02 — Livestock nutrition, energy and herd review of cleaned v0.7.0

Reviewer role: ruminant nutrition / livestock systems (adversarial). Source: CIAT/cleaned @ 800d53a (read-only). How I checked it: I sourced R/feed_quality.R and R/energy_requirement.R into Rscript and ran them on inst/extdata/example_input.json. The example needed two patches before it would run (both are findings: C-01 and M-01). I reproduced the livestock-N block of R/ghg_emission.R by hand and compared it with IPCC 2006/2019 Tier 2. Scripts are in the session scratchpad (run2.R, run3.R). I did not change the source.

Verdict

The feed-demand engine has a basic unit error. It adds up IPCC net energy requirements (NEm, NEa, NEg, NEl, NEp) and divides them by the feed's metabolisable energy density. There is no km/kl/kg or REM/REG conversion anywhere. This makes energy-driven DMI about 35–50% too low for cattle, and too low by a larger margin for growth and for small ruminants. The error carries through to land requirement, enteric CH4, VS and N excretion. In the example the bias is hidden because DMI is taken as max(DMI_energy, DMI_CP) with no intake cap. For low-CP tropical diets this gives physically impossible intakes: 4.4% of body weight for a 600 kg cow on a 5.3 MJ ME/kg DM ration. In the IPCC N-excretion block, retention is in kg N/day but the code uses it as a fraction. For dairy cows it is also divided by 365 twice. So N excretion comes out 30–45% too high and N retention is effectively zero. Neither shipped example dataset runs end-to-end: the column names, the DM units and the livestock labels don't match the code. Herd dynamics (mortality, replacement, offtake) are not modelled at all. As it stands, I would not trust absolute DMI, land, CH4 or N outputs from this model. Scenario differences inherit the same biases, but less severely.

Findings table

Counts: Critical 2, High 5, Medium 9, Low 8.

Detailed findings

C-01 NE requirements met with ME (no efficiency conversion). Critical, CONFIRMED

• Evidence. energy_parameters.json "Table 10.4" gives the IPCC NEm Cfi (0.322/0.386/0.370/0.217/0.236/0.315). L79 computes er_maintenance = Cfi*BW^0.75, which is NEm. L93-101 compute NEg (IPCC 10.6/10.7), L106-111 NEl (10.8), L117 NEp and L122 NEwork. L140 sums them and L170 does fresh_intake_required_e = energy_required_by_season/average_me, where average_me is ME from me_content (MJ ME/kg DM). I found no km, kl, kg, REM or REG anywhere in R/.
• Expected. IPCC 2006/2019 Eq 10.16: GE = [(NEm+NEa+NEl+NEwork+NEp)/REM + NEg/REG]/(DE%/100). Or, in an ME system (AFRC 1993/CSIRO), MEm = NEm/km (≈0.70), MEl = NEl/kl (≈0.62), MEg = NEg/kg (≈0.3–0.45).
• Magnitude. I ran the example (run3.R). Code DMI as a share of IPCC DMI: cow 0.58/0.61/0.63, heifers 0.51/0.55/0.58, calves 0.49/0.53/0.56, at DE = 55/60/65% respectively. The DB's own ME system (lkp_livetype) gives about 119 MJ ME/d for the example cow; the code uses 82.8. For small ruminants: DB ewe me_maintenance is 8.64 MJ ME/d, while the code gives 0.217·53^0.75 = 3.76 (−56%).
• Impact. Energy-limited DMI is about 35–50% too low, which drives land requirement, enteric CH4 (GE = DMI·18.45), VS and N intake down. Pigs are internally consistent (Cfi 0.44 is ME, NRC 1998), so ruminant and pig results are on different bases.
• Fix. Implement IPCC Eq 10.14–10.16 (REM/REG from diet DE) and derive GE→DMI. Alternatively, convert each NE term with AFRC km/kl/kg from qm = ME/GE. Then use the same DMI for land and GHG.

C-02 IPCC N retention used as a fraction and double-annualised. Critical, CONFIRMED

• Evidence. ghg_emission.R:226: n_retained = ifelse(!Pigs, ifelse(annual_growth==0, milkN/6.38/no_days, milkN+growthN)/no_days, …). The /no_days outside the ifelse applies again to the dairy branch, so dairy retention is divided by 365². Then L233 does n_excretion_rate = n_intake*(1-n_retained)*no_days.
• Expected. IPCC 2019 Eq 10.31 is Nex = (N_intake − N_retention)·365, with both terms in kg N/animal/day. The 2006 form uses N_intake·(1 − N_retention_frac) with a fraction from Table 10.20.
• Magnitude (example). Nex code vs correct: cow 89.4 vs 68.2 kg N/yr (+31%); heifers (herd) 49.7 vs 43.3 (+15%); calf 12.8 vs 8.8 (+46%). Cow retention used: 0.00016 against the correct 0.058 kg N/d. The overstatement propagates to direct and indirect manure N2O, and to total_n_from_manure_mgmt.
• Fix. Compute everything per animal per day (milk/365, WG = annual_growth/365), use Nex = N_intake − N_ret, then multiply by herd.

H-01 DMI = max(energy DMI, CP DMI), uncapped. High, CONFIRMED

• Evidence. L174 dmi_s = ifelse(dmi_required_cp>dmi_required_e, dmi_required_cp, dmi_required_e) and L180 me_intake_s = dmi_s*…. Example (5.8% CP, 5.3 MJ ME/kg DM): every season is CP-limited, and the cow DMI is 26.6 kg DM/d (4.4% BW), the calf 4.2% BW.
• Expected. A 600 kg cow on NDF-rich forage can eat roughly 1.2% BW as NDF, which is about 2–2.5% BW as DM (Mertens; NASEM 2021 DMI equations). An animal on a CP-deficient diet eats less and produces less; it does not over-eat to meet CP.
• Impact. On low-CP tropical rations, DMI, ME intake, enteric CH4, manure and land are all inflated. The error also partly and accidentally cancels C-01, so results depend on which constraint happens to bind. Scenario comparisons can flip sign.
• Fix. Base DMI on energy, apply an intake cap (BW or NDF fill), and report CP deficit and energy/CP balance instead of force-feeding. Optionally back-calculate the milk or growth the diet can support.

H-02 Small-ruminant growth and lactation silently collapse with default weights. High, CONFIRMED

• Evidence. L96-101: NEg = annual_growth·(a + 0.5·b·(BW_weaning + BW_year_one))/365. The example and the DB defaults have body_weight_weaning = body_weight_year_one = 0, so NEg = annual_growth·a/365 (e.g. castrates 4.4 MJ/kg, where IPCC gives about 14 MJ/kg at 20→40 kg, and the DB me_growth is 46 MJ ME/kg). L109: if annual_milk == 0, ewe/doe NEl = 5·BW_weaning/365·EV, which is 0.
• Impact. Sheep/goat growth energy is 3–10× too low and suckling-ewe lactation energy is zero unless the user fills fields that are not in lkp_livetype.
• Fix. Validate that these fields are > 0 for sheep and goats, or fall back to the DB ME values. Use WG = BWf − BWi, consistent with eq 10.7.

H-03 Column contract broken between the code and every data source. High, CONFIRMED

• Evidence. Running on the example stops at L106 with object 'fat_milkcontent' not found. The example JSON, lkp_livetype and data/mappings/input_mappings.csv all use fat_content, cp_pregnancy, cp_growth, lw_gain, proportion_growth and n_content. The code and vignettes/input-contract.Rmd use fat_milkcontent, cp_lys_pregnancy, cp_lys_growth, lw_gain_piglets, proportion_growth_piglets_milk and n_manure_content. The same break is in land_productivity.R:60.
• Refutation attempt. I searched R/ for a rename or mapping step and found none apart from the livetype_code rename. The app may rename columns upstream; this is not visible in the package.
• Fix. Add a validation and rename layer at the entry point, and ship a runnable example with a regression test. The only test file is a context() line with no tests.

H-04 Growth N retention unit error. High, CONFIRMED

• 268 − 7.03·er_growth/annual_growth mixes MJ/d with kg/yr. Heifer example: NEg/WG should be 3.86/0.329 = 11.7 MJ/kg, which gives 186 g protein/kg. The code gives about 268 g/kg (+44% protein in gain). This is partly masked by C-02. Fix: WG = annual_growth/365.

H-05 Two inconsistent manure-N estimates. High, CONFIRMED

• energy_requirement.R:203 sets manure DM = DMI·0.365 (fixed, whatever the diet digestibility; example DE is 36%, which implies faecal DM of about 60% of DMI). L207/L211 then set N = manure DM·n_manure_content. nitrogen_balance.R:135-148 uses that value, while GHG uses the IPCC Nex (C-02). In the example, manure_in_stable = 0 with time_in_stable = 1, so the N balance gets 0 collected manure N while GHG N2O is computed on 89 kg N/yr. Fix: derive a single Nex from N intake minus retention, and partition it by time and location for both modules.

M-01 Shipped examples are not runnable, and their units are inconsistent. Medium, CONFIRMED

• feed_quality.R:44 needs crop_name, which the example lacks. Example dm_content values are 0.3, 0.29, 0.15 and 0 (fractions); the DB and contract use % (87, 15, 90). DMI is invariant to DM scaling (it cancels: DMI = E·Σf·DM/Σf·DM·ME), but fresh_intake_required_* comes out 100× too high (about 5.5 t fresh/cow/day in S1). Concentrate DM = 0 zeroes its ME and DM silently, and L132 turns NA/NaN into 0. mufindi.rda livetype labels lack the "Cattle - " prefix, so every Cfi is NA.

M-02 ME requirement data in the DB bypassed. Medium, CONFIRMED

• lkp_livetype has me_maintenance (≈0.50 MJ ME/kg^0.75 for cattle and sheep), me_lactmilk 5.5 MJ/kg, me_growth 50/46/45 MJ/kg and me_pregnancy 1500/170. None are read by the code. Pig values are hard-coded (45 at L92, 171 at L116), which hides where they came from. Either wire this ME system in or drop it; right now the app shows parameters that have no effect.

M-03 Herd vs per-animal mixing in N. Medium, CONFIRMED

• ge_intake and dmi_tot are herd totals (energy × herd_composition at L140); annual_milk and annual_growth in n_retained are per head. The mismatch scales with herd size.

M-04 No herd dynamics; meat from LWG. Medium, CONFIRMED

• herd_composition is a static count. There is no mortality, fertility-driven calf crop, replacement rate or offtake. meat_production_animal = number·annual_growth·carcass_fraction counts replacement heifer growth as meat and gives zero meat from culled cows. For East/West African herds, where mortality is 5–15%/yr and offtake is low, per-kg-product intensities are therefore biased. The direction depends on the herd, but typically too little product is credited to the breeding herd.

M-05 Small-ruminant lactation CP double-counted. Medium, CONFIRMED

• L133 computes cp_lactation (overwriting the input column). L143 adds cp_lactation/BI, and L144 adds annual_milk*cp_lactmilk without making the two exclusive. The energy side (L109) makes them exclusive; the CP side does not.

M-06 Feed library (cleaned.sqlite lkp_feeditem) plausibility. Medium, CONFIRMED

• Brachiaria brizantha CP 7.2548 = ME 7.2548, and the hybrid's CP = ME + 1: a copy error. Oats straw CP 0 (expected about 3–4% DM), which forces CP-limited DMI through H-01. Lupin grain ME 19 and groundnut kernel ME 21.2 MJ/kg DM give de_fraction 1.25 and 1.40; lupin should be about 13–14 (Feedipedia). Maize stover ME 9.13 (codes 48 and 53) against about 6.5–7.5 in the ILRI/Feedipedia range (+25%). Sweet-potato tuber "fresh" DM 59%, where fresh tubers are about 25–35%. I did not audit all 132 rows.

M-07 ILRI and ERA libraries (parent repo data/). Medium, CONFIRMED

• ilri_feed_db.xlsx/ilri_feed_db: 995 rows have ME 16–74 alongside IVDMD < 20, a column swap (e.g. "Food oats 1": ME 73.7, IVDMD 10.6). DM is on a lab-sample basis (median 91%) and can't be used as fresh DM for forages. ERA_nutrition_library.xlsx/Means: CP is in g/kg (max 941 g/kg, which is implausible), DM mixes "% as fed" and "g/kg", and there is Desmodium ME 16.1. These must be converted before they enter feed_items (the contract says % DM). These files are not read by R/, so the risk arises at app or connector level.

M-08 Pig protein and growth basis. Medium, PLAUSIBLE

• DB sow cp_maintenance 0.030 kg/d and grower cp_lys_growth 0.05 kg/kg look like lysine-scale values, yet they are compared against feed CP%. The contract itself calls them "CP/lysine". If they are lysine, the pig CP requirement is about 15× too low. Pig growth at 45 MJ ME/kg gain resembles whole-diet ME per kg gain (which includes maintenance), which would mean maintenance is counted twice. The authors need to confirm the source.

M-09 No seasonal feed balance. Medium, CONFIRMED (design)

• The annual requirement is split by season length (L167), and DMI is whatever the basket needs, so dry-season deficits, weight loss and compensatory gain can't appear. This matters for Sahelian and East African dry seasons, where cattle routinely lose 10–20% of liveweight.

Low items (brief)

• L-01. sum(dmi_required_e, na.omit=TRUE): na.omit is swallowed by ... as the value 1. Verified: sum(c(1,2),na.omit=TRUE) returns 4.
• L-02. IPCC C = 0.8 for females; mixed "steers/heifers" and calves use 1.0.
• L-03. IPCC Table 10.7: sheep and goats 0.077/0.126. The pregnancy term is also applied to any class with birth_interval > 0; the finite guard only catches 0.
• L-04. Sheep Ca values. Unsure of the goat source; please cite.
• L-05. Eq 10.10 WG_wean is the gain from birth to weaning, not the weaning weight.
• L-06. FPCM uses FAO/GLEAM 0.337 + 0.116F + 0.06P, which is correct for cow milk. Protein should be taken from raw milk. The fallback at merge_outputs:457 is per head.
• L-07. AME divisor is 2500 vs 2100.
• L-08. No days normalisation; season_length() omits the end day and is unused.
• Also. All "Cows" get the lactating Cfi of 0.386 all year, including the dry period (slight overestimate). distance_to_pasture is unused.

Done well

• The IPCC Tier 2 equation forms (10.3, 10.4, 10.6 with C = 0.8/1.0/1.2, 10.7 a/b constants for males, castrates, females and goats, 10.8, 10.11, 10.12, 10.13) are transcribed correctly. The error is in how they are combined, not in how they are written.
• The sow lactation ME follows NRC 1998 correctly: (4.92·gain − 90·n)/0.72 → 6.83 and 0.125 Mcal, ×4.2 MJ.
• Pigs use an ME Cfi (0.44 MJ/kg^0.75, matching NRC 106 kcal) that is consistent with ME feed values.
• DMI is invariant to feed-allocation scaling and DM units, because the ratio construction cancels them (only fresh intake is affected).
• The milk "× herd number" change is applied once in land_productivity.R:60 and summed once in merge_outputs.R:502-506. Energy uses per-head annual_milk × herd_composition once (L140). I found no double multiplication.

Open questions for authors

1. Was the NE-divided-by-ME approach deliberate (e.g. inherited from the CLEANED Excel tool), and was it ever benchmarked (auxilary_info/CLEANED_benchmark_RGQ.xlsx)?
2. Is feed_basket.allocation meant to be % as fed or % DM? The code treats it as as-fed (fraction_as_fed). App users usually think in DM.
3. Is annual_milk offtake milk or total yield? Suckling calves' milk intake isn't credited in the calves' energy balance.
4. Where do the pig 45 MJ/kg and 171 MJ values, the goat Ca of 0.019, and the lamb a = 2.3, b = 0.4 come from?
5. Are the pig cp_* fields CP or lysine?
6. Which renaming layer (app side?) maps fat_content → fat_milkcontent and similar, and why does the package ship an example that fails?
7. Should there be a DMI cap and explicit deficit reporting instead of max(E, CP)?