The split is drawn along two failure lines, not down the middle.

The compromise reading

Half-automated systems usually satisfy nobody, and for a reason.

A framework described as part discretionary and part systematic invites an obvious suspicion: that it is a hedge, taking some structure to reassure the analytical and leaving some freedom to reassure the intuitive, and delivering the weaknesses of both. That suspicion is well earned, because most hybrids are exactly that. They split by *proportion* — automate the easy decisions, leave the hard ones to the operator — which places the boundary wherever automation happened to be convenient rather than wherever it was warranted. The result is a system whose structure covers the parts that were never the problem.

Where discretion fails

Nobody audits themselves under pressure.

Discretionary trading fails at honesty rather than at insight. An experienced operator reads structure well, adapts to conditions competently, and recognises setups no rule enumerated — and the same operator, in drawdown, cannot reliably assess whether their sizing has crept, whether a losing stretch is variance or deterioration, or whether the reason for taking a trade was the reason they will later record. This is not a character flaw and it does not improve with experience, because the failure is not of skill but of position: the person under pressure is the person being assessed. Every honest account of discretionary failure lands here.

Where automation fails

Markets present context no backtest enumerated.

Full automation fails at the opposite end. A rule set encodes the situations its author anticipated, and live markets routinely produce conditions that fit none of them — a setup technically valid in a regime that invalidates it, a structure the indicators score well and an experienced operator would decline on sight. Automated systems handle these by applying the rule anyway, because that is what they are. The failure is not in the arithmetic, which is exactly as good as its specification. It is in the assumption that a specification written in advance can enumerate a live market, which nothing has yet managed.

The placement

Each half is given the territory the other loses in.

Put the two failure lines side by side and the boundary places itself. Trade selection, entry judgement, and reading structure go to the operator, because that is where human context beats enumerated rules and where automation's failure is expensive. Sizing, deployment authority, drawdown response, and the assessment of whether the edge is still working go to the arithmetic, because that is where self-audit is required and where the operator's failure is expensive. The line is not negotiated. It is derived, and it lands in a specific place that would be hard to move without moving it toward one of the two known failure modes.

The test the design has to pass

Each side must be strongest exactly where the other is weakest.

This construction only works if the failures are genuinely complementary rather than overlapping, and it is worth stating what would break it. If human judgement were also unreliable at reading live structure, no territory would be safe to hand over and full automation would win. If arithmetic were also unreliable at self-audit, the governance layer would be decoration. The design's whole bet is that the two failure profiles are close to disjoint — that people are good at the things rules cannot enumerate, and rules are good at the things people cannot honestly assess about themselves. That bet is old, well evidenced, and the reason the split holds.

• Judgement keeps what requires unenumerated context: selection, entry, reading structure.
• Arithmetic takes what requires self-assessment: sizing, authority, drawdown response, edge verdicts.
• The boundary is derived from the failure profiles, not negotiated between preferences.

The key idea

A good hybrid is defined by where it divides, not by how much it divides.

Any system can be described as part human and part rule; the proportion tells you almost nothing. What determines whether a hybrid is coherent is whether the boundary tracks the two approaches' failure modes or merely their convenience. Draw it along the failure lines and each half is doing work the other genuinely cannot. Draw it anywhere else and the system inherits both sets of weaknesses while claiming both sets of strengths, which is the more common outcome and the reason the description invites suspicion in the first place.