In most plants I have worked in, functional safety is treated as someone else's problem. The safety engineers write the safety case, the software team handles ASIL ratings, and the quality department carries on issuing certificates of conformity. That division is comfortable, and it is wrong. A safety case is not a document that sits in the requirements domain; it is an argument that runs all the way down to the shop floor, to the torque curve of a press-fit machine and the firmware flashing station at end of line.
The moment a safety goal depends on hardware produced in series, the quality system becomes a supplier of evidence. ISO 26262 makes this explicit in Part 7, which deals with production and operation. Yet audit after audit, I find plants where the functional safety audit trail and the production quality audit trail are two separate, non-communicating hierarchies. The safety case claims a certain fault detection rate; the production data that would substantiate it sits in a quality database nobody from the safety side has ever queried.
The interlock points are specific and buildable: what the safety case needs from production, how ASIL decomposition imposes requirements on manufacturing processes, and what evidence the quality system must be architected to produce. This is written from the production side of the fence, which is where the argument usually fails.
What a Safety Case Demands From Manufacturing
A safety case is a structured argument: here are the hazards, here are the safety goals, here is the evidence that the mitigations work, and here is the evidence that they will keep working. The last clause is the one that drags production in. Claiming that an item achieves a particular diagnostic coverage or hardware failure metric is only credible if the manufacturing process is demonstrably capable of producing that item consistently, and if deviations are detected and contained.
Concretely, the quality system must supply three classes of evidence. First, process capability data for every characteristic that carries a safety requirement — not generic capability studies, but evidence tied to the specific safety mechanisms in the item. Second, traceability records that bind each serialised unit to the process parameters under which it was built, so a field incident can be investigated against actual production data rather than nominal assumptions. Third, change control records showing that whenever a process, tool or supplied part changed, someone assessed the impact on the safety case.
The third class is where I see the most failures. A tooling modification on a machining line changes a burr profile; the burr affects a connector insertion force; the insertion force affects a contact resistance that a safety-relevant diagnostic reads. Nobody connected the engineering change order to the safety case. The safety case is now silently out of date, and it stays that way until an auditor or an incident finds it.
The interlock must be procedural, not cultural. No ECO closes without a documented functional safety impact statement, signed by someone with authority over the safety argument. Anything softer than a hard gate in the change process decays within a year, because production pressure always wins against a discretionary review step.

ASIL Decomposition and Production Independence
ASIL decomposition is one of the more elegant and more abused mechanisms in ISO 26262. Where a safety requirement at ASIL D is allocated to two sufficiently independent elements, each may be developed at a lower ASIL — D decomposed into two B elements, or into C and A — provided independence is demonstrated and the original ASIL D requirements are still verified at the integrated level. The concept works. The problem is what happens on the production side.
Independence that exists in design does not survive a shared manufacturing process. If both decomposed elements are produced on the same line, with the same operators, the same handling equipment and the same calibration chain, a single systematic fault — a mis-calibrated tester, a wrong solder paste batch, a contaminated cleaning agent — can defeat both channels simultaneously. Common cause failure does not respect design intent.
Where independence cannot be maintained in production, the decomposition is invalid and the elements must be treated at the original ASIL. The consequences are direct: ASIL D production requirements in Part 7 demand demonstrable control of systematic faults — proven-in-use arguments, validated process parameters, specific end-of-line test coverage — that ASIL B does not. Plants that assumed the lower rating for production purposes and built their process controls accordingly have stored up a serious nonconformity.
Decomposition Independence: Design vs Production
What the rationale document usually covers
- Architectural separation of the two elements
- Software independence and diverse algorithms
- Different sensor or actuator channels
- Independent verification at integration level
What production must also demonstrate
- Separate lines, handling equipment or calibration chains
- Distinct operators or validated automation isolation
- No shared consumables across both channels
- Documented common-cause analysis of manufacturing faults
Proven in Use and the Long Memory Problem
Functional safety leans on systematic faults being managed through process discipline. One mechanism is the proven-in-use argument: a process or component is safe because field history shows no relevant failures. From the quality side, that argument is only as good as the records behind it. A proven-in-use claim requires demonstrating what was actually produced, what unchanged configuration it ran, and what the field return data shows — segmented, not blended into a general warranty pool where safety-relevant and non-relevant failures are indistinguishable.
Field data segmentation is a genuine weakness in most quality systems. Warranty codes are written for commercial purposes, not safety analysis. A return coded as ECU replacement, no fault found tells a safety engineer nothing; the underlying root cause may or may not relate to a safety mechanism. To feed the safety case, the return handling process needs a path that triggers technical analysis rather than scrap-and-replace whenever the part touches a safety function. That path must be defined in advance, because it cannot be reconstructed retroactively when a Part 7 audit or a field investigation demands it.
Change control deserves the same long memory. A safety case written against a process validation campaign of three years ago describes a process that no longer exists if there have been twelve changes since. Each change is individually justifiable; the accumulation is the problem. A living cross-reference between the safety case's process assumptions and the current process documentation must be reviewed at every significant change and at least annually. When the safety case references process step numbers that no longer exist in the control documentation, the decay is well advanced.
Common cause failure does not respect design intent. Production independence must be argued with the same rigour as design independence.
End-of-Line Testing as a Safety Mechanism
End-of-line testing sits in an awkward position. If the safety concept credits end-of-line tests with detecting certain fault modes before delivery — and hardware metrics often implicitly do — then those tests are safety mechanisms and must be validated, controlled and maintained as such. In practice, plants treat EOL testers as ordinary quality equipment: calibrated annually, program changes handled under normal software change rules, test coverage never formally documented against the fault models in the safety analysis.
The interlock requirement is specific. For every safety-relevant test at end of line, there must be a documented link to the fault model it is claimed to detect, a validation that the test actually detects that fault, and control of the tester itself — including its own failure modes. A tester that silently degrades, passing units it should reject, is a common cause failure shared across the entire production output. Tester self-check routines, master-unit verification intervals and tester fault modes belong in the safety analysis explicitly.
Validation must be by fault insertion, not golden-unit correlation. Take known-bad units or inject the fault deliberately — a shorted diagnostic line, an open sensor input, a corrupted firmware image — and prove the station catches it and the containment path works. Repeat it periodically, not once at validation. The frequency should derive from the detection claim's contribution to the safety argument: the more the safety case leans on that test, the tighter the verification interval. Where a plant cannot show fault insertion records for safety-relevant EOL checks, the safety case's production evidence is effectively unverified.
Validating a Safety-Relevant EOL Test
- 01Map fault modelLink each safety-relevant test to the specific fault model it claims to detect.
- 02Insert the faultProve detection with injected faults, not golden-unit correlation.
- 03Verify containmentConfirm failed units are quarantined and routed to technical analysis.
- 04Repeat on intervalFrequency scaled to how much the safety case depends on the test.
- 05Audit the testerSelf-checks and master-unit checks under the tester's own failure-mode control.
Structuring Quality Records as Safety Evidence
Evidence has to be retrievable, attributable and intelligible to someone who was not in the plant when it was created. Quality records in most plants were designed to satisfy product conformity and customer requirements, not safety argumentation. Retention periods align with contractual warranty terms, which may be shorter than the safety-relevant field observation period. Data formats are proprietary or tied to MES systems that will be replaced within a decade.
The practical fix is a jointly owned evidence register: a list, agreed between the safety and quality functions, of every production record the safety case depends on, its retention period, its storage format and its owner. Traceability data for safety-relevant characteristics should be retained for the full safety-relevant lifetime of the fleet, in a format that survives system migrations. Serialisation must be granular enough to bind units to process data — batch-level traceability is inadequate where the safety argument depends on unit-level parameters such as flashed firmware versions or calibration values.
Internal audit is where the interlock gets maintained. At least one audit stream per year should walk the safety case's production claims into the plant and verify them against reality: is the claimed test still running, at the claimed coverage, with the claimed containment? Are the decomposition independence assumptions still true after last year's line consolidation? The safety case will decay without this.
The Quality Function as Custodian of the Argument
Processes change, people move, and assumptions quietly expire. The quality system is the only function positioned to see that happening and to produce the evidence that keeps the safety argument standing. Treating Part 7 as an appendix to the development standard, rather than as a set of requirements on the production system itself, is the root error behind most of the findings I have described.
The corrective sequence is the same in every plant that closes the gap. Inventory what the safety case claims about production, map each claim to the record that substantiates it, gate every process change against that map, and audit the map annually against physical reality. None of this requires new technology; it requires the quality function to accept that its data is safety evidence, with everything that implies for retention, traceability and change control.
Across two decades in automotive and aerospace quality, I have watched safety cases fail audits for exactly these reasons far more often than for design flaws. The design was sound; the evidence chain was not. Building that chain is unglamorous work, and it belongs to the quality department. Own it before an auditor, or a field incident, does the inventory for you.
