Loop-Engineering Alignment — Full Rules (dev-pabcd §11)

Source: dev-pabcd/references/loop-engineering.md

Loop-Engineering Alignment — Full Rules (dev-pabcd §11)

Canonical full text for the §11 stub in ../SKILL.md. Rule IDs and section numbers

are stable; the router keeps a one-line summary per rule and this file holds the

complete wording. Genealogy and sources: backlog/260702_loop_engineering_backlog.md

in the initiative repo.

Context

PABCD is the macro loop: phase-level control with gates. Loop engineering supplies the

micro-loop rules inside each phase. This file holds the adopted subset in full.

§11.1 Loop values (DEFAULT)

Feedback must change the next action. A cycle whose result does not alter the next

step is not a loop — it is a retry. Read the failure delta before acting again.

The verifier outranks the prompt. Prefer deterministic evidence (tests, exit codes,

diffs) over model self-assessment; the maker does not grade its own homework when a

stronger verifier exists.

Memory lives on disk, not in the transcript: worklog, devlog, attestations. The

next iteration must be able to resume from artifacts alone.

Budget exhaustion ≠ done. Never report a budget stop as success.
Context pressure ≠ budget exhaustion. Compaction is survivable BY DESIGN because

memory lives on disk: an approaching context limit means checkpoint durable state

(goalplan, ledger, devlog) and continue after the flush — it is never grounds to close

the goal, shrink the plan, or report DONE/BUDGET_EXHAUSTED. BUDGET_EXHAUSTED

requires a bound the plan actually stated (tokens, cost, wall-clock).

Interview does not "solve" intent transfer — it produces a testable initial

loop-spec. Later evidence may force an Interview return; that return path existing is

the design's own admission that intent is refined cyclically, not settled up front.

§11.2 Terminal-state vocabulary (DEFAULT)

D is the success exit, not the only exit. Reports must name which terminal state the

close-out actually represents: DONE (verified success) · NOOP (nothing needed) ·

BLOCKED (external dependency) · UNSAFE (needs a human risk decision) · NEEDS_HUMAN

(judgment only the user can make) · BUDGET_EXHAUSTED (resources ran out — adopt

best-so-far and say so). These are agent-recognized report states, not FSM states: the

FSM still closes via D or reset; the D summary states the real terminal state.

§11.3 Repair-loop discipline (C→B returns)

The B/C inner loop is: implement → run verifier → read the failure delta → repair only

the failing delta → re-verify.

DEFAULT (LOOP-REPAIR-01): 2 consecutive failed repairs of the same failure → stop

patching, enter root-cause mode (dev-debugging). 3 → escalate: replan (return to P)

or Interview return. "One more attempt" past these thresholds is the doom loop.

HEURISTIC (LOOP-DOOM-01): 3 attestation failures in the same phase within one

work-phase → treat as no-progress and force an Interview return. Self-applied for now;

server-side enforcement is recorded as backlog.

§11.4 Loop archetype by problem type (DEFAULT, LOOP-ARCHETYPE-01)

Classify the work-phase's problem type before choosing the inner loop shape:

Spec-satisfaction — the verifier defines *done* (tests, contracts, tsc). Use the

§11.3 repair loop; it converges. Iterate until green.

Open-ended optimization — the verifier only defines *better* (scores, win rates,

adversarial opponents). Repair loops plateau at local optima here: they polish the

current strategy instead of finding the missing one. Use an **explore-and-select

loop**: generate diverse candidates in parallel (different strategies, not parameter

tweaks — LOOP-CANDIDATE-ANCHOR-01 in §10), evaluate all on the same instances, keep

best-so-far, regenerate from the winner. Stop on plateau (LOOP-PHASE-DEATH-01) or

budget; the exit state is BUDGET_EXHAUSTED with the best candidate adopted — not

DONE.

A repair loop applied to an optimization problem is a category error; the fix is loop

shape, not more cycles. A full paper-grounded explore-and-select workflow proposal

(LOOP-EXPLORE-SELECT-01: StrategyCard candidates, descriptor archive, opponent

portfolio, evaluation cascade, restart ladder) is recorded in

backlog/260702_explore_select_research.md — pending adoption via a normal PABCD pass.

§11.4a Analysis-before-regeneration (DEFAULT, LOOP-REANALYZE-01)

Repair fixes actions; only analysis revises the model that generates them (single- vs

double-loop learning). In an explore-and-select loop, every generation MUST begin with

an analysis deliverable — not a patch:

Updated problem/opponent model from the evidence (telemetry, replays, failure

deltas): what did the opponent/environment actually do, and what does it imply?

Capability-gap hypotheses: what can the artifact not currently *sense* or *do*

(e.g. opponent tracking, scouting, income estimation)? A gap hypothesis may propose

expanding the allowed patch surface — that is a P-level amendment, which is

exactly why new capabilities can never emerge from a surface-bounded repair loop.

Candidates are then sourced from these hypotheses (§10 LOOP-CANDIDATE-ANCHOR-01), and

the next P quotes them (§10 LOOP-CONTINUITY-01). A generation that skips analysis and

regenerates straight from scores is a repair loop wearing an explore costume.

§11.4b Mechanism activation proof (DEFAULT, LOOP-MECHANISM-PROOF-01)

Scores select candidates; only traces verify mechanisms. When a candidate's value is a

new branch (a defense mode, an endgame converter, a special-case handler), adoption

requires activation evidence: a debug counter, stderr line, or replay trace showing

the branch fired on the instances it was built for, with the intended effect.

Failure signatures that mandate instrumentation before the next candidate:

Zero-delta ablation: the solo candidate scores exactly baseline on its target

instances. The correct read is "the mechanism probably never ran", not "the

mechanism is weak". A branch that runs and loses moves the metric somewhere

(different HP, different turn counts); byte-identical outcomes mean dead code path.

Combo masking: the mechanism ships inside a multi-feature combo whose aggregate

score improves. The combo's gain proves nothing about THIS branch; without

activation evidence a dead mechanism rides along and its target failures persist

as "residual" losses that look opponent-caused.

Real case (2026-07, NEXT NATION bot): an endgame HP-race branch (bank gold from T150,

upgrade HQ to win TURN_LIMIT ties) was adopted inside a combo that raised the loss-gate

1.5 -> 7.5. Its solo ablation had scored baseline-exact 1.5 — recorded as "weak alone"

and never instrumented. The branch was in fact never arming: two early-return trainers

drained the bank every turn, and a warriors-at-HQ precondition kept cancelling the

reserve. The residual draws were attributed to the opponent ("they also stall") until a

colleague's hand-tuned bot won the same official battle with plain late HQ upgrades.

One debug stderr line (gold/reserve/atHQ per turn) exposed the leak in minutes; the

same line at adoption time would have caught it one work-phase earlier.

Corollary (LOOP-RESIDUAL-TRACE-01): every residual failure carried through D gets a

mechanism-level trace ("branch X armed at T150, banked 2400 by T174, upgraded") or the

explicit label unexplained. Plausible environment stories are hypotheses, not

evidence. And when a peer artifact achieves the objective on an instance we fail

(LOOP-PEER-CONTRAST-01), the next generation's first deliverable is the behavioral diff

of the two traces — the peer's trace is a free counterexample to the environment story.

§11.5 Unattended-loop resource policy (DEFAULT)

Goal-mode (unattended) loop-specs must state: tool/credential scope (what the loop may

touch), token/cost budget, and a wall-clock bound. For C4 surfaces, an unattended loop

with unstated scope is an ESCALATE-class omission — stop and ask before running it.

§11.6 Continuation doctrine (DEFAULT, LOOP-CONTINUE-01)

Exception to this file's completeness claim: §11.6 full text currently lives in the

§11 router index (../SKILL.md §11.6). Stub recorded here so the section sequence

stays visible from the reference file; consolidation into this file is a pending

follow-up, not an editorial accident.

§11.7 Divergence/collapse (DEFAULT)

Exception to this file's completeness claim: §11.7 full text currently lives in the

§11 router index (../SKILL.md §11.7). §11.8 below refines its cost model; the one

provenance widening §11.8 makes is named explicitly inside §11.8.

§11.8 Divergence cost tiers (DEFAULT, DIVERGE-TIER-01)

Divergence (§11.7) defaults to CONCEPTUAL candidates, not implemented ones. Choose the

cheapest tier that can kill the wrong option:

Tier 0 — inline brainstorm. The planning session itself lists options with

trade-offs inside the plan/interview. No dispatch. Default for ordinary uncertainty.

Tier 1 — conceptual candidate docs (the divergence default). 2-3 parallel

candidate lanes each produce ONE one-page candidate direction doc (no code, no

worktrees) with mandatory front-matter: assumptions, risks, kill-criteria,

evidence-needed. Lane research is read-only; the doc itself is written by the

collapse owner from lane findings, or by a lane worker whose write scope is the

plan archive — whichever the runtime's dispatch doctrine allows. The MAIN session

(collapse owner) performs critique/triage directly — it holds the most context; a

separate cross-critique round is waste. Collapse gate: N candidate docs with

filled front-matter AND per-candidate provenance — search provenance per §11.7

for externally-sourced candidates; for candidates grounded in the local codebase

a repo-evidence path is acceptable, an EXPLICIT §11.8 AMENDMENT to §11.7's

search-only wording (named here so the widening is deliberate). The mandatory

front-matter keeps the gate otherwise stricter than §11.7. Cross-critique rounds

are NOT a gate condition.

Tier 2 — implementation spike (rare escalation). Parallel worktree

implementations judged by the same verifier, ONLY when both hold: (a) the choice is

load-bearing and Tier-1 candidates genuinely conflict on it, and (b) judging

requires running code (performance assumptions, live API contracts, deceptive local

metrics). Expected frequency: 0-1 per unit. Tier-2 entry is a recorded P-level

decision.

Budget rationale: subagent tokens may be near-free, but wall-clock and the collapse

owner's triage attention are not. Tier inflation (defaulting to Tier 2 because

subagents are cheap) is a discipline violation; so is tier deflation that lets a

load-bearing conflict collapse from paper arguments alone.

Unknowns lane: the first Tier-1 dispatch of a research-heavy or unfamiliar-surface

unit SHOULD be a blindspot/unknowns pass (known unknowns, unknown knowns recoverable

from references, unknown unknowns from codebase/web search), so candidates are sourced

from evidence, not parameter tweaks (§10 LOOP-CANDIDATE-ANCHOR-01).

Topology: explorers neither message each other nor spawn their own workers (star, not

mesh). Candidate/critique exchange is file-mediated through the unit's archive; the

collapse owner schedules rounds and owns the collapse.

Genealogy: adopted 2026-07-07 from the codexclaw harness-divergence interview

(codexclaw devlog/_plan/260707_harness_divergence_interview/), which distilled the

Fable field-guide unknown-discovery practices into a cost-tiered divergence default.