External Methodologies Cross-Check — IACDM & AI-DLC 2026
Two independent 2026 methodology papers mapped onto iterion's mechanisms: what they validate, the handful of rules worth importing (folded into workflow_authoring_pitfalls.md), and what iterion deliberately does differently. Read this when weighing the next methodology paper against iterion, or when talking to operators who arrive speaking these frameworks' vocabulary.
Sources.
- IACDM — Interactive Adversarial Convergence Development Methodology (J. Moreira, arXiv:2604.16399). Core claim: the verification gap is structural — an LLM is a stochastic generator with no internal
V(artifact) → {correct, incorrect}, so "the tool is irrelevant; the process is determinative." Mechanisms: 8 phases with discrete gates held by Verification Agents (VA-automatic = tests/linters/compilers with a binary verdict; VA-human = the operator, for semantic adequacy), adversarial critique through specialized lenses (each owning an exclusive failure class), a persistentspecs/knowledge repo, context granularization. Evidence: 20+ projects at a single institution (N=1, self-acknowledged, with testable hypotheses). - AI-DLC 2026 — AI-Driven Development Lifecycle (AWS + The Bushido Collective). Work decomposed into Intents → Units → Bolts, driven by programmatically verifiable completion criteria and backpressure (quality gates that reject non-conforming work without prescribing the approach); three human-involvement modes (HITL / OHOTL / AHOTL) matched to work verifiability; knowledge artifacts; ops-as-specs.
What they independently validate (no action needed)
Both papers converge — from different starting points — on choices iterion pre-arbitrated and shipped:
| Their concept | iterion's shipped form |
|---|---|
| Backpressure over prescription; programmatic completion criteria | Termination contract (axis_complete ∧ gates green) + verify_run on the REAL exit code (ADR-044/058) |
| Verification is always external to the generator (GA/VA model) | "Gates stay deterministic — never an LLM judgment" (C103–C106) |
| "Ralph Wiggum pattern" / the 19-agent trap: a simple loop with good feedback beats complex multi-agent architectures | ADR-058 v2: ONE campaign agent + deterministic gate + bounded continuation loop, which replaced the reviewer/fixer relay fleet-wide |
| Validation theater / RLHF sycophancy; request refutation, never validation | Goodhart + façade (workflow_authoring_pitfalls.md); adversarial review posture |
| "Scope presence blindness": modules that were never implemented do not fail tests | The goai→claw façade case study + the right-artifact discipline (git add -N / git add -A) |
| Convergence gate + the "convergence perfectionism" antipattern | The asymptote rule + bounded max_passes + iterion bench asymptote |
| A gate verifies completeness of critique, not absence of findings | sec-audit's scan_health: hard-fail when the scanner floor produced no output — the zero-finding façade, caught deterministically |
| Ops-as-specs (scheduled / reactive triggers) | iterion schedule, the pkg/trigger spine, manifest invocations: |
Persistent knowledge layer (specs/, knowledge artifacts) | The three knowledge channels: workspace memory / board issues / committed bilans |
| Context granularization ("one session per phase, one phase per module") | Fresh context per pass, per-node sessions, sub-agent offloading |
| Mode matched to work verifiability, not just risk | permission: precedence, interaction: modes, supervisors — per node, not per run |
The shared thesis — quality comes from the process, not the model — is iterion's product thesis (philosophy.md): a deterministic orchestration engine over interchangeable backends.
The imported delta
Folded into the pitfalls doc as cross-checked rules:
- Teach-back, not confirmation — invert the agreement bias: the agent restates the mission; the human judges the restatement.
- Verify divergence, not correctness — refutation framing for every self-check.
- Scope inventory — absence doesn't fail tests; check presence deterministically.
- Cost-tier switch point — cheap models on well-specified units once context is externalized.
- A lens must own a failure class + concentration analysis at synthesis.
- The ~40–60% context budget — cap unbounded inputs before the LLM.
- Criteria escape rate — a defect that passed the gates is a gate bug.
First application: the feature-dev and whole-improve-loop campaigns teach-back an ambiguous mission via ask_user_async (ADR-081) and keep working under stated assumptions — a non-blocking variant that IACDM's operator-gated Phase 0 cannot express.
Parked candidates (worth a look when the need shows up): visual backpressure (vision-model comparison of implementation screenshots against a design reference, for UI bots — the Playwright/browser-pane plumbing exists); pre-launch adversarial spec review (a cheap judge attacks the mission statement — contradictions, hidden complexity, unvalidated assumptions — before an expensive campaign burns budget on the wrong problem).
Vocabulary mapping (for operators arriving from these frameworks)
| Their term | iterion mechanism |
|---|---|
| HITL — human validates each step | permission: ask, interaction: human, human nodes, plan gates |
| OHOTL — human observes live, may interrupt, no blocking gate | supervisor blocks, operator chat/steering, interaction: async |
| AHOTL — human sets criteria, reviews output | campaign + deterministic verify gate + termination contract |
| Completion criteria | termination-contract flags + the gate's all-of conjunction |
| Backpressure | verify gates / scanners / diagnostics — they reject, never prescribe |
| VA-automatic / VA-human | tool & compute gate nodes / human gates + review scope |
| Bolt | one pass of a continuation_loop |
Knowledge artifacts / specs/ | memory scopes, bundle skills, committed bilans |
| Criteria escape rate | gate escapes, noted in a bot's bilan ("Findings / misses") |
Deliberately not adopted
- The 8-phase liturgy with VA-human at most gates (IACDM) — operator-centric by construction (its own L5 limitation concedes team scalability is open); iterion runs headless, so human gates are the exception (review scope, plan gates) and deterministic gates the rule.
- "<15% structural change between versions" as a stopping criterion — author-calibrated, N=1; a machine-checkable done-flag ∧ green gates is stronger, and is what ships.
- Unconditional Phase-0 discovery on every non-trivial task — the wrong default for bounded/mechanical missions; iterion scopes teach-back to ambiguous missions, first pass only, non-blocking.