An ASIL rating is not a stamp you note in the validation pack header and move past. It is a statement about how much rigour the evidence behind a perception function must demonstrate — and a reviewer who sees uniform depth across functions of different integrity will ask why.

That gap is where most perception teams first encounter the automotive safety integrity level as something more than a number inherited from the system spec. Someone upstream assigned an ASIL during hazard analysis; it arrived in a requirements document; the perception team copied it into the pack and assumed the rating itself was the deliverable. It isn’t. The rating is an instruction about where to spend evidence depth, and reading it as a label produces packs that are over-documented on the functions nobody is worried about and thin exactly where the reviewer’s risk attention concentrates.

What Does the Automotive Safety Integrity Level Actually Govern?

ISO 26262 defines four integrity levels — ASIL A through ASIL D — plus QM, “quality management,” for hazards that carry no safety integrity requirement at all. ASIL D is the most stringent and ASIL A the least; QM means the standard’s safety lifecycle does not bind that function, though ordinary quality processes still do. That ordering is the part everyone learns. The part that matters for a perception team is what the ordering demands.

Each step up the integrity scale is a step up in the rigour the evidence must show: more independent verification, tighter coverage targets, more conservative assumptions about failure, and more explicit traceability from the integrity requirement down to the test that produced the supporting result. The level is, in effect, a dial on evidence depth. A function rated ASIL B and a function rated ASIL D can use the same detector architecture and still owe the reviewer very different evidence packs, because the integrity expectation behind them differs.

This is the reframe the whole article rests on: ASIL is a decomposition discipline, not a stamp. A team that internalises that builds evidence proportionally; a team that treats the rating as a header field builds evidence uniformly and gets challenged on the high-integrity functions where uniform depth reads as insufficient.

How Is an ASIL Assigned to a Perception Function?

The rating does not originate with the perception team. It comes out of ISO 26262’s hazard analysis and risk assessment — HARA — performed at the vehicle level, long before anyone is profiling a detector. HARA evaluates each hazardous event along three axes and combines them into an integrity level:

• Severity (S) — how bad the harm is if the hazard occurs, from no injury to life-threatening or fatal.
• Exposure (E) — how often the operating situation that enables the hazard arises in normal driving.
• Controllability (C) — how reliably a driver (or the system) can act to avoid the harm once the hazard is present.

The three ratings index into a table that yields the ASIL. A high-severity hazard that occurs in a common driving situation and is hard to control lands at ASIL D; lower the severity, the exposure, or raise the controllability, and the level drops, eventually to QM. The crucial point for a perception team is that the S/E/C reasoning already happened — by the time an integrity level reaches you, the severity, exposure, and controllability judgments are baked in. Your job is not to re-derive the rating. It is to read the rating as the rigour budget it encodes and trace each integrity expectation to a specific evidence surface.

That tracing discipline is the same one that underpins a perception validation evidence package reviewers trust — ASIL simply tells you how deep each surface in that package has to go.

Evidence Depth by ASIL Level

The table below maps each integrity level to the kind of evidence depth a reviewer expects on a perception function carrying that rating. It is a planning rubric drawn from how these reviews tend to run, not a clause-by-clause restatement of ISO 26262 — the standard’s normative requirements are the authority; this is the practitioner’s read of what “proportionate rigour” looks like in a pack (observed across our automotive validation engagements; not a published benchmark).

ASIL	What the rating says	Evidence depth the reviewer expects
QM	No safety integrity requirement; ordinary quality processes apply	Functional test coverage and basic performance evidence; no independent safety-case argument required
A	Lowest safety integrity tier	Documented requirements traceability, nominal-condition test results, defined acceptance thresholds
B	Moderate integrity	The above plus structured edge-case coverage and a stated rationale for test-set composition
C	High integrity	The above plus degraded-condition and adversarial-stress evidence, with independence between development and verification visible in the pack
D	Highest integrity	The above plus the most conservative failure assumptions, the widest operating-condition coverage, and explicit traceability from each integrity requirement to the test that satisfies it

The shape of the table is the lesson. Depth accumulates as the level rises; an ASIL D function carries everything a lower-rated function carries and more. A reviewer reading a pack does not expect the QM functions and the ASIL D functions to look the same. When they do, the uniformity is itself a finding — either the low-integrity functions are over-documented, which wastes effort, or the high-integrity functions are under-documented, which fails the review.

For what ASIL D specifically demands when a perception function lands at the top of the scale, our breakdown of ASIL D for perception evidence goes a level deeper than this hub does.

How Do You Let the Classification Shape the Pack?

The practical move is to stop treating evidence depth as a global setting and start treating it as a per-function decision keyed to the integrity level. A pack assembled this way reads differently: the high-ASIL functions carry visibly more — more conditions tested, more independence between who built the model and who verified it, more explicit links from requirement to result — and the low-ASIL functions carry proportionately less, with the difference deliberate rather than accidental.

We see the same pattern repeatedly: the clarification rounds that drag a release out are concentrated on the high-integrity functions where the evidence depth did not match the rating. Concentrating rigour where the ASIL is highest is what reduces those rounds. It is also why uniform-depth packs feel safe to produce and then stall in review — they spend effort evenly and the reviewer’s attention is not even.

This proportionality is not unique to automotive. It is a graded instance of the general reliability-evidence discipline that any safety-relevant AI system needs, which is why the automotive perception validation package reviewers sign against frames ASIL as the automotive projection of a broader pattern: match evidence depth to consequence. ASIL just puts a formal scale on the consequence.

How Does ASIL Decomposition Change the Pack Structure?

ISO 26262 allows a high-integrity requirement to be decomposed into lower-integrity sub-requirements allocated to sufficiently independent elements. A single ASIL D requirement can, under defined conditions, be split — for example into two redundant, independent sub-functions each carrying a lower ASIL — such that the combination still satisfies the original integrity target. This is ASIL decomposition, and it is the reason the discipline is named one.

For a perception team this has a direct structural consequence. If a high-integrity perception function is decomposed — say, a primary detector and an independent confirmation path — the evidence pack is no longer one monolithic high-ASIL argument. It becomes a structured set of sub-function arguments plus an argument that the decomposition itself is valid: that the sub-functions are genuinely independent, that their combination meets the original target, and that the independence claim holds under the failure assumptions the original ASIL demanded. The pack mirrors the decomposition. A reviewer who understands decomposition will look for exactly that structure, and a flat pack that asserts a single high-integrity claim without showing how the parts combine will draw questions.

This is also where the work connects to the broader engineering of computer vision systems that have to survive review, not just demonstrate accuracy in a notebook — the decomposition shapes the architecture and the architecture shapes the evidence.

Why Doesn’t Naming the ASIL in the Header Satisfy a Reviewer?

Because the rating is a claim, and a claim needs evidence proportionate to what it asserts. Writing “ASIL D” at the top of a pack tells the reviewer how much rigour you owe; it does not pay it. The reviewer’s question is never “what rating did you write down” — it is “does the depth of this evidence match the integrity you claimed.” A correctly stated ASIL with thin evidence behind a high-integrity function is worse than an honest QM, because it sets an expectation the pack then fails to meet.

The teams that clear review fastest are the ones whose packs make the proportionality obvious: each integrity requirement points to the test that satisfies it, the depth visibly tracks the level, and a decomposed function shows its sub-arguments. That is the difference between treating the automotive safety integrity level as a stamp and treating it as the decomposition discipline it was designed to be.

FAQ

How does automotive safety integrity level work, and what does it mean in practice?

ISO 26262 defines four levels — ASIL A through D — plus QM, where ASIL D is the most stringent and QM carries no safety integrity requirement. In practice the level is a dial on evidence depth: each step up demands more independent verification, wider coverage, more conservative failure assumptions, and tighter traceability. For a perception team it means the rating tells you how much rigour the evidence behind a function must demonstrate, not just what number to record.

How is an ASIL assigned to a perception function, and what do severity, exposure, and controllability contribute?

The rating comes from ISO 26262’s hazard analysis and risk assessment (HARA) at the vehicle level, not from the perception team. HARA rates each hazardous event on severity (how bad the harm), exposure (how often the enabling situation arises), and controllability (how reliably the harm can be avoided), then combines them into an ASIL. High severity in a common, hard-to-control situation lands at ASIL D; relax any axis and the level drops toward QM.

What does each ASIL level demand in terms of evidence depth for a perception validation pack?

Depth accumulates as the level rises. QM needs functional coverage and basic performance evidence; ASIL A adds requirements traceability and defined thresholds; B adds structured edge-case coverage; C adds degraded and adversarial-stress evidence with visible development/verification independence; D adds the most conservative failure assumptions, the widest operating-condition coverage, and explicit requirement-to-test traceability. An ASIL D function carries everything below it plus more.

How do we let the ASIL classification shape which evidence surfaces need the most rigour?

Stop treating evidence depth as a global setting and key it to each function’s integrity level. High-ASIL functions visibly carry more — more conditions, more independence, more explicit links from requirement to result — and low-ASIL functions proportionately less, with the difference deliberate. Concentrating rigour where the ASIL is highest reduces the clarification rounds that otherwise pile up on exactly those functions.

How do we trace an ASIL expectation to the specific test that produced the supporting evidence?

Each integrity requirement should point directly to the test result that satisfies it, so a reviewer can follow the claim to its evidence without inference. The higher the ASIL, the more explicit that traceability needs to be. This per-requirement linkage is what turns a stated rating into a demonstrated one.

Why does naming the ASIL rating in the pack header not, on its own, satisfy a reviewer?

Because the rating is a claim about how much rigour is owed, and a claim needs evidence proportionate to what it asserts. The reviewer’s question is whether the evidence depth matches the integrity claimed — not what number appears in the header. A stated ASIL D with thin evidence is worse than an honest QM, because it sets an expectation the pack then fails to meet.

How does ASIL decomposition let a high-integrity function be split, and what does that imply for the pack?

ISO 26262 permits a high-integrity requirement to be decomposed into lower-integrity sub-requirements allocated to sufficiently independent elements, so the combination still meets the original target. For a perception function this means the pack becomes a set of sub-function arguments plus an argument that the decomposition itself is valid — that the sub-functions are genuinely independent and combine to satisfy the original ASIL. The pack mirrors the decomposition rather than asserting one monolithic claim.

How does the ASIL classification relate to ISO 26262’s HARA, and where does the S/E/C rating originate?

The ASIL is the output of HARA, performed at the vehicle level before perception development begins. The severity, exposure, and controllability ratings are assigned there and combined into the integrity level. By the time a rating reaches the perception team, the S/E/C reasoning is already baked in — the team’s job is to read the rating as a rigour budget and meet it, not to re-derive it.

The closing question for any perception team is not “what ASIL did we inherit” but “does the depth of every evidence surface in this pack match the integrity each function was assigned” — and where a high-integrity function was decomposed, whether the pack shows the independence its rating demands.