Maturity moves settled work out of runtime

Early Meal Companion runs treat every week as novel. The agent generates candidates, assesses each evening, composes, critiques, patches, and repeats. Households accept plans. Costs accumulate. Variance shows up in busy weeks that feel similar to last month.

After dozens of accepted plans, patterns stabilize. Busy evenings cluster. Spinach deadlines recur. Leftover preferences repeat. Re-running the full deliberation path every Tuesday spends money and introduces jitter where the terrain is familiar and strongly patterned.

White Paper Principle 8 names the mature response. Maturity moves settled work out of runtime deliberation. Decisions that closed through evidence migrate toward deterministic code, endorsed expertise, or promoted library versions. Measurement continues. Guards return the system to full deliberation when terrain shifts.

This page connects Decisions closing, Mirror expertise, and Economics cost fall. How decisions close explains closing dynamics. Closed decisions belong in code states the destination for exact work. unit cost falls with history covers economics.

Three destinations for settled work

Settled work can leave expensive runtime paths in three ways. The destination depends on the character of the decision, not on team impatience.

Deterministic code receives exact, verifiable work. Portion math for four people never belonged in cognitive units. Schema validation, allergy enforcement against endorsed knowledge, transactional patching, and idempotent purchase execution belong in the deterministic shell from day one. Maturity here means recognizing closure early, not waiting for a model to pretend arithmetic is judgment.

Endorsed expertise receives fuzzy but patterned work. The Busy Week Pattern v0.2 in the grounding reference includes applicability conditions, strategy skeleton, checks, failure modes, and fallback. Before the guard passes, the agent runs full deliberation. After promotion, strategy compresses: anchor busy evenings first, place spinach early, compose around variety, run one critique, patch locally if needed.

Library promotion receives stable cognitive unit behaviour behind unchanged contracts. AssessMealPracticality v2 may replace v1 when suites show improvement without class regressions. Callers keep the same name and decision sentence. Implementation and figures change with semantic versioning discipline.

The cognitive units still exist after compression. Agent strategy changes when expertise licenses a shorter path. That distinction prevents false maturity that deletes names and suites while smearing judgment into one prompt.

What maturity is not

Teams confuse maturity with several harmful shortcuts.

Removing evaluation because "it works now" is not maturity. Closed decisions still need exact tests. Promoted expertise still needs guards and periodic remeasurement. Hard-coding this week's preferences as eternal rules is not maturity. Context belongs in working state. Endorsed knowledge requires approval paths (knowledge is endorsed). Collapsing distinct judgments into one template because the model got bigger is not maturity. It violates one cognitive unit, one open decision and destroys locality. Skipping human gates because automation is smoother is not maturity. It violates human judgment remains part of the system.

Maturity compresses known paths. It does not smear architecture.

Maturity should move settled work out of runtime deliberation.

Thoughtware White Paper · Pr. 8

Guards license compression

The applicability guard for Busy Week Pattern is concrete architecture, not folklore.

Before reusing the pattern, the guard checks: at least two low-time evenings exist, busy evenings are clearly identified, no unfamiliar medical restriction is present, leftovers are acceptable this week, temporary requests do not conflict with the strategy skeleton. Guard passing licenses compressed strategy. Guard failing returns work to full deliberation without shame. Guards are how maturity stays safe when the world changes.

Expertise without guards produces brittle shortcuts. A pattern learned from autumn weekday traffic fails when school schedules shift. Guards force explicit re-entry to deliberation.

What this looks like in practice

Compare two runs with the same originating sentence.

Full deliberation path. Generate candidates broadly. Assess each evening with AssessMealPracticality. Compose with competing goals. Critique completely. Patch locally if one weakness remains. Accept after review stop conditions pass.

Compressed path after promotion. Applicability guard passes. Anchor three busy evenings first. Place spinach before Wednesday. Use one accepted leftover subplan. Compose remaining evenings for variety. Run exact checks and one complete critique. Patch locally if a single weakness remains.

Cost drops because repeated cognitive work shrinks. Variance drops because the skeleton is stable. Names remain because AssessMealPracticality, CritiquePlan, and judges still exist when guards fail or critique finds material weakness.

Portion calculation and allergy enforcement never entered the deliberation path. That closure was immediate. Maturity is the gradual discovery of additional closures, documented with evidence.

Relationship to architecture persistence

Architecture outlives implementation vocabulary gives maturity something stable to compress toward. Without stable cognitive unit maps, compression becomes vendor-specific hacks.

Evaluation part of architecture verifies promotions. Compression without figures is hope. Promotion criteria block silent regressions when expertise or library versions advance.

Enterprise parallel

Invoice intake matures similarly. Extraction fields that stabilize move toward deterministic validators. Classification that remains contextual stays in cognitive units with suites. Expertise patterns emerge for recurring document layouts. Cost per decision falls when history endorses closure (cost per decision).

Common mistakes

Premature compression. Pattern promotion after five demos, not after labeled evidence.

Frozen prompts instead of expertise. A static template replaces a guarded pattern without applicability conditions or fallback.

Deleting cognitive units after promotion. Names disappear. Incidents become untraceable.

Ignoring remeasurement after world change. Pandemic schedules, allergy updates, or new household members invalidate guards silently.

Measuring compression honestly

Compression saves cost only when figures prove quality held. After Busy Week Pattern promotion, compare trajectory metrics before and after: number of critique cycles, patch count, time-to-acceptance, regression cases on allergy edges. Compression that hides regressions is false maturity.

Expertise records carry version numbers for a reason. v0.2 differs from v0.1 because applicability conditions or skeleton steps changed with evidence. Treat expertise like library semantic versioning. Roll back when guards fail more often than full deliberation would have.

Library promotion for cognitive units follows the same discipline. v2 behind the same contract still requires stratified suites. Maturity without measurement is optimism. Principle 8 pairs with Principle 6: move work out, keep eval hooks active.

When compression fails

Guards exist because compression fails in the real world. A household adds an unfamiliar medical restriction mid-season. Leftovers become unacceptable for a month. School schedules change and busy evenings move. Failed guards log why full deliberation returned and feed expertise review.

Teams celebrate guard failures as signal, not embarrassment. A guard that never fails might be too loose. A guard that always fails was promoted too early. Each guard failure contributes to the evidence record that informs the next version of the expertise pattern, turning production surprises into structured improvement rather than ad hoc prompt edits. Principle 8 is iterative closure, not a one-way door.

Principle 8 keeps runtime lean while names and suites remain honest about what still deliberates.

Expertise review cadence

Schedule periodic expertise reviews even when guards pass frequently. Household patterns drift. Medical restrictions appear. Expertise records that never update become hidden technical debt with friendly names.

Reviews ask whether applicability conditions still match observed failures and whether compression still beats full deliberation on cost-adjusted quality metrics.

What to do next

Track which judgments repeat with stable outcomes on labeled cases. Promote to code, expertise, or library versions with explicit evidence and guards. Keep guards that return to full deliberation when terrain shifts, and remeasure after promotion because compression can hide new failure modes until traffic diversifies.

Review maturity proposals with one question: which decision closed, and what evidence closed it? If the answer is vague, the work still belongs in runtime deliberation.

Read next: A cognitive unit is one call.