# Paper 08: Retrospectives, Gardening, and the Comprehension Budget

**Series:** Aigile deep dives, paper 08 of 08. Elaborates working paper Sections 5.3 and 11. Depends on Papers 02 and 05.

## Abstract

The retrospective is aigile's learning organ, and it differs from its classic ancestor in three ways. First, its outputs must land in artifacts, because half its participants have no memory: routing learnings to the correct layer is the entire trick. Second, it produces a category of learning that human teams never had to write down, collaboration knowledge, which must compound in the durable layer or evaporate daily. Third, it carries a maintenance duty its ancestor lacked: the gardening rule, under which the document layer is refactored with the seriousness of code, because it is load-bearing. Woven through all three is the comprehension budget: the deliberate expenditure that keeps human mental models accurate in a system humans no longer implement, without which the feedback quality every other paper depends on silently decays.

## 1. Position in the framework

The retrospective is the terminal learning edge (Paper 01, R1): the channel through which process reality flows back into the durable layer. It executes Paper 05's promote-and-demote signals, refactors Paper 03's constitution, recalibrates Paper 04's sizing parameters, and hosts, at the agentic end, the self-improvement circle that Paper 06 anchors.

## 2. The problem in full

### 2.1 Learning with amnesiac colleagues

In a human retrospective, the learning is applied by the same heads that generated it; artifacts are optional reinforcement. In a mixed or agentic team, the entity that will do next week's work was not in the room and retains nothing. A retro insight that does not become an artifact in the correct layer has a lifespan of one human memory and zero agent sessions. The retro is therefore not a discussion with notes; it is a pipeline whose output format is spec-diffs, amendment proposals, parameter changes, and prompt-norm updates.

### 2.2 The unwritten curriculum

Human teams never documented how to phrase things so a colleague understands, because colleagues learned: the phrasing curriculum was absorbed socially and permanently. Agents reset. Which spec wordings produce misreadings, which task shapes exceed a session, which prompt patterns hold constraints, which examples anchor behavior: this **collaboration knowledge** is genuinely novel as a documentation category, and it compounds only if written into the durable layer (constitution content type five, Paper 03). A team that loses it daily pays the same tuition every day.

### 2.3 The atrophy problem

Classic agile built system understanding in developers' heads as a free by-product of implementation. When agents implement, the by-product disappears, and the humans' mental models thin. This is not sentimental: every load-bearing human act in the framework, validation verdicts (Paper 06), andon decisions (Paper 05), amendment judgment (Paper 03), review sampling (Paper 07), depends on the human holding an accurate model of the system. Thin models produce confident, wrong verdicts, and the framework cannot detect the thinning from inside, because the verdicts still arrive on schedule. Comprehension must be budgeted, spent, and audited like any other scarce input.

### 2.4 Monotonic growth

Retro outputs accumulate. Twenty features in, an ungoverned durable layer is a statute book: amendments upon amendments, half obsolete, consuming agent context in every session and human patience at every reading. Lessons must be composted into principles, not stockpiled as incidents, and composting is scheduled work, not an aspiration.

## 3. The mechanism

### 3.1 The three-output retrospective

Held after each validated feature (and doubling, in small teams, as the Paper 05 conformance sweep), the retro routes everything it produces to one of three destinations:

1. **System learnings** ("the ingestion module actually behaves like X"; "this vendor API has a hidden rate shape") become spec-diffs, decision records, or, when fundamental, amendment proposals. Routing test: which layer would an agent need to read to avoid the surprise next time?
2. **Collaboration learnings** ("acceptance criteria phrased as scenarios misread less than phrased as rules"; "slices touching two aggregates exceed one session") become prompt norms, sizing-parameter changes, and spec-writing guidance in the durable layer. This is also where Paper 07 incidents deposit their lessons (Orderly's confirmation-framing drift).
3. **Drift findings** from the sweep are batch-triaged per Paper 05's table, and the deviation log is read for over-specification patterns.

### 3.2 The gardening rule

> On a defined cadence (for example every Nth retrospective, or when Layer 1 grows past a size tripwire), the retrospective is a gardening session: clauses are consolidated into principles, superseded law is deleted, chronically drifting sections are promoted or demoted per the Paper 05 signals, and the whole ships as an ordinary amendment PR with the version history intact.

Gardening is refactoring applied to the document layer, and it carries refactoring's discipline: behavior-preserving by intent (the law's meaning survives consolidation), reviewed as a diff, reversible through versioning. The metric of a good gardening session is negative net lines with zero lost enforcement.

### 3.3 The comprehension budget

The budget is spent through mechanisms the framework already contains, named here as deliberate expenditures rather than side effects:

- **Brains-eyes-hands demos** (Paper 06): unmediated contact with the running system, the highest-density comprehension purchase available.
- **Review sampling** (Paper 07): reading real diffs keeps the code-level model alive; this is a second function of sampling beyond calibration.
- **Retro participation**: the retro's narrative reconstruction of the feature (what surprised us, what deviated, what drifted) is where the humans' system model is explicitly rebuilt; agents attend as reporters, humans attend as learners.
- **Derived-layer reading** (Paper 02): regenerated overviews give honest, current descriptions on demand, and reading one before an andon decision should be normal practice, not an admission of ignorance.
- Optionally, **deliberate hand implementation** of selected slices, chosen for comprehension value rather than throughput, in areas where verdict quality has been weak.

The budget is audited at the retro with an uncomfortable question asked explicitly: could each of us still give a competent andon decision in the areas we validated this cycle? Weak answers reallocate the next cycle's spend.

## 4. Normative protocol

> **R29 (Artifact-terminal retros).** A retrospective's outputs are artifacts routed to defined layers before the session ends; an insight without an artifact and a destination is recorded as not yet learned.
> **R30 (Collaboration corpus).** Collaboration learnings are written to the durable layer's process section and are subject to the same amendment and gardening lifecycle as all law.
> **R31 (Gardening cadence).** Gardening runs on a standing cadence or size tripwire, whichever fires first; its PRs target negative net lines with preserved meaning; skipping two consecutive gardening windows is an incident.
> **R32 (Budget audit).** Each retro records comprehension spend (demos driven, samples reviewed, slices hand-built) per human and flags areas where validation authority is exercised without recent contact.

## 5. Orderly worked example

Feature retro after consolidated invoicing. System learnings: the pilot's ERP batches order events hourly, invalidating a freshness assumption (spec-diff to ingestion criteria; one sentence proposed for the constitution and rejected under the R7 andon test, landing as a decision record instead). Collaboration learnings: scenario-phrased criteria misread zero times versus three misreadings for rule-phrased ones (prompt norm updated, R30); two-aggregate slices exceeded sessions twice (sizing guidance updated, Paper 04 parameter note). Drift sweep: four descriptive findings batch-approved, one demotion flagged. Gardening (sixth retro, cadence fires): the constitution shrinks from 31 clauses to 24 by consolidating four retry-related clauses into one idempotency principle and deleting three superseded by v2.0.0; net minus 41 lines, all enforcement preserved, shipped as v2.5.0. Budget audit: one human validated billing features all cycle without reviewing a single billing diff; next cycle's sampling is reallocated toward billing, and she hand-builds one small slice there. The audit question was uncomfortable and five minutes long, which is the intended price.

## 6. Failure modes

- **Discussion-terminal retros:** insights nodded at, artifacts absent, the same lesson relearned quarterly. Counter: R29's closing check is the agenda's last item.
- **Stockpiling:** every lesson appended, nothing composted, the statute book (Paper 03) by another road. Counter: R31's incident framing for skipped gardening.
- **Budget denial:** comprehension assumed permanent because the humans were once the implementers. Detection: verdict quality decay in Paper 06 metrics correlated with zero contact in R32 records.
- **Retro capture at the agentic end:** the self-improvement circle sets its own agenda and grades its own proposals. Counter: Paper 06, R24; verdict-derived outcomes remain the only admissible fitness evidence, audited at gardening.

## 7. Metrics

Artifact yield per retro (outputs with destinations, by layer); relearning rate (lessons appearing more than once, target zero); Layer 1 net lines per gardening session; comprehension spend per human per cycle against verdict quality; time from retro to artifact merge.

## 8. Open questions

The empirical comprehension floor (shared with Paper 06): how much contact keeps a validator competent, by system size and role; whether collaboration knowledge transfers across agent generations or must be re-derived; optimal gardening cadence as a function of amendment rate; whether an agent can usefully draft the budget audit without softening it.

## 9. Derivative artifacts

Retro runbook with the three-output agenda and closing artifact check; gardening session checklist and PR template; comprehension budget worksheet; collaboration-corpus starter file (prompt norms, wording guidance) for the repository template; the uncomfortable-question card for facilitators.