PhytoRemedy

Method — how this tool produces its numbers

This page is the full account: what goes in, what happens to it, what comes out, what the model deliberately refuses to compute, and where it is most likely to be wrong.

The governing principle

Every coefficient is traceable to a named source, or it does not exist. Where a quantity is needed but nothing publishes it, the model returns a gap — a visible statement of what is missing and which output it blocks — rather than a plausible-looking number. EDTA and EDDS are named in the sources as tested and effective on lead, but neither comes with a published fold-change, so both are offered in the form and apply no numeric effect at all. That refusal is the feature.

A separate model block holds structural choices — the simulation horizon, how sub-scores combine — which assert no empirical fact. They are kept apart so it is always obvious which numbers are evidence and which are architecture.

The sources

Copies of all seven are in research/.

Step 1 · Location becomes a climate

The location string resolves to coordinates: postal codes via Zippopotam.us, place names via the Open-Meteo geocoder, or raw latitude/longitude accepted directly. Those coordinates pull ten years of daily ERA5 reanalysis from the Open-Meteo historical archive — maximum and minimum temperature, precipitation, and FAO-56 reference evapotranspiration (ET₀).

Ten years of daily records reduce to a 365-day normal series, from which the model derives:

• USDA hardiness zone — from the mean annual coldest night, in 10 °F bands from −60 °F.
• Thermal growing season — the longest run of days whose 5-day running mean temperature exceeds 5 °C. In the southern hemisphere the year is rotated by half before searching, so a season straddling 1 January is found correctly.
• Annual rainfall and ET₀, and the ratio between them.

Step 2 · Site match

Each species is scored on five envelopes: winter hardiness, growing-season temperature, water supply, soil pH, and whether the season is long enough to complete one growth cycle. Each envelope is a trapezoid — the score is 1.0 between the published optimum bounds and ramps to 0 at the published tolerance limits.

Where a source gives an optimum but no tolerance limit, and the site falls outside that optimum, the factor returns unknown and is excluded from the score rather than guessed. Hardiness is skipped entirely for annuals, which are grown inside the frost-free season — season length already covers that.

Surviving factors combine as an unweighted geometric mean. Geometric because these factors are limiting rather than additive: a plant that cannot survive the winter is not "70% suitable", and a zero in any factor correctly drives the whole score to zero. Unweighted because no source ranks the factors against one another, and inventing weights would be exactly the false precision this tool avoids.

Step 3 · Water

The FAO-56 single crop coefficient method:

ETc  = Σ over growth stages ( ET₀ × Kc )
need = max(0, ETc − rainfall)

The cycle splits into FAO's four stages — initial, development, mid-season, late — because evaporative demand swings widely across a season and a single average would smear that out. K_c ramps linearly through development and late season, so each ramp contributes the midpoint of its endpoints.

Only sunflower has a direct FAO-56 entry. B. juncea borrows "Rapeseed, Canola" — same family, and FAO's 0.6 m crop height matches the 0.3–0.6 m your comparison table reports. That substitution is flagged on every result. Amaranth has no FAO-56 entry at all, so no water requirement is computed for it.

Rainfall is counted at face value. No "effective rainfall" discount is applied because no source here quantifies one, which means the modelled irrigation requirement is a lower bound — real losses to runoff, deep percolation and canopy interception typically run 15–35%.

Step 4 · Lead uptake

The core of the model, and the part built entirely on [R13]:

shoot Pb (mg/kg) = BCF × fpH × amendment × soil Pb (mg/kg)

The bioconcentration factor is shoot lead divided by soil lead, measured in the unfertilised control at soil Pb 40.2 mg/kg and season-mean pH 6.88:

A trap in [R13] Table 3: its "Total" column is the root concentration plus the shoot concentration — for sunflower at N 380, 6.1 + 15.5 = 21.6 mg/kg. Adding concentrations from two tissues is not a meaningful concentration (the real whole-plant figure is 14.3 mg/kg), and using it would overstate the BCF by about 60%. This model uses the Shoot column only, because roots stay in the ground and their lead never leaves the site.

The pH multiplier, and why it is derived rather than assumed

[R13] attributes its nitrogen effect to acidification: the transfer factor "was much higher for Pb than for Cu... attributed to high Pb mobility in the soil due to the decreasing soil pH." That gives a measured pair. At 380 mg N/kg, season-mean pH fell 6.88 → 6.08 (Δ = −0.80) while mean shoot lead rose ×1.90. So:

fpH = exp( k × (6.88 − site pH) )
k    = ln(1.90) / 0.80 = 0.802

Checked against the other rate: at 190 mg N/kg, Δ = −0.64 predicts ×1.67 against a measured ×1.75 — inside the species spread (×1.69, ×1.84, ×1.67).

Because pH carries the entire nitrogen effect, the nitrogen amendments apply no separate multiplier. Doing both would double-count. This is also why soil pH is the highest-leverage input on the form.

No clamp is applied to f_pH. Outside the 6.08–6.88 range over which it was measured, the model reports that it is extrapolating instead of silently capping.

Amendments

Two mechanisms, both sourced, and a third state for honesty:

• pH shift — nitrogen at 190 and 380 mg N/kg shifts season-mean pH by −0.64 and −0.80 [R13] Table 1. The uptake effect then follows from f_pH above.
• Direct multiplier — citric acid, ×7.0, from the 6.7–7.2-fold shoot lead increase [PCT] reports on vetiver. The only chelate with a published fold-change anywhere in these sources.
• Unquantified — EDTA and EDDS are named as effective but no source gives a magnitude. They are offered in the form, apply no numeric effect, and say so. In reality EDTA raises uptake substantially and mobilises lead into groundwater, where it persists for years.

Step 5 · Lead removed, and why yield is your input

soil mass (kg/ha)  = 10,000 m² × depth (m) × bulk density (kg/m³)
Pb removed (kg/ha) = yield (t/ha) × shoot Pb (mg/kg) ÷ 1000 × cycles/year

Soil mass defaults to a 5 cm treatment depth. [MC] collected cores "from a depth of 5 to 10 cm (depending on penetrability)", so 5 cm is the shallow end of the range those lead measurements describe, and therefore the most optimistic reading of them. Because this input moves the answer more than any other, results always show 5 / 7.5 / 10 / 15 cm side by side rather than one figure alone. Use 10–15 cm if your plot is tilled — tillage genuinely mixes surface lead downward, and the deeper slab is then what you farm.

One caution on the shallow end. The lead concentration is an average over the whole sampled slab. Halving the depth halves the lead in this arithmetic, but physically the lead does not vanish — if it is concentrated near the surface, which [MC] explicitly suspects ("surface deposition of Pb would have been diluted"), then the top few centimetres carry a higher concentration and the mass falls far less. A shallow depth paired with a slab-average concentration takes the optimism twice.

Bulk density is the NRCS "ideal for plant growth" figure for your texture — the lowest of the three NRCS values — so modelled lead mass and clean-up time are both lower bounds: a compacted site holds more lead and takes longer.

Yield is a user input because no source in this set publishes field dry-matter yields for these species. [R13] reports grams per pot with no pot volume given, which cannot be converted. Rather than invent a tonnes-per-hectare figure — a number that multiplies every lead result linearly — the tool asks for it. Leave it blank and you still get shoot concentration and lead removed per tonne of harvest, both fully sourced; annual totals and the timeframe are withheld.

The plants account for 0.08–0.29% of the lead in the pot. Table 4 reports 20–54%. Either that percentage is measured against the plant-available fraction rather than total lead, or it reflects a soil re-measurement in which most of the change was not plant uptake. Either way it cannot be applied to total soil lead, which is exactly what a reader would do with it.

Tables 2 and 3 are mutually consistent — shoot dry matter × shoot concentration reproduces the reported shoot accumulation in 8 of 9 treatments. The bioconcentration factors rest on that consistency, and it is why they are trusted while the removal percentage is not.

The model checked against the experiment

Removal per cycle is plant mass × BCF ÷ soil mass — soil lead cancels out, so a pot and a hectare can be compared directly:

                    pot        field @5cm       field @7.5cm
Sunflower        0.1160%    0.2713% (2.34×)   0.1809% (1.56×)
Amaranth         0.0770%    0.1100% (1.43×)   0.0733% (0.95×)
Indian mustard   0.0671%    0.0783% (1.17×)   0.0522% (0.78×)

At either depth the model lands within about a factor of two of what the experiment measured — the closest thing this tool has to a validation. Note that the agreement is tightest around 7.5–10 cm, where the three ratios straddle 1.0 rather than all sitting above it. That is a mild hint the deeper end of the sampled range is the more realistic one, though it depends on the assumed field yields, so it is not decisive.

It also corrects a tempting explanation. The plant-to-soil mass ratio is the same order of magnitude in a pot and a field — 0.0055 against 0.0129 kg of shoot per kg of soil at 5 cm, a factor of 2.3, not 100. Phytoextraction is not slow because fields scale badly. It is slow because plants remove roughly 0.1% of soil lead per crop anywhere. The 42% headline is the outlier, not the field projection — and that agrees with [Z24], which lists "the slow pace of phytoremediation" as a primary weakness of the technique.

Step 6 · Draw-down simulation

The model steps year by year. Each year's removal lowers soil concentration; next year's shoot concentration is computed from the lower soil concentration, so removal falls too. Draw-down is therefore roughly exponential and the tail is long — the last stretch to a low target takes disproportionately longer than the first.

Because shoot lead is strictly proportional to soil lead, and nothing else in the model depends on concentration, the draw-down is exactly geometric and is solved in closed form rather than iterated:

f     = yield × cycles × BCFeff × 1000 ÷ soil mass     (fraction removed per year)
C(t)  = C₀ (1 − f)^t
years = ln(target ÷ C₀) ÷ ln(1 − f)

So the tool always reports a specific number of years, however large. A projection of 250 years tells you far more than "more than 100". Past 100 years the result carries a warning — that is where a projection stops being a plan and becomes an illustration — but the number is still shown, and it is exact for the model's assumptions.

How candidates are ranked

1. Species with a measured bioconcentration factor above those without.
2. Shortest time to target, where a yield makes that computable.
3. Otherwise, most lead removed per tonne of harvest.
4. Otherwise, best site match.

What was deliberately removed

An earlier build of this tool carried a cost model, phytotoxicity thresholds, water-stress yield penalties, harvest-loss factors, per-texture lead bioavailability multipliers, per-species soil-texture preference scores, EDTA and EDDS magnitudes, and nine plant species. None of it had a source. All of it is gone.

Peat and urban fill are no longer offered as soil textures because NRCS Table 1 does not cover them — which is a real loss, since [MC] found Oakland's worst lead in made ground. Irrigation efficiency tables were replaced by a single input: millimetres available per season, which is a site fact rather than a constant.

The model also no longer scales yield down for water stress. Instead it reports the supply-to-demand ratio and the shortfall in millimetres, and leaves the judgement to you.

Making it faster — what the arithmetic says

The removal rate is a single product: f = yield × cycles × BCF ÷ soil mass. Every term multiplies, so it does not matter which one you improve — only by how much. Working backwards from the Fremont example (273 → 80 mg/kg, sunflower at 9 t/ha × 4 harvests, effective BCF 0.286, 250 years):

Look at the 5-year row: you would need shoot tissue above 1,000 mg/kg — the hyperaccumulator threshold itself. Finishing a site in a handful of years requires a genuine hyperaccumulator, by definition. (These figures assume the optimistic 5 cm depth; at a tilled 15 cm every required BCF triples.)

Why a higher BCF alone may not help

The obvious move is to find a plant that concentrates lead harder. But BCF is only one of three multiplying terms, and the species with extraordinary BCFs tend to be small alpine herbs with very little biomass. Run the arithmetic on that trade-off:

‡ Illustrative arithmetic only. These rows are not in the tool — no source in research/ gives a soil-paired BCF for a lead hyperaccumulator, and the EDTA multipliers are unverified figures from a study whose full text was not accessible.

The second row is the point. A plant with fourteen times the bioconcentration factor finishes slower than sunflower, because it grows twenty-four times less biomass. Classic lead hyperaccumulators such as Thlaspi rotundifolium — reported around 8,200 mg/kg shoot lead — are small alpine Brassicaceae yielding perhaps 1–2 t/ha against sunflower's 36 t/ha/yr across four harvests. Concentration and tonnage matter equally, and the literature selects hyperaccumulators on concentration alone.

Which is why the strongest lever here is not a different plant but a chelating amendment on a high-biomass crop — the only route in this table that reaches a human timescale. It carries a real cost: mobilised lead moves toward groundwater, and EDTA persists for years.

The levers that do not involve plants at all

Measuring the depth of contamination is free and moves the answer by a factor of three across the 5–15 cm range — more than any plant choice in this tool. Choosing a defensible target halves it again.

Where this is most likely to be wrong

1. BCF extrapolation. Every bioconcentration factor was measured at soil Pb 40.2 mg/kg. A typical contaminated site runs 5–25× that. Bioconcentration factors generally fall as soil concentration rises, so the model likely overestimates uptake on hot sites. Results flag the extrapolation factor.
2. Pot to field. [R13] is a 56-day pot trial. Field rooting, competition, weather and soil heterogeneity all differ.
3. Only phytoextraction is modelled. Vetiver grass was removed in the 2026-08-29 revision: it is a phytoSTABILISER, putting roughly 6% of plant lead in the harvestable shoot, so it answers a different question from the three species here. Stabilisation and capping remain valid strategies — this tool simply does not score them. See the note below.
4. One pH curve, one soil. The pH sensitivity comes from a single sandy loam. Lead speciation differs enormously between soils — carbonate-bound lead in alkaline urban fill behaves nothing like lead in an acid sand.
5. Spatial variability is ignored. Real sites are patchy and hot spots dominate risk. [MC] found lead lognormally distributed with a skewness near 4 — a site mean tells you much less than you would expect.
6. Depth. Lead below 15 cm is not addressed at all.

Implementation

Static site, no build step and no dependencies — plain ES modules served over HTTP. Data lives in src/data/ (constants, soils, plants), the calculation in src/lib/engine.js, and the interface in src/ui/. Climate comes from the Open-Meteo historical weather and geocoding APIs and Zippopotam.us; none needs an API key. Nothing you enter leaves your browser except the location, which is sent to the geocoder.

Results — what building it under that constraint produced

PhytoRemedy was built around a single constraint: every coefficient must be traceable to a published source, or the tool returns no number at all. Where a quantity is required but nothing in the literature supplies it, the model records a gap — naming what is missing and which output it blocks — rather than substituting a plausible value. This rule shaped the tool more than any design decision, and it was expensive. An earlier build carried a cost model, phytotoxicity thresholds, water-stress yield penalties, per-texture bioavailability multipliers and nine plant species; all of it was removed for lacking support, leaving three species and roughly thirty parameters that survive scrutiny.

The lead-uptake core rests on [R13], a 56-day pot trial of amaranth, Indian mustard and sunflower grown on industrial sandy loam at soil Pb 40.2 mg/kg and initial pH 6.9, under three nitrogen rates. Its Table 3 supplies every bioconcentration factor here — shoot lead divided by soil lead, taken from the unfertilised control and from the Shoot column rather than the Total column, which sums root and shoot concentrations and would overstate the factor by 62% while counting lead that never leaves the site. Its Table 1, in which nitrogen fertiliser depressed season-mean soil pH from 6.88 to 6.08 while shoot lead rose 1.90-fold, yields the model's one derived constant: a pH sensitivity of ln(1.90) ÷ 0.80 = 0.802, which was then validated at the paper's second nitrogen rate, predicting a 1.67-fold rise against a measured 1.75-fold. Everything else is read straight off a source. Species growth envelopes come from [PCT] and its underlying horticultural sources; crop coefficients and rooting depths from [FAO56]; bulk density from [NRCS]; and soil lead scenarios, regulatory targets and sampling depth from [MC]. Climate is not stored at all — ten years of daily ERA5 reanalysis are fetched for your coordinates at run time, so any location worldwide can be assessed without adding a single stored number.

The model was then validated against the experiment that produced it. Because removal per cycle is plant mass × BCF ÷ soil mass, and soil lead cancels from both sides of that ratio, a 2 kg pot and a one-hectare field are directly comparable; the field projection reproduces the pot trial's own measured uptake within a factor of two for all three species. The same arithmetic exposed an inconsistency in the source. [R13]'s Table 4 reports 20–54% of soil lead removed per cycle, but the plant tissue in its Table 2 accounts for only 0.075–0.286% of the 80.4 mg of lead present in each 2 kg pot — a discrepancy of 188 to 315 times. That headline percentage is therefore excluded from the model, and the exclusion, with its reason, is stated in the tool rather than buried in a comment.

What the tool does with those parameters is score three candidate species against a site, then project lead removal, timeframe, water demand and — from rates you supply — cost. Site matching combines five envelopes as an unweighted geometric mean, so a single disqualifying factor collapses the score rather than being averaged away: a plant that cannot survive the winter is not seventy per cent suitable. Factors without a published envelope are excluded from the mean and reported as excluded, not assumed favourable, and that coverage is uneven — amaranth has no published temperature tolerance at all. Draw-down is solved in closed form rather than simulated: because shoot lead is strictly proportional to soil lead, the yearly decline is exactly geometric, so the tool reports a specific number of years however large it is, instead of truncating at some arbitrary horizon and saying "more than fifty years". Each result carries the model's uncertainties on its face — the extrapolation distance from the 40.2 mg/kg at which every BCF was measured, whether site pH falls outside the 6.08–6.88 range over which the pH response was derived, and whether predicted shoot lead exceeds the 1,000 mg/kg hyperaccumulator threshold that none of [R13]'s species reached. And because treatment depth moves the answer more than any other input while being the least certain, every result is reported across a depth ladder — 5, 7.5, 10 and 15 cm — rather than at one assumed depth.

Four examples show what that produces. Screening a site: for ZIP 94539 (Fremont, California) at 273 mg/kg soil lead, the West Oakland median, sunflower ranks first, removing 2.81 kg Pb per hectare per year across four harvests and reaching California's 80 mg/kg screening level in roughly 83 years at 5 cm depth, or 250 years if the plot is tilled to 15 cm. The tool simultaneously flags that the site's 14.75 °C growing-season mean sits below sunflower's published 18–25 °C optimum, and that 273 mg/kg is 6.8 times the concentration at which its bioconcentration factor was measured.

Comparing amendments: nitrogen fertiliser at 380 mg N/kg acidifies the soil from pH 6.5 to 5.7, nearly doubling shoot lead from 78 to 148 mg/kg and halving the timeframe to 43 years. The tool reports in the same breath that pH 5.7 sits at the floor of sunflower's tolerated range, collapsing its site match to zero — surfacing the central tension of induced phytoextraction, that the acidity which mobilises lead also stresses the crop meant to take it up.

Testing whether a better plant exists: working backwards from the removal equation, finishing the Fremont site within five years would require an effective BCF of 4.23, equivalent to 1,155 mg/kg of shoot tissue — above the hyperaccumulator threshold, by definition. But because BCF, yield and harvest frequency multiply, a hypothetical species with fourteen times sunflower's BCF and one twenty-fourth its annual biomass finishes slower, not faster: 142 years against 83. The measured values make the same point, since sunflower has the lowest shoot-to-root ratio of the three extractors yet the highest BCF and the highest biomass, and outperforms both of the others. Concentration and tonnage matter equally, though the hyperaccumulator literature selects on concentration alone.

Identifying what is not known: harvest yield illustrates the refusal to guess. Annual lead removal and every timeframe scale directly with tonnes of dry matter per harvest, and no source consulted here publishes a field yield for any of these species — [R13] reports grams per pot without giving the pot volume. Rather than assume one, the tool asks for it, and withholds annual totals and the timeframe until it is supplied while still reporting shoot concentration and lead per tonne of harvest, both fully sourced. The gap is visible in the interface rather than hidden in the data files.

Footnote — vetiver grass, and why it is not here

Earlier revisions of this tool carried a fourth species, Chrysopogon (Vetiveria) zizanioides. It was removed on 2026-08-29 on grounds of scope: vetiver is a phytostabiliser, not an extractor. Soil-grown measurements put roughly 6% of its lead in the harvestable shoot — Rotkittikhun et al. (2010) report 38 mg/kg in shoot against 629 mg/kg in root — and every source consulted agrees it restricts root-to-shoot translocation. This tool models lead leaving a site in a harvest, so vetiver was answering a different question from the three species that remain.

A correction to what earlier versions of this page claimed: published lead data for vetiver does exist. It is not usable here, for three separate reasons. Wilde et al. (2005) report a bioconcentration factor of 88, but that is hydroponic tissue ÷ solution with a chelator, where [R13]'s factor is shoot ÷ soil — the same word for quantities about 400-fold apart. Ng, Boyce, Rahman & Abas (2016) and Ng et al. (2019) are soil-grown from the same laboratory as [R13], and the 2019 paper computes a bioconcentration factor directly, but every figure reachable is EDTA- or nitrogen-amended rather than the unfertilised control this model is calibrated against. And all of them work at soil lead of 200–1000+ mg/kg, where this model hard-codes a single global measurement concentration of 40.2 mg/kg for every species. Since bioconcentration factors fall as soil lead rises, importing one measured at 25× that concentration would rank vetiver as a weak extractor for a reason that has nothing to do with the plant.

Stabilisation and capping remain legitimate responses to lead contamination, and for a site where children are exposed they act far faster than any extraction crop. This tool does not score them. That is a boundary, not a judgement.

Applied case — a West Oakland yard, the most at-risk soil in [MC]

Everything above is general; this section runs the tool on the one place the source data actually describes, and on the worst of it. [MC] sampled Oakland at three scales — 113 city-wide points, 116 West Oakland yards and 11 gridded sites — and found risk unevenly distributed: lead concentrated in West Oakland, the oldest and most industrial part of the city, and near San Leandro Bay and the airport, while the newer hills were cleanest, and by land use vacant lots and gardens carried more than parks and parks more than open space, so the ground people would plant on is the ground carrying the most lead. Both distributions were strongly lognormal (city skewness 3.957, neighbourhood 4.16), so the median is the honest central figure: 273 mg/kg, the West Oakland median, against California's 80 mg/kg CHHSL. The tool was run at ZIP 94607 (37.81 °N, −122.29 °E) on sandy loam — the [R13] trial texture, and so the least extrapolated choice — with no pH measurement, so the [R13] reference pH of 6.88 is assumed, and a treatment depth of 5 cm. Ten years of ERA5 reanalysis for those coordinates give a mean annual temperature of 14.2 °C, 665 mm of rain against 1,122 mm of reference evapotranspiration, a mean coldest night of 1.6 °C (USDA zone 10a) and, because the daily mean never falls below 5 °C, a year-round thermal growing season at a season mean of 14.3 °C. A hectare of that soil 5 cm deep weighs 700,000 kg at the NRCS ideal bulk density and therefore holds 191 kg of lead, of which 135 kg must leave the site. Each species was then run twice: once unamended, and once with every lever the sources quantify stacked together — nitrogen at 380 mg N/kg, which drops season-mean pH to 6.08 and multiplies uptake by 1.90; the 1.25-fold biomass gain [R13] measured at that same rate; and irrigation meeting the FAO-56 crop water deficit in full.

‡ Harvest yield is a user input — no consulted source publishes field dry-matter yields for these species. Sunflower is set at 9 t/ha per harvest and the other two scaled from it by the [R13] Table 2 pot shoot dry matter (11.0 / 10.0 / 9.8 g per pot); every timeframe moves inversely with that assumption. † Amaranth's cost is lower only because FAO-56 publishes no crop coefficients for Amaranthus, so its irrigation cannot be computed and the largest single line is absent — it is not cheaper, it is less known. § Costs use rates supplied by the author, not by any source: $85/ha per planting for seed, $1.20/kg for nitrogen, $1.50/m³ for irrigation water, $100/t for biomass disposal. ¶ The irrigation and cost rows were computed against an earlier ten-year ERA5 window. That window rolls forward, so they no longer reproduce exactly — re-running the tool today gives sunflower about 3,125 m³/ha/yr against the 3,480 shown. Every other row in this table was recomputed on 2026-08-29 and is current.

The table makes three points that the headline figures alone would hide. First, amaranth ties sunflower — 113 years against 112 — despite a bioconcentration factor 27% lower, because its 56-day cycle banks six harvests a year against sunflower's four. That is the concentration-versus-tonnage argument from the previous section appearing in measured data rather than in a hypothetical, and it is the reason the model multiplies BCF by yield by cycles rather than ranking on BCF alone. Second, the site-match scores are not comparable across the row: sunflower's 0.913 averages four scored factors, both 1.000s average two. Oakland's 14.3 °C growing-season mean is below the published optimum for all three species, but only sunflower has a published lower tolerance (6 °C) that lets the model score it — amaranth and mustard have none, so the factor is excluded rather than guessed, and their perfect scores mean "nothing contradicted them", not "ideally suited". Third, the two bottom rows move the answer further than any choice of species: tilling to 15 cm triples every timeframe, and capping the schedule at the two harvests a year the tool considers defensible doubles or triples it again. Between them, the treatment depth and the harvest count span 47 to 390 years across a table whose species differ by at most 16%.

Results, outcome and limitations. Sunflower ranks first on every measure the model can defend, but the margin is small and the absolute answer is discouraging: a West Oakland yard at the neighbourhood median takes 112 years unamended, and stacking every sourced lever — nitrogen at 380 mg N/kg, the biomass gain that comes with it, and full irrigation — brings that to 47, a 2.4-fold gain that is still longer than a working lifetime, at roughly $11,300 per hectare per year ($340 seed, $1,277 nitrogen, $5,220 irrigation, $4,500 disposal) and about $533,000 across the project. The outcome across the rest of [MC]'s range is sharper than the median suggests: at the citywide median of 64 mg/kg the site is already below target and needs no remediation at all, which is McClintock's own headline surviving the model intact; at the worst gridded yard (1,023 mg/kg) the projection is 234 years, and at the West Oakland maximum of 3,329 mg/kg it is 342 years while extrapolating 83× beyond the soil concentration at which every BCF here was measured — arithmetic, not evidence, and a capping-and-removal problem rather than a planting one. The limitations are correspondingly heavy. The four-to-six harvests a year assume continuous cropping with no fallow, no rotation and no yield decline, which nothing here supports. Harvest yield is assumed, not measured, and every timeframe scales inversely with it. The nitrogen rate is agronomically extreme — 266 kg N/ha per application at 5 cm, over 1,000 kg/ha/yr, roughly five times a heavily fertilised maize crop — with real nitrate-leaching risk, and it triples if the plot is tilled. Every price above is mine rather than a source's, and the three costs most likely to dominate are missing entirely: labour for cutting and baling 45–61 t of dry matter a hectare a year for five decades, whether lead-bearing biomass is even accepted as green waste rather than hazardous waste, and soil testing — the one expense that genuinely changes the answer, since measuring the depth of contamination costs a few hundred dollars and moves the result threefold, more leverage than every amendment priced here combined. The one route to a human timescale is a chelate: citric acid at [PCT]'s sevenfold multiplier gives sunflower 766 mg/kg shoot tissue and six years to target, but that figure was measured on vetiver, not sunflower, and mobilising lead sevenfold in a residential yard moves it toward groundwater.

Feedback & contact

Found a wrong number, a missing source, or a species worth adding? Write in. Corrections that cite a source get made.

Email feedback Saketh.Balivada@gmail.com

What helps most

• A correction — the number you think is wrong, and the source that says otherwise. Every coefficient here is traceable, so a citation can change one.
• A gap — a soil texture, amendment or species the tool should cover but does not.
• A bug — what you entered, what you expected, what you got.
• Field trials and experiments — results from a plot, a pot trial or a classroom experiment, whether or not they agree with the tool. Measured harvest yields are the model's biggest missing input; observed shoot lead, biomass and cycle lengths all help too.

Every number this tool produces is traced to a named source in the Method tab — equations, coefficients and the limits of each. Check there first; it likely already answers the question.