Decision cost falls with history

Software built from named cognitive units has a property ordinary software lacks: marginal cost can fall as the system ages, if someone runs the closing loop. Without that loop, every case keeps paying for cognition that has become arithmetic. The bill never concentrates on genuinely open terrain because the system treats every decision as equally open regardless of how many times it has answered the same question correctly.

Helpful context: Maturity moves settled work out of runtime explains where closed work goes. How expertise forms and how decisions close supply the closing loop. Paying for reliability covers spend on the open remainder. This page is the ROI story for governance skeptics.

Ordinary features accumulate complexity and usually accumulate cost. Named cognition can do the opposite when closing is deliberate and measured. The mechanism is straightforward: work that was once open becomes closed, and closed work moves to code or cache where its marginal cost approaches zero.

Before and after call counts

In month one of the Meal Companion's busy week feature, every run receives full deliberation. GenerateCandidates runs for 7 evenings times 2 loop iterations, full compose and critique follow. Roughly seventy cognitive unit calls produce one accepted plan. Minutes pass before acceptance. The system is exploring because it has no history to draw on.

By month six, with the Busy Week Pattern promoted and guards passing, the picture has changed. The agent anchors busy evenings, composes around accepted work, runs one critique, and applies local patch if needed. Roughly twenty calls produce an accepted plan. Seconds replace minutes on the compressed path.

The cognitive units still exist. What changed is strategy: loop compression when guards pass, deterministic guards shrinking population, and portion math never having lived in a cognitive unit at all because it was recognized as closed work from day one. The cost difference comes from fewer calls on paths that no longer require open judgment.

1. 01Open judgmentHighest marginal cost
2. 02Cached answersRepeat classes stop paying full price
3. 03Extracted rulesClosed classes with cognitive fallback
4. 04Deterministic guardsPopulation shrinks before expensive calls

Closing moves work down the stack from open cognition toward cache, rules, and code.

How the fall happens

Closing is not automatic. Teams that never retire spend, never watch hit rate per class, and never promote expertise patterns pay forever for work that stabilized months ago. The mechanism requires active ownership at each step.

Cases that answer once and return from cache when the class has closed represent the first reduction. Closing classes becoming rules with cognitive fallback represent the second. Deterministic guards that shrink the population before expensive cognitive units run represent the third. Expertise patterns that compress loops when applicability guards pass represent the fourth. New territory appearing as new named cognitive units rather than silent accumulation inside old ones prevents complexity from creeping back in.

Each step requires instruments: cache hit rate, guard clearance, guard pass/fail on expertise patterns, loop iteration counts per plan. Without signals, teams cannot tell whether cost is falling or merely shifting to a different edge. The fall is real only when the instruments confirm it.

Its marginal cost falls as it ages.

The cognitive unit, Ch. 14

Maturity has ROI

Maturity moves work out when evidence supports promotion to code, expertise, or library versions. The ROI appears in call counts and variance reduction, not in slide decks about "learning." Governance skeptics respond to numbers. Call counts per accepted outcome at month one versus month six is the number that makes the case.

Busy Week Pattern v0.2 did not remove AssessMealPracticality. It changed agent strategy when applicability guards pass: anchor busy evenings first, skip regenerating accepted evenings, one critique instead of loop until perfect. Eval confirms compression did not hide new failure modes. The cognitive unit remains available for cases that fall through the guard, which means the system gracefully degrades to full deliberation on unfamiliar terrain while saving budget on the familiar.

Portion calculation closed into deterministic code from day one never contributed to falling marginal cost because it never should have been open. The fall comes from work that was open and became closed or patterned through history. That distinction matters because it means governance and eval investment produce economic return over time, not on day one.

Paying for the open remainder

Money still belongs on consequential open judgments: wrappers up to the determinacy ceiling where consequence justifies it. The fall in decision cost is not an argument for starving open work or skipping evaluation as engineering. It is an argument for stopping payment on work that is no longer open, so that budget concentrates where genuine judgment still operates.

What this looks like in practice

An operations review for the Meal Companion after six months shows the fall in concrete numbers:

Metric	Month 1	Month 6
Mean cognitive unit calls per accepted plan	68	22
Median time to accepted plan	4.2 min	1.1 min
Busy Week Pattern guard pass rate	n/a	61%
Allergy violations in production	0 (exact guard)	0
Open spend per plan (illustrative)	$0.29	$0.11

Governance skeptics watching these numbers see call counts and open spend per plan falling while exact predicates remain at zero violations. Discipline paid in cost and variance reduction. The allergy row staying at zero across both periods confirms that guards maintained safety while cost optimisation happened elsewhere in the system.

Connection to closed work left open

The mirror failure is paying forever because closed work stayed inside a cognitive unit. The price of closed work left open quantifies that waste across the system's lifetime. Decision cost falls when closure happens. It flatlines when closure does not. The two pages together form the economic argument for active closing: one shows what the system gains, the other shows what it pays by not acting.

Variance reduction is part of ROI

Falling cost also means fewer sources of variance. Fewer cognitive unit calls on stable classes mean Tuesday plans stop drifting when busy-week compression applies. Households experience consistency. Support tickets fall. ROI appears in operations metrics alongside inference bills.

Tracking variance alongside mean cost reveals whether the system is truly maturing or merely getting cheaper while wobbling. Standard deviation of call count per accepted plan, abstention rate stability, and regression recurrence tell the stability story. A system that costs the same but wobbles week to week has not matured. A system that costs less and also stabilizes has genuinely moved settled work into deterministic paths.

When cost fall stalls

If call counts flatline after month three, the closing loop stalled. Common causes include no owner for promotion reviews, expertise patterns that were never guarded, cache without hit-rate instrumentation, or new features adding edges faster than closure removes them. Diagnosing with the same instruments used to prove the fall (calls per outcome, guard pass rate, cache hit rate, closed-work extraction tickets closed versus opened) reveals which mechanism stopped working.

Expertise without eval is compression theatre. Promoting the Busy Week Pattern without applicability guards and regression suites may reduce cost briefly while failures rise. Maturity requires evaluation as engineering on the compressed path: same busy-week cases, same allergy exact checks, faster time to acceptance. Without that confirmation, the cost savings are an accounting illusion that externalize to incidents later.

Closing loop ownership

Falling marginal cost requires an owner for the closing loop. Without one, busy-week compression never promotes, cache hit rates stay unmeasured, and closed-work extraction tickets linger while new features add edges. A named role reviews monthly: which case classes hit one hundred percent on exact lane, which expertise patterns have stable guard pass rates, which cognitive units still run closed checks inside templates. The owner does not need to close everything. They need to prevent silent stagnation where month-six looks like month-one on call counts.

The invoice intake parallel

Invoice intake shows the same closing economics on different terrain. Vendor identification that stabilized after ten thousand examples may promote to deterministic routing with cognitive fallback on novel formats. Field interpretation cognitive units shrink to the remainder only. Finance sees intake cost per document fall while exception rate stays visible. Closure without measurement is automation theatre. Closure with suites and guard pass rates is architecture maturity paying back in operations.

Cache hits on repeat case classes appear in operations review alongside cognitive unit call counts. Meal Companion may cache accepted busy-week anchor patterns when household constraints match prior weeks. Invoice intake may cache vendor identification for recurring suppliers. Each cache class needs invalidation rules when terrain shifts. Uncached repetition is money left on the table. Uninvalidated cache is silent wrongness. Both failures have the same root: missing ownership of the closing loop.

Reuse spreads the investment

Reusing sealed cognitive units spreads eval and engineering cost across callers. Rediscovering prompts per feature keeps unit cost high even when model prices drop. Library maturity metrics predict falling unit cost at the organisation level. Immature fork culture predicts flat or rising cost despite cheaper models, because each fork requires its own eval investment and its own closing loop.

Tracking cost per successful decision (not cost per token alone) reveals whether maturity investments are paying back. Successful decision includes abstention when abstention was correct. A system that refuses when it lacks authority is cheaper than a system that guesses and triggers human review on the exception path.

What to do next

Instrumenting call counts and loop depth per accepted outcome on one feature is the starting point. Identifying one repeating case class eligible for cache, rule extraction, or expertise promotion reveals the first closing opportunity. Promoting with eval evidence and guards that return to full deliberation when terrain shifts protects against compression theatre. Remeasuring after promotion confirms that compression produced genuine savings rather than hidden failures.

See price of closed work, how decisions close, and cost per decision.

Read next: The price of closed work left open.