experiments (academic partner)
- Agtech investor with soil biology interest (funding)
- Cannabis/MJ grower for pilot validation (commercial wedge proof)
6. Funding Path
Status: pre-funding. Science first.
Sequence:
- Paper 1 on bioRxiv → establishes credibility + FEDS reference point
- Paper 2 (methods, ADC comparison + light stimuli) → shows experimental rigor
- AMF detection pilot results → first commercial signal
- Agtech VC raise — likely Series A target. Frame: detection product for growers + platform play underneath
Grant angle: NSF / USDA / EU Horizon for soil biology + sensor technology. Not yet pursued. Low priority until papers are published.
7. Moat Development
| Asset | Status | How it grows |
|---|---|---|
| Labeled electrophysiology dataset | Early — hundreds of hours of P. ostreatus | Every experiment adds to it |
| FEDS standard | Staked (feds-standard.org) | Papers cite it → external adoption follows |
| Detection algorithm | Not yet trained | Matures as dataset grows; accuracy = moat |
| Adamatzky relationship | Active | Potential affiliation strengthens paper credibility |
| Scientific IP | Papers 1+ | Establishes priority; basis for patents if warranted |
What's exposed: Hardware is commodity (ESP32 + ADS1220). Anyone can replicate the sensor. The data and the model trained on it are the defensible assets.
8. Risks / Unknowns
| Risk | Category | Status | Severity |
|---|---|---|---|
| Passive signal doesn't carry same info as active EIS | Scientific | Open question | High |
| Soil moisture dominates, drowns biological signal | Scientific | Known confound, not yet solved | High |
| Near/far gradient confounded by soil heterogeneity | Scientific | Open question — Cseresnyés didn't test this geometry | Medium |
| P. ostreatus → AMF transfer assumption fails | Scientific | Hypothesis, not confirmed | High |
| Plant root electrical activity masks AMF signal in soil | Scientific | Open question (Moshe's hypothesis: geometry solves it) | Medium |
| n=3+3 too underpowered to detect modest AMF effect | Scientific | Known; treat first AMF exp as feasibility | Medium |
| Independent researcher affiliation for Paper 1 | Strategic | Open — resolve before bioRxiv | Medium |
| "FEDS already published" claim in vision doc | Strategic | Needs softening — site exists, adoption zero | Low |
Filter
Before any new experiment, feature, or direction:
"Does this move us closer to a validated, fundable, scalable system?"
If not → deprioritize or discard.
Cross-references
- Hypothesis tree — 8-rung scientific claim chain, confidence levels
- Bridge gap — honest framing of Paper 1 vs. vision
- AMF experiment — detailed experimental plan
- Todo — live punch list
- Decisions log — what was decided and why
Hypothesis tree
Pulled from earthtalk/wiki/hypothesis-tree.md
Hypothesis tree
Last updated: 2026-05-10
The vision document is prose; this is the load-bearing structure underneath. Each claim is rated: falsified / unfalsified-no-evidence / unfalsified-some-evidence / supported / proven.
Root: a real-time map of soil biology is achievable using electrical sensors + AI
H1. Fungal mycelial networks produce measurable, structured electrical signals
- Status: Supported.
- Evidence: Adamatzky lab (2018, 2022); Buffi et al. 2025 (FEMS Micro Rev); Paper 1 directly.
- Key uncertainty: Which signal modality dominates (action-potential-like spikes, ion-current oscillations, membrane-potential rhythms, streaming potentials) — see methods.md. Vision doc currently glosses this; the field has not settled.
H2. Those signals contain biologically meaningful, decodable information
- Status: Unfalsified-some-evidence.
- Evidence: Paper 1 demonstrates phase-locked 24h coherence — i.e., the signal has temporal structure. That is not the same as decodable meaning.
- Gap: No paper yet shows "signal feature X → biological state Y" mapping.
H3. Signals are robust to field conditions (vs. lab substrate blocks)
- Status: Unfalsified-no-evidence yet.
- Gap: All published work to date (including Paper 1) is laboratory substrate. Soil, with its electrochemical complexity, moisture variation, ion concentration shifts, root exudates, and bulk-impedance variability, is a different beast.
- Required experiment: field-deployed electrode in soil with active AMF, demonstrating signal extraction comparable to lab.
H4. The relevant biology in agricultural settings is AMF specifically
- Status: Unfalsified-no-evidence in our work.
- Gap: Paper 1 uses Pleurotus ostreatus — a saprotroph, not mycorrhizal. Its electrical behavior may not generalize to AMF. AMF cannot be cultured in pure form (obligate symbionts), making electrophysiology on AMF specifically much harder.
- Required experiment: AMF-colonized root/soil system with electrode recording.
H5. Sensors can be built at smartphone-accessory price points and deployed at scale
- Status: Unfalsified-no-evidence yet.
- Evidence: ESP32 + ADS1115 BOM is genuinely cheap. Field-deployable, weatherproof, low-power, networked version with calibration and durability — different problem.
- Required: field-grade hardware prototype + cost-down analysis.
H6. AI can map signal → soil-biology state with enough labels
- Status: Unfalsified-no-evidence yet.
- Gap: No labeled dataset of "signal X observed when AMF stressed in way Y" exists. The interpretation engine is chicken-and-egg with deployment.
- Initial route: controlled greenhouse experiments with imposed stresses (drought, pathogen, nutrient deficit) and parallel signal recording.
H7. The market values continuous soil-biology readouts at scale
- Status: Plausible but unsupported.
- Gap: Insurance, food security, carbon markets — each named in vision-v3 — has very different sales cycles and willingness-to-pay. No customer-discovery interviews yet documented.
H8. An open data standard (FEDS) + proprietary value layer is the right business model
- Status: Plausible (pattern proven elsewhere — Linux/RedHat, weather data) but unsupported here.
- Gap: FEDS adoption outside the founding team is zero based on 2026-05-10 web search. The standard exists; the coalition does not yet.
How to read this
- H1 is real. The rest is the research program.
- Paper 1 advances H1 and weakly H2. Everything else is still open.
- The investment case currently rests on H1 being foundational enough that the rest can be built — Karpathy/Tan-style "build infrastructure while science finishes."
Punch-list
Pulled from earthtalk/wiki/todo.md
TODO — live punch list
Last updated: 2026-05-10
Sorted by priority and dependency. Check off as completed; do not delete — move to "Done" section at bottom.
Paper 1 — before bioRxiv submission
Substance edits
- Consolidate temporal-separation-of-controls defense into a single Discussion paragraph (currently scattered). Explicitly acknowledge "truly concurrent inoculated/control recordings would further strengthen the conclusion." (See reviewer-attacks.md #1.)
- Add one-line discriminator callout adjacent to the "dead-mycelium controls show 62.5% Lomb-Scargle detection" number — point to phase-coherence (R=0.054 vs 0.83) as the resolution. (Attack #3.)
- Introduce pilot-study framing in Intro or Discussion to preempt n=4 attack. (Attack #2.)
- Gesture at a candidate mechanism in §4.5 (membrane potential, ion transport, or growth-driven streaming potentials). (Attack #6.)
- Justify the 30-second window length and z>2 threshold briefly in §2.4 or supplementary. (Attack #7.)
References (See references-to-chase.md.)
- Add Lomb 1976 + Scargle 1982 for periodogram methodology.
- Add Berens 2009 (or Fisher 1993) for circular statistics / Rayleigh / Watson-Williams.
- Add 2–3 more Adamatzky papers beyond 2018 Sci Reports.
- Add Olsson lab citation(s).
- Add a more recent fungal circadian review beyond Dunlap 1999.
Operational
- Pre-deposit data + analysis code on Zenodo or OSF. Get a real DOI. Replace
[repository URL]placeholder in Data Availability before submission. Non-negotiable. - Resolve Adamatzky co-authorship / affiliation question before bioRxiv. See adamatzky-thread.md. Author-list changes after preprint are awkward.
Paper 2 — ADC methods + light stimuli (planned, experiment 2026-05-11)
- Pre-specify analysis plan before running experiment — what exactly is being tested: (a) 16-bit vs 24-bit resolution difference in stimulus-response capture, (b) light stimulus detectability in live fungus vs control, or both? Define hypothesis, metrics, thresholds, and analysis before data collection.
- Define light stimulus protocol — wavelength, intensity, duration, interval, on/off cycles. Must be non-environment-changing (consistent with stimulus category rules).
- Design control — dead mycelium or uninoculated substrate as control, same stimulus applied. Hardware-matched across 16-bit and 24-bit channels if possible.
- Run experiment (Moshe setting up 2026-05-11)
- Write short methods paper
Vision document
- Reconcile "FEDS is already published" claim with reality (web search 2026-05-10 surfaced no external usage). Either soften to "we are proposing FEDS" or, if the standard genuinely exists, link to where it lives.
- Source the strong stats: "84% of agricultural inputs fail," "40% of crops lost," "three weeks before any human eye." Citations or softening.
- Commit explicitly to which electrical phenomenon (action-potential-like spikes vs impedance vs streaming potentials) — currently the doc waves at "electrical activity."
- Name the smartphone-accessory price target explicitly so it can be argued.
- Add a "why not them" paragraph addressing Biome Makers / Trace Genomics / CropX / Indigo / hyperspectral incumbents.
Project infrastructure
- Stand up
earthtalk/workspace per Karpathy LLM Wiki pattern. - Schedule Sunday 7am research brief (see research-log).
- Import Adamatzky email correspondence into
sources/adamatzky-correspondence/when Moshe shares. - Decide on hypothesis-tree status reviews — monthly recalibration.
Done
(none yet)