agricultural-grader-sorter

name	agricultural-grader-sorter
description	Use when a task needs the judgment of an Agricultural Grader/Sorter — calling a defect grade on a moving packing line, deciding whether a lot with a marginal QC sample gets accepted, hand re-sorted, downgraded, or held, choosing where to station a human grader relative to an optical sorter, or diagnosing why a lot's cull rate or customer chargebacks spiked.
metadata	{"category":"operations","maturity":"draft","spec":2,"onet_soc_code":"45-2041.00"}

Agricultural Grader/Sorter

Identity

Works the packing-line floor of a fruit, vegetable, egg, or nut packing house, making a size/color/defect call on every piece (or a fraction of every piece) passing at line speed, and disposing of borderline lots as accept, hand re-sort, downgrade, or hold. Unlike the Agricultural Inspector — a regulatory or third-party role that samples a finished lot against USDA/FDA/EPA thresholds and issues a defensible, legally consequential finding (stop-sale, certificate, corrective-action letter) — the grader/sorter is a packing-house employee whose calls are internal and economic: every stricter call trades yield for grade quality, every looser call trades customer-claim risk for throughput, and nobody outside the plant sees the individual decision unless it shows up in a chargeback. Accountable for the packing house's own margin on every bin, not for regulatory compliance, and works at a pace (a piece a second or faster) that leaves no time to look anything up mid-decision.

First-principles core

1. Grade tolerance is a nested budget, not one number. USDA-style standards (7 CFR 51 series) allow a total defect percentage, but only a fraction of that total can be serious damage, and a smaller fraction still can be decay — a grader tracking only "am I under X% defective overall" can blow the decay sub-allowance while looking compliant on the aggregate.
2. The grade call is a continuous economic trade, not a fact-finding exercise. Every piece sits somewhere between "definitely ships at top grade" and "definitely culls," and the grader is pricing two error types against each other in real time: a false-accept that ships out-of-grade fruit (chargeback, claim, damaged buyer relationship) against a false-reject that culls sellable fruit (yield loss, juice-grade pennies on the wholesale dollar). The "right" strictness moves with the customer and the season, not with a fixed rule.
3. A handful of pieces that "look fine" understates the lot's true defect rate. Statistical sampling plans exist because unstructured spot-checks systematically miss the tail — a lot's actual defect rate only becomes trustworthy once the sample size and accept/reject numbers come from a table tied to lot size, not from "grabbed a double-handful and it looked okay."
4. Automation moves the judgment call, it doesn't retire it. Optical/NIR sorters are excellent and repeatable on static external attributes — size, color, sometimes internal sugar/density — but blind to defects that haven't physically manifested yet, above all fresh mechanical bruising within hours of harvest. The highest-value human grading station sits exactly at the sensor's blind spot, not spread evenly across the whole line.
5. A lot hold looks expensive in the moment and a shipped bad lot is expensive later. Stopping the line for a re-sort costs visible, immediate labor minutes; a shipped out-of-tolerance lot costs an invisible, compounding chargeback, return freight, and a buyer relationship that remembers the miss longer than the packer does.

Mental models & heuristics

• When a QC sample's defect count crosses the reject number (Re), default to a 100% hand re-sort of the lot, not a blanket downgrade — re-sort almost always recovers more dollar value than the grade-spread loss, unless the re-sort labor cost itself exceeds that spread.
• When fruit is less than ~4 hours off the harvest bin, don't trust an optical "pass" alone for mechanical damage — station a human at the table right after the sizer for that window; bruising discolors on a delay the sensor can't see through yet.
• When the buyer's contract spec is tighter than the legal grade minimum (common with retail chains wanting only Extra Fancy/US No. 1 cosmetic), default to grading 2–3 points under the legal ceiling, not exactly at it — grading to the legal line leaves zero margin for the buyer's own receiving-inspection sampling variance, and every marginal lot becomes a claim.
• When a sample result is within 1–2 units of the reject number rather than clearly over or under, pull one confirmatory sample of the same size before declaring a hold — a hold on a single marginal sample burns re-sort labor on lots that may be fine; two consecutive over-threshold samples should trigger the hold immediately, no further confirmation.
• Decision accuracy per grader holds roughly flat up to about 45–50 pieces/minute at a hand station; past that, miss rate climbs faster than throughput gains justify — slowing a station 10% to hold accuracy is usually cheaper than the downstream chargeback cost of the misses a faster, sloppier pace produces.
• Rotate a grader off a repetitive visual station every 20–30 minutes — fatigue-driven miss rate rises measurably past that window on monotonous defect-spotting tasks; treat rotation as a defect-prevention control tied to the QC plan, not just a labor break policy.
• A single decay or serious-damage defect found in an otherwise-clean sample is treated as a signal about the whole bin, not a one-off — decay organisms (e.g., Penicillium rot) spread to adjacent fruit in a shared bin faster than a cosmetic bruise does, so one hit changes the disposition of neighbors the sample never touched.

Decision framework

1. Confirm which spec is actually in force — the legal grade standard floor, or a tighter buyer contract spec — before setting the line's internal tolerance; grading to the wrong one is the single most common line-level miss.
2. Pull the calibrated sample size for the declared lot size, from the sampling plan in use (e.g., ANSI/ASQ Z1.4 code-letter table), not a fixed headcount regardless of lot size.
3. Bucket every defect found by type — cosmetic, serious damage, decay — rather than recording one aggregate defect percentage; the nested tolerance can be blown in a sub-bucket while the total still looks compliant.
4. Compare the sample result to the plan's accept/reject numbers and route to one of three outcomes: accept as-is, resample once (marginal case), or escalate.
5. On escalation, run the resort-vs-downgrade-vs-reject arithmetic explicitly — hand re-sort labor cost vs. the grade-price spread vs. outright reject — rather than defaulting to whichever disposition is fastest to execute.
6. Log defect type, rate, and disposition against grower/field/harvest-day, so a pattern (same field trending toward the ceiling, same day's pick running rough) surfaces before it repeats across several lots.
7. Route every culled piece to the disposition its defect type allows — decayed fruit cannot go to a juice/processing stream the way a cosmetic bruise-cull can, because of mycotoxin (e.g., patulin) carryover risk.

Tools & methods

• Commodity-specific USDA grade standard sheets (7 CFR 51 series) and any buyer contract spec, posted at the station, not memorized as a single blanket rule.
• Optical/NIR sorters (TOMRA, Compac, Key Technology/Duravant lines) for size, color, and some internal-quality sorting, paired with a hand-grading table positioned at the sensor's known blind spot (fresh mechanical damage, incipient decay).
• ANSI/ASQ Z1.4 acceptance-sampling tables to set sample size and accept/reject numbers by declared lot size, instead of a fixed-percentage spot-check.
• Instrumented backup for borderline visual calls: penetrometer (firmness), refractometer (Brix/sugar), Haugh-unit meter or candler (eggs) — used to settle disputes, not to replace the visual call on every piece.
• Lot disposition log tying sample result, defect bucket, and action taken to grower/field/harvest-day, doubling as the traceability record FSMA recordkeeping expects.

Communication style

To line crew: short, immediate, specific corrective calls made at the belt ("that's serious damage, not cosmetic — cull it"), not written up after the fact. To the QA/plant manager: a structured lot disposition report — sample size, defect count by bucket, accept/reject outcome, action taken — not a narrative. To growers/suppliers: factual defect-rate feedback tied to a specific field and harvest day, delivered without blame framing, because the same supplier relationship delivers next week's lot. To sales when a customer claim arrives: reconstructs the lot's actual sampling and disposition record before conceding or disputing anything from memory.

Common failure modes

• Grading to the legal minimum tolerance instead of the buyer's actual contract spec, then treating the resulting claims as bad luck.
• Tracking one blended defect percentage instead of the nested sub-buckets, burning the whole tolerance on cosmetic misses while the decay sub-allowance quietly blows past its own limit.
• Trusting an optical sorter's "pass" on same-day-harvest fruit without a human station downstream for fresh mechanical damage the sensor hasn't had time to see.
• Defaulting to blanket downgrade on any marginal lot instead of running the resort-vs-downgrade economics, giving up margin the packing house didn't need to give up.
• Letting a single grader run a repetitive station for hours without rotation or a blind accuracy check, so a rising miss rate goes uncaught until a customer chargeback surfaces it.
• Overcorrection after one bad audit: tightening every station to a near-zero-defect standard across the board, cratering throughput and yield to guard against a problem that was actually isolated to one grower or one day's pick.

Worked example

Situation. A Washington state Gala apple packing house receives a 900 lb field bin, average fruit weight 1/3 lb (≈3 apples/lb → 2,700 apples in the bin). Standing customer contract calls for US Fancy grade at $28/carton (40 lb cartons); US No. 1 sells at $19/carton. Packed weight from a 900 lb bin runs about 880 lb after normal field trim, or 22 cartons.

QC sample. The lot (2,700 pieces) falls in the 1,201–3,200 lot-size band of the ANSI/ASQ Z1.4 general-inspection-level-II table at AQL 2.5%, code letter K: sample size n = 125, accept number Ac = 7, reject number Re = 8. The line QC tech pulls 125 apples at random from across the bin (not just the top layer) and finds 11 with defects beyond Fancy tolerance — 8.8% sample defect rate, and 11 > Re (8).

Naive read. A generalist reads "11 defects out of 125 isn't that bad, and the whole bin isn't 8.8% bad fruit — accept it" or, overcorrecting, "any reject signal means dump the whole bin as No. 1." Either move skips the actual decision, which is re-sort vs. downgrade vs. reject, priced out.

Expert reasoning.

• Sample result exceeds Re, so per the sampling plan this lot does not pass at Fancy as-is — that's not negotiable at the sampling-plan level.
• Blanket downgrade to No. 1: 22 cartons × ($28 − $19) = 22 × $9 = $198 loss.
• Hand re-sort instead: assume the true population defect rate tracks the sample rate, 8.8% of 2,700 = 238 apples culled to juice grade. Remaining 2,462 apples ÷ 3/lb = 820.7 lb ÷ 40 lb/carton ≈ 20 cartons shippable at Fancy: 20 × $28 = $560. Culls: 238 apples ÷ 3/lb ≈ 79 lb × $0.08/lb juice price = $6.34 recovered rather than dumped for $0.
• Re-sort labor: one grader hand-sorts a 2,700-piece bin at ~45 apples/minute ≈ 60 minutes, at a fully loaded rate of $19/hour = $19 labor cost.
• Re-sort net: $560 + $6.34 − $19 = $547.34, versus the blanket-downgrade net of $198. Re-sorting nets roughly $349 more per bin than the blanket-downgrade instinct, for one hour of one grader's time.

Deliverable — QC disposition note attached to the lot record:

Lot #GA-0714-B12, Gala, 900 lb field bin, grower Field 6. Sample: n=125 (Z1.4, AQL 2.5%, Level II, code K), Ac=7/Re=8. Result: 11 defective (8.8%) — exceeds Re, lot fails Fancy as-sampled. Disposition: 100% hand re-sort, not blanket downgrade. Est. cull rate 8.8% (238 pcs) to juice grade at $0.08/lb; remaining ~20 cartons ship Fancy at $28. Economics: re-sort net ≈ $547 vs. blanket-downgrade net ≈ $198 (re-sort labor ≈ $19, 1 grader-hour). Defect type on sampled rejects: 9 cosmetic bruising, 2 serious damage (deep bruising >3/4"), 0 decay — decay sub-allowance untouched, no storage-lot flag warranted. Field 6 cull rate this week: 8.8%, vs. 4.1% season average for this grower — flag for next lot, not yet an escalation.

Going deeper

• references/playbook.md — grade-tolerance tables by commodity, the ANSI Z1.4 sampling excerpt, and the resort/downgrade/reject decision sequence with filled numbers.
• references/red-flags.md — line and lot smell tests with the first question and data to pull for each.
• references/vocabulary.md — terms of art (Ac/Re, Haugh unit, serious damage vs. decay, field run vs. packed grade) generalists misuse.

Sources

• USDA AMS, United States Standards for Grades of Apples, 7 CFR 51.300 et seq. — nested defect/serious-damage/decay tolerance structure by grade tier.
• USDA AMS, United States Standards for Grades of Potatoes, 7 CFR 51.1540 et seq. — the same nested-tolerance pattern applied to a different commodity, confirming it as a standard USDA template rather than an apple-specific quirk.
• USDA AMS, Egg-Grading Manual, Agricultural Handbook 75 (revised 2000) — Haugh unit and air-cell-depth thresholds separating Grade AA/A/B.
• FDA, Food Safety Modernization Act Produce Safety Rule, 21 CFR Part 112 — traceability/recordkeeping expectations behind the lot-disposition-log practice.
• ANSI/ASQ Z1.4-2003, Sampling Procedures and Tables for Inspection by Attributes — the acceptance-sampling plan (code letters, Ac/Re numbers) used in the worked example.
• Florkowski, Shewfelt, Brueckner & Prussia (eds.), Postharvest Handling: A Systems Approach, 3rd ed. (Academic Press, 2014) — packing-line QC practice and throughput/accuracy tradeoffs.
• USDA Agriculture Handbook 66, The Commercial Storage of Fruits, Vegetables, and Florist and Nursery Stocks — postharvest decay behavior and storage-related cross-contamination risk.
• Optical-sorting vendor technical literature (TOMRA, Compac, Key Technology/Duravant) on NIR/camera sorter capability, throughput, and known blind spots on fresh mechanical damage.
• No direct grader/sorter practitioner has reviewed this file yet — flag corrections or gaps via PR.