Beacon Rev A: Design Review Before Fab
A review is not a meeting where the designer defends the board. It is a meeting where the board defends itself.
🎯 By the end you'll understand
How to run a pre-fabrication review that produces decisions rather than agreement: who does what, which artefacts have to exist beforehand, and the rules that stop it degenerating into a walkthrough. How to trace every requirement, R1 through R9, to the specific design feature that delivers it and the evidence that it does. How to record a finding as fix, accept-with-mitigation or defer, and why "accept" is a legitimate engineering decision when it is written down and a defect when it is not. And what Beacon Rev A actually ships with — including one defect the review flags and the team knowingly accepts, one it can only convert into a bring-up measurement, and one nobody catches at all.
KEY INSIGHTThe review's product is a decision list, not a feeling of confidence
Everyone agreed the board looked good. Nobody had opened the layout. Two of the three defects that shipped were visible in it at normal zoom.
Design reviews fail in a specific and recognisable way. The designer shares a screen, drives the mouse, and narrates the design. The reviewers, who received the artefacts twenty minutes earlier, ask about the parts they happen to know. Everything the designer is confident about is skipped, because a person narrating their own work naturally moves quickly through the areas they are sure of — and those are exactly the areas where a wrong assumption has had the longest time to solidify. Ninety minutes later the meeting ends with "looks good to me", which is a mood, and the board goes to fab.
A useful review inverts almost every part of that. The artefacts are circulated in advance and read alone. The designer answers rather than presents. Someone who has not seen the board drives the checklist, item by item, so coverage does not depend on anyone's interest. Every observation becomes a numbered finding with an owner and a written decision. And at the end there is no verdict on the board's quality — there is a list, and each item on it says fix, accept, or defer, with a reason attached.
That last part is what makes a review honest about shipping with known defects. Rev A goes to fab with three of them. Two are engineering decisions made deliberately, in writing, with mitigations. One is not a decision at all, because nobody saw it. The difference between those two categories is the entire subject of this lesson.
A finding you accept with a written mitigation is a managed risk. The identical finding, unrecorded, is a defect. The board is the same board. What differs is whether the next person — the one doing bring-up, or writing the manual, or diagnosing a field return — knows what they are looking at.
HOW IT WORKSHow to run the review
Three roles, and one that surprises people
- The reviewer drives. Someone who did not design the board holds the checklist, opens the files, and moves through the domains in order. This is the counterintuitive one and it is the single highest-value change you can make: coverage stops depending on the designer's sense of what is interesting.
- The designer answers. They explain intent when asked and do not defend. Their job in the room is to be the fastest available source of "why", not to be persuasive.
- A scribe records. Every finding gets an identifier, a one-line statement, an owner, and a decision. Nothing is closed verbally. If the scribe cannot write it down, it was not a finding, it was a conversation.
Artefacts, circulated 48 hours ahead
| Artefact | What it lets a reviewer do alone |
|---|---|
| Schematic PDF, all sheets | Read every net, check every strapping pin and pull-up without a live tool |
| Layout PDF, one page per layer, plus an assembly drawing | See layer 2's continuity, the antenna keep-out, and what is on the bottom side |
| 3D renders from several angles | Catch connectors facing inwards, collisions, parts overhanging the outline |
| BOM with supplier part numbers and DNP lines | Check part choices, temperature ranges, and cost against R5 |
| DRC report with the exclusion register | Read what was excluded and why — the most information-dense page in the pack |
| The requirements table and the design-intent notes | Trace R1–R9 to features, and see which decisions were deliberate refusals |
If the artefacts do not exist, the review cannot happen yet. That is not bureaucracy; a reviewer who has not read the board alone cannot do anything in the meeting except react to what they are shown.
Domains, in a fixed order
Coverage comes from the list, not from anyone's memory. Beacon's checklist has seven domains, and every one of them is walked whether or not it seems relevant:
Power
Every rail's source, sequencing, and decoupling. Every regulator's enable pin. Every leakage path in sleep, named individually and summed. The current budget compared against the measured numbers from Stage 4.
Digital and strapping
Every strapping pin on the module — on the ESP32-C3 those are GPIO2, GPIO8 and GPIO9 — checked for what it sees at power-up and what can pull it. Every pull-up and pull-down, with its value and its purpose. Every unused pin's state. And a pass over the pin map for pins that do not exist to be used: GPIO11–GPIO17 are bonded to the in-package flash on the MINI-1, so a schematic that assigns one is a schematic that cannot work.
RF
Keep-out on all four layers, ground under the module body, stitching pitch, distance to the battery and to the regulator, enclosure materials, and the link budget as arithmetic.
Sensors and analogue
Thermal isolation, I²C addresses and pull-up sizing, interrupt paths traced from pin to firmware, reference and divider accuracy.
USB, charging, and protection
CC resistors, series resistors, TVS placement, charger programming resistor arithmetic, protection thresholds, connector mechanical anchoring.
Test, debug, and mechanical
Probe access, Tag-Connect clearance in three dimensions, pogo pad geometry, mounting, enclosure fit, part heights.
DFM and outputs
Exclusion register read aloud, mask and silk constraints, edge keep-outs, panelization, and the assembler's rendered preview.
Fix now — change the design before the order. Cheap at this moment and never cheaper again.
Accept with mitigation — ship it deliberately, with a written mitigation: a documentation line, a bring-up check, a firmware workaround, a manual instruction. This is a real engineering decision, not a euphemism, and a design that accepts nothing is a design nobody was honest about.
Defer to the next revision — record it on the Rev B change list and do nothing now.
The failure mode is using only one of them. A team that fixes everything never ships; a team that defers everything ships a list of surprises. What makes the middle option legitimate is the word written.
CORE CONCEPTTracing R1 to R9
The requirements were fixed in Lesson 1.9 and everything since has been an attempt to satisfy them. The trace is the part of the review that asks, for each one: which feature delivers this, what evidence says so, and does it hold. It takes twenty minutes and it is the only mechanism that catches a requirement quietly abandoned three stages ago.
| # | Requirement | Design feature that delivers it | Evidence in the pack | Verdict |
|---|---|---|---|---|
| R1 | ≥ 6 months at 1 report/min on 1000 mAh | TPS62840 buck (60 nA IQ), deep sleep, divider switched by a PMOS, LED rail switched by a high-side PMOS | Stage 4 budget: 8 µA sleep + 45 ms at 40 mA per minute ⇒ 38 µA average ⇒ ~26,000 h | Conditional — finding P-1 |
| R2 | ≥ 10 m through one plaster wall | Module at the board edge, keep-out on all four layers, cell at the far end, buck ~9 mm away, stitching at 3 mm | Link budget: 22.8 dB of fade margin at 10 m; keep-out zone opened on each layer in turn | Pass |
| R3 | Temp ±0.3 °C, RH ±3%, motion wake ≤ 100 ms | SHT40 19.5 mm from the charger with a thermal moat; LIS2DH12 INT1 on GPIO3; 4.7 kΩ I²C pull-ups on GPIO6/GPIO7 | Thermal reasoning from 5.5; the interrupt path traced from pad to ISR | Conditional — finding S-1 |
| R4 | Report within 2 s of motion | INT1 on GPIO3, inside the RTC domain (GPIO0–GPIO5); FSM WAKE → SAMPLE → ADVERTISE; 45 ms active window | Latency budget from 3.4; the pin checked against the RTC-domain list, because a motion interrupt on a non-RTC pin is an interrupt nobody is awake to hear | Pass |
| R5 | ≤ IDR 150,000 (~USD 9.5) BOM at 25 units | Single-sided assembly, consolidated passive values, module rather than bare chip | Quote structure from 5.8: ~USD 9.40 at 25 units, ~USD 20 at 5 | Pass at 25 |
| R6 | USB-C charging at 500 mA with protection | MCP73831 with RPROG; 5.1 kΩ CC pulldowns; PRTR5V0U2X TVS; 22 Ω series on D±; DW01A + FS8205 | RPROG arithmetic done in the room | Fail — finding C-1 |
| R7 | OTA with rollback | 4 MB module flash, two OTA app partitions, opt-in by holding the button at wake | Partition table from 3.2 | Conditional — finding I-1 |
| R8 | ≤ 35 × 25 mm, four layers | Outline measured at 35.0 × 25.0 mm; JLC04161H-7628 named on the order | Board outline dimension and the stackup entry in the project file | Pass |
| R9 | Indoor, 0–45 °C, IP54 optional | Every part's temperature range tabulated; no battery thermistor | BOM temperature column; the MCP73831 has no battery-temperature sense input, accepted for an indoor 0–45 °C product | Pass, assumption recorded |
Two rows are worth dwelling on. R6 fails, and it fails to a piece of arithmetic anyone could have done at any point in the last three weeks: the MCP73831 sets its charge current with RPROG according to IREG = 1000 V / RPROG, so 500 mA needs 2.0 kΩ and the schematic has 2.2 kΩ — a value that was in the parts drawer. That gives 455 mA. The review catches it because the checklist line says "charger programming resistor: state the equation and evaluate it", not "check the charger".
R9 passes with an assumption recorded, which is a different and underused verdict. The MCP73831 has no battery-temperature sense input, so Beacon charges a lithium cell without measuring its temperature. For an indoor 0–45 °C product that is defensible, and writing the assumption down means that the day someone proposes an outdoor variant, the sentence "we assumed indoor use when we chose a charger with no thermistor input" is already in the file.
One row passes for a reason nobody in the room can take credit for on the day. R1's battery sense sits on GPIO4, and the checklist asks why. The answer is two constraints intersecting: ADC1 on the C3 covers GPIO0 to GPIO4 (GPIO5 belongs to ADC2), and the pin has to be in the RTC domain — which leaves GPIO0 to GPIO4, minus GPIO2, a strapping pin that must read high at reset. Putting the divider on GPIO2 would have been quietly fatal: with the switch off, the lower 1 MΩ pulls the node to ground and latches the wrong boot mode, and even with the switch on, a full cell gives 4.2 V ÷ 2 = 2.1 V against a VIH of 0.75 × 3.3 = 2.48 V. It boots wrong at every battery voltage. Somebody worked that out in Stage 3 and wrote it on the sheet, and the review's job here is only to confirm that the reasoning still holds — which is what a good design-intent note buys you.
SEE ITThe review board
Beacon Rev A's review, as a decision surface. First choose how deep the review goes — this is the choice most teams make implicitly, by running out of time. Then decide each finding you can see. Then freeze the design and find out what Stage 6 makes of it.
Run it once at "schematic only" and freeze. Five findings never appear at all, and the board goes to fab with them — which is what "the review looked at the schematic and everybody was satisfied" produces. Then run it at full depth and fix everything you can see, and notice that the outcome still contains one defect. That defect is not a rule you failed to apply or a distance you failed to measure. It is a four-pad LED that was never on anybody's list of risky parts, and it is the reason the Rev B checklist says "3D viewer on every polarised part, ticked one at a time", with the word every underlined.
BEACONBeacon Rev A's findings, and what the team actually decided
Here is the real record, in the form the scribe produced. Nine findings, seven decisions the team is proud of, and two entries that are the reason Rev A is teaching material.
| ID | Domain | Finding | Decision |
|---|---|---|---|
| C-1 | Charging | RPROG is 2.2 kΩ, giving 455 mA against R6's 500 mA | Fix now. Change to 2.0 kΩ, 1%. Zero cost, half a day, before the order |
| R-1 | RF | Stitching along the pour boundary near the module is at 5 mm, not the 3 mm the λ/20 rule asks for | Fix now. Twenty minutes of via placement |
| S-1 | Sensors | The SHT40's thermal moat is bridged on L3 by the power pour, which re-filled after the keep-out was drawn on L1 only | Fix now. Extend the rule area to all layers, re-pour, re-check |
| M-1 | Mechanical | C12 sits 0.8 mm from the scored panel edge | Fix now. Move it 2 mm inboard |
| F-1 | DFM | Two DRC exclusions carry no written reason | Fix now. One was a deliberate 0.15 mm via in the module fan-out; the other was a genuine 0.09 mm track, corrected |
| T-1 | Test | A 1.2 mm capacitor sits inside the Tag-Connect probe's body clearance — invisible in 2D, obvious in the 3D viewer | Fix now. Found only because the 3D pass covered the debug connector |
| I-1 | Digital | The user button is on GPIO9, a boot-strapping pin: held at power-up it puts the module into serial download mode | Accept with mitigation. See below |
| P-1 | Power | The battery-sense divider's switch has a gate with no defined state in deep sleep | Accept with mitigation. See below |
| — | Assembly | The WS2812B's placement rotation follows KiCad's convention rather than the assembler's | Not found. See below |
I-1 — the GPIO9 button: flagged, argued, accepted
GPIO9 on the ESP32-C3 is a strapping pin, sampled at reset to select boot mode. Beacon's user button pulls it low. Press and hold the button while the unit powers up — which is exactly what a person does when a device seems dead — and the module enters serial download mode instead of running the application. It looks like a brick.
The review flags it in the digital domain, because the checklist line is "every strapping pin: what does it see at power-up, and what can pull it". The argument in the room is short and the conclusion is not obvious:
- Fixing it now means a schematic change, a placement change, a re-route, and a change to the enclosure drawing because the button's position is a mechanical feature. Call it a day and a half, plus the risk that a rushed re-route introduces something new — which is a real cost, not a rhetorical one.
- Accepting it means Rev A ships with a documented behaviour: the bring-up procedure checks it deliberately, the pilot instructions say "do not hold the button while connecting power", and the firmware reports boot cause so a unit that came up in download mode is identifiable in the logs rather than mysterious.
The team accepts. The reasoning that makes it defensible is that Rev A is five boards plus a twenty-five unit pilot, all deployed by people who will be given instructions, and the failure is recoverable by power-cycling without the button held. Rev B moves the button to a non-strapping pin, and the change is on the list from the moment the decision is made rather than discovered later. That is the whole difference between a managed risk and a defect.
P-1 — the divider leak: the finding a review cannot close
The battery-sense divider is two 1 MΩ resistors with a 100 nF capacitor, sensed on GPIO4 and switched by a high-side PMOS whose gate is driven from GPIO1 with a pull-up to VBAT — the same arrangement as the LED rail's switch, which holds itself off through reset and sleep. Drive the gate low and the divider is on. To turn it off, the pin has to be released to high impedance so the pull-up carries the gate all the way to VBAT and VGS reaches zero.
Which is where the trap is. The intuitive way to turn off a switch is to drive its gate to the opposite rail, and a 3.3 V GPIO driving high leaves the gate 0.9 V below a 4.2 V source. VGS = −0.9 V is not off for a low-threshold PMOS; it is subthreshold, partially conducting. The pull-up is simultaneously carrying a standing current from the cell into the pin. Neither path is a fault and neither is visible in a schematic, which is correct as drawn.
The review asks the right question — "what holds every gate at a defined level during deep sleep?" — and gets a plausible answer: GPIO1 is set to its off state before sleeping. What the review cannot do is verify which off state, or whether the pad keeps it through deep sleep at all. That depends on firmware configuration and on pad-hold behaviour, and the difference between a released pin and a driven one is a microamp or two. It is not a drawing; it is a measurement.
So the decision is accept with mitigation, and the mitigation is a number in the bring-up procedure: measure deep-sleep current on the first board, budget 8 µA, escalate above 12 µA. Stage 6 measures 9.8 µA. The extra 1.8 µA raises the average from 38 µA to 39.8 µA, which takes the theoretical life from about 26,000 hours to about 25,100 — from roughly three years to 2.9. R1 asks for six months, so the requirement holds with an enormous margin and the defect is real anyway. It goes on the Rev B list with a specific fix, and the fix is a part, not a setting: a small N-channel FET (or NPN) as a level shifter, with a 4.7 MΩ pull-up from the PMOS gate to VBAT. The GPIO drives the small FET, the small FET pulls the PMOS gate to ground to switch the divider on, and when it releases, the pull-up rests the gate at VBAT so VGS is exactly zero. And there is no firmware stopgap for the boards already built, which is the part worth sitting with. Rev A fits no gate pull-up, so releasing GPIO1 does not park the gate anywhere — it floats, and an indeterminate switch is worse than a measurably leaky one. Driving it high is the only defined state Rev A can express, and that is exactly the −0.9 V that leaks. Fit the pull-up alone and you trade the leak for a pad clamping its own ESD diode. Only the level shifter gives VGS = 0. Draw the level shifter and its pull-up onto the schematic sheet beside the divider, where the next person will meet them.
Some findings cannot be closed by inspection. The correct output is not a guess about whether it is fine; it is a named test, with a budget and an escalation threshold, added to the bring-up procedure before the boards arrive. A review that converts an unanswerable question into a scheduled measurement has done its job better than one that answers it confidently.
The one nobody catches
The WS2812B-2020's footprint has its pin-1 mark on the wrong corner (Lesson 5.3), and the placement file exports the footprint's orientation, so every LED is placed rotated. The review's 3D pass exists — it is what found the Tag-Connect collision, T-1 — but it covered "the parts we flagged as risky": the module, the connectors, the regulator. A four-pad square LED is the definition of a part nobody flags. Nobody opens the assembler's rendered preview either, because the order goes out at the end of a long day and the preview is the last click before Confirm.
Twenty-five boards arrive with LEDs that are dead and slightly warm. It costs a rework pass and a footprint correction, and it goes into the Rev B change list alongside the button and the pull-down. The honest lesson is not "check your footprints", which everyone already believes. It is that the review's checklist said 3D viewer on the parts we flagged, and a checklist that contains the words "the parts we" has a hole in it shaped exactly like the part you did not think about. Rev B's line reads: 3D viewer on every polarised part, ticked individually, plus the assembler's rendered preview before Confirm.
EXAMPLEWhat Rev A costs, and what the review saved
The review takes ninety minutes for three people plus about two hours each of preparation — call it eleven person-hours. The fixes it produced took two working days. Against that:
| Finding | Cost caught here | Cost if it had shipped |
|---|---|---|
| C-1 · charge current 455 mA | One resistor value, half a day | R6 missed on 25 units; a BOM change and a rework pass, or a specification quietly rewritten to match the hardware |
| S-1 · thermal moat bridged on L3 | A rule area and a re-pour, one hour | Humidity error while charging on every unit; weeks of field data to notice and a respin to fix |
| M-1 · capacitor at the scored edge | Move a part, ten minutes | Cracked ceramics appearing as intermittents across the pilot — the most expensive defect class there is, because it is not reproducible |
| T-1 · Tag-Connect blocked | Move a part, ten minutes | No debug access on any Rev A board; every bring-up problem debugged by soldering wires |
| F-1 · undocumented exclusion hiding a 0.09 mm track | Read two lines, five minutes | Intermittent open on some boards, found warm, blamed on everything else first |
The comparison is not there to prove reviews are worthwhile, which is unfalsifiable in the abstract. It is there because the numbers are the argument you will need the next time someone proposes skipping one to save an afternoon. Lesson 5.1's cost-of-change ladder says the same thing in general; this is the specific instance, on your own board, with names attached.
The Rev A order goes out: five boards, four layers, ENIG, assembled top side only, at roughly USD 20 each against a target of USD 9.5 at twenty-five. Stage 6 opens the box.
WEAKNESSWhere reviews bite
The most common failure and the one this lesson's widget is built around. Schematics are comfortable: they are readable, they fit on a screen, and a room full of engineers can discuss them fluently. Layouts are not, so they get a quick scroll and a nod. But by Stage 5 most of the remaining risk is geometric — thermal distances, keep-outs, edge clearances, probe access, panel edges — and none of it is visible in a schematic. If a review does not have the layout open, on a projector, layer by layer, it is a schematic review being described as a design review.
The person who drew the board navigates it the way they think about it, which means they move fastest through the regions they are most certain about — and certainty that has never been checked is exactly the thing a review exists to find. Give the mouse and the checklist to somebody else. It feels inefficient for the first ten minutes and then it starts finding things.
A review held after the boards are ordered is a post-mortem with better lighting: every finding it produces is now a change order or a defect. And a review with no written findings produces nothing that survives the meeting — the accepted risks become forgotten risks, the deferred items never reach a Rev B list, and six months later somebody asks why the button is on a strapping pin and nobody remembers that it was a decision. The scribe is not administrative overhead. The scribe is the deliverable.
KEY INSIGHTHow this connects
- Beacon's Requirements: From Wish to Specification (1.9) — R1 to R9 were written to be traceable, and this is the lesson where that pays.
- Schematic Capture Done Right (5.2) — the schematic domain of the checklist, and the design-intent notes that let a reviewer read intent instead of guessing it.
- Placement: The Decision That Determines Everything (5.5) — the thermal and mechanical findings are placement findings, which is why the checklist names distances in millimetres.
- RF and Antenna Layout for the ESP32 Module (5.7) — the RF domain: keep-out on every layer, stitching pitch, and the link budget as arithmetic rather than assertion.
- DRC, DFM, and Getting a Board Manufactured (5.8) — the exclusion register and the rendered preview: the first is the review's densest page, the second is the gate Rev A skipped.
- Board Bring-Up: Smoke Test to First Blink (6.1) — five boards, a written defect list, and a bring-up procedure containing the measurements this review could not make.
Stage 5 is finished. Beacon Rev A exists as a reviewed design with a named stackup, a keep-out that DRC enforces, a routing plan with reasons attached, five assembled boards on order, and a defect list that is honest about two entries and silent about a third. Stage 6 opens the box and starts measuring.
GLOSSARYKey terms introduced
- Requirement trace
- A row-by-row mapping from each numbered requirement to the design feature that delivers it and the evidence that it does. The mechanism that catches a requirement quietly abandoned.
- Finding
- A numbered, written observation from a review, with an owner and one of three decisions. Anything not written down was a conversation, not a finding.
- Accept with mitigation
- A decision to ship a known issue together with a specific written defence — a documentation line, a bring-up check, a firmware workaround. Legitimate when written; a defect when not.
- Exclusion register
- The list of DRC violations deliberately excluded, each with a reason and an approver. Read aloud in review because it is where deliberate rule-breaking hides.
- Strapping pin
- A pin sampled at reset to configure boot behaviour. Anything that can pull it during power-up is part of the boot design whether it was meant to be or not.
- Rev B change list
- The running list of accepted and deferred items, opened during the review rather than reconstructed afterwards from memory.
- Escalation threshold
- The number attached to a bring-up measurement that says when a result stops being acceptable — the thing that turns "measure sleep current" into a test that can fail.
- Design freeze
- The declared point after which changes require a new revision. Placing the order without declaring it is the same event with no record.
ACTIVE RECALLCheck your understanding
Answer in your own words first, then reveal.
1Beacon ships with the button on a strapping pin. Explain why that is a managed risk rather than a defect, and what would have made it a defect.›
Because the review found it, stated it, argued it, and recorded a decision with a mitigation. GPIO9 is sampled at reset to select boot mode, so a unit powered up with the button held enters serial download mode and looks bricked. Fixing it before fab would have cost a schematic change, a re-route and an enclosure drawing change — about a day and a half, plus the risk a rushed re-route introduces something new. The team accepted, with three written mitigations: the bring-up procedure exercises the behaviour deliberately, the pilot instructions say not to hold the button while connecting power, and the firmware reports boot cause so an affected unit is identifiable in the logs. Rev B moves the button, and that entry exists from the moment of the decision. The identical board with none of that written down is a defect, because the next person to meet the behaviour has no way to know it was ever considered.
2The review could not determine whether the battery-divider switch leaks. What is the correct output of a review for a finding it cannot close, and what did that produce here?›
A named measurement with a budget and an escalation threshold, added to the bring-up procedure before the boards arrive — not a confident guess either way. The divider's PMOS gate is driven from GPIO1 against a pull-up to VBAT, and whether the pin is released to high impedance in deep sleep or driven to 3.3 V depends on firmware configuration and pad-hold behaviour. Released, VGS is zero and the switch is off; driven, VGS is −0.9 V against a 4.2 V source, which is subthreshold rather than off, and the pull-up carries standing current as well. The schematic is identical in both cases. So the review recorded: measure deep-sleep current on the first board, budget 8 µA, escalate above 12 µA. Stage 6 measured 9.8 µA, identified the partially-on switch, and put the pad-hold configuration on the Rev B list. The extra 1.8 µA raises the average from 38 µA to 39.8 µA, taking theoretical life from about 26,000 to about 25,100 hours — R1 asks for six months, so the requirement holds comfortably and the defect is still real.
3Why does the reviewer, rather than the designer, drive the checklist — and why does it matter more for the layout than for the schematic?›
Because a person narrating their own work moves fastest through the parts they are most certain about, and unchecked certainty is precisely what a review exists to surface. Handing the mouse and the checklist to someone who has read the pack alone makes coverage a property of the list instead of a property of anyone's interest. It matters more for the layout because a schematic is legible to a room — people will genuinely read it and ask questions — whereas a layout is dense and unfamiliar, so a designer-led walkthrough becomes a scroll and a nod. By Stage 5 most remaining risk is geometric: thermal distances, keep-outs on inner layers, edge clearances, probe access in three dimensions. None of it is visible in the schematic and none of it gets discussed unless a list forces it.
4Beacon's review ran a 3D pass and still missed the rotated LED. Diagnose the process defect and write the corrected checklist line.›
The checklist line said "3D viewer on the parts we flagged as risky", and the parts anyone flags are the module, the connectors and the regulator. A four-pad square LED is by construction the part nobody flags: it is cheap, it is simple, and its footprint gives no geometric hint about which corner is pin 1. The defect is not in anyone's attention, it is in a checklist containing a self-referential filter — any line of the form "check the parts we consider important" can only find what you already suspect. The corrected line removes the filter and adds the free gate that was skipped: "3D viewer on every polarised part, ticked individually against the datasheet's top view; then open the assembler's rendered preview and inspect every polarised part again before pressing Confirm." Both halves matter — the second one is the only artefact anywhere in the flow that shows your data after the assembler's rotation conventions have been applied.