Originally written in French. Translated by AI — the meaning has been preserved, not the prose.
Angle
A quality mechanism placed after delivery — the team that takes other people's bugs, QA as the last barrier, the defect queue, the rework metric — does not merely catch defects: it diverts to a third party the feedback that would have taught their author how not to produce them. That is what makes these mechanisms structurally losing propositions, not their direct cost: they protect the product once, and they degrade production every time. Quality belongs to those who ship in the most literal sense — they are the only ones occupying the position from which the defect could still have not existed, and every move downstream empties that position of what made it useful.
Synthesis
The movement repeats at four points in the chain, and each time it takes on the appearance of good practice.
At assignment, first. Handing a defect to whichever developer is available optimizes a schedule and cuts the link between a delivery and its cost. The author does not see what their delivery produced; whoever fixes it reconstitutes a context they did not create and finds a technical cause, never a decision cause. The opposite rule — whoever creates the defect fixes it — is then worth less for the work it assigns than for the two effects it restores: an incentive that acts before delivery, and a learning available nowhere else.
At the entrance, next. An ambiguous specification passed on as it stands does not disappear: it gets settled later, silently, by whoever writes the code. The right to refuse to produce in the fog is the exact counterpart of the duty to fix one's own defects — in both cases, you refuse to let the decision and its consequence come apart. The test written before the code makes the same move in a tooled form: it fixes the expected behaviour at the moment when it is still debatable, rather than verifying it once it has become expensive to change.
At the exit, finally, in the most institutionalized form. A QA placed as the last barrier can only block or let through; the same profession placed upstream writes the criteria, the thresholds and the risks to cover. Position, not mission, decides whether the function frames or catches up — and automation, by absorbing the repetitive part, makes that move practicable at the precise moment it was becoming indispensable.
What connects these four points is a temporal asymmetry. The cost of quality is minimal at the instant of production and grows thereafter without ever coming back down; the information needed to exercise it follows exactly the inverse curve. Every downstream mechanism therefore chooses scheduling convenience over the one position where the defect could have failed to be born.
Two conditions bound the movement, and they hold together. The first is that the responsibility handed back to whoever ships arrives as ownership and not as fault: as soon as it serves to designate a culprit, defects leave the conversations before they leave the product. The second is that their zone stay distinct from that of the product — the dialogue between product choice and technical quality informs the decision, it does not pool it, failing which responsibility dissolves without anyone having moved it.
The final reckoning reads in the legacy. What we call inheritance is in part current production: every transferred defect, every accepted ambiguity, every unwritten test, every regression replayed by hand is a catch-up chosen today, whose trace the product will keep long after the constraint that motivated it has been forgotten.
Tensions / contradictions
Putting quality back with whoever ships and leaving them the latitude to exercise it do not always go together. The rule "whoever creates the defect fixes it" works as an incentive on condition that the developer had a grip on what they were shipping; imposed under a deadline constraint they did not set, it becomes a sanction for a decision made elsewhere. These notes do not say where the threshold lies.
There is also a tension with the method for clearing an existing backlog: dealing each week with support's three most frequent items is a downstream mechanism, owned as such, useful, and contradicting the thesis as long as it stays in place. It justifies itself as a transition, which presupposes a bound that nothing here sets.
Questions
- In a team of several dozen developers, what minimal form of feedback from production maintains the link between a defect and its author, when having that author fix it is no longer practicable?
- Which external constraints — certification, contractual commitment, audit — make a final check non-negotiable, and how do you hold it without its becoming the only quality mechanism again?