Convergence Mechanics — How the Loop Compounds Instead of Drifts

The /evolve loop only compounds when each cycle reads prior cycles' outcomes and lets them change behavior. Append-only ledgers that no step reads are write-only artifacts — they accumulate without compounding.

This reference documents the four feedback mechanisms that turn raw cycle output into next-cycle behavior change.

Mechanism 1: Step 0 reads prior-failure surface

In Step 0 (Setup), after mkdir -p .agents/evolve, the loop reads the last 3 entries of cycle-history.jsonl. For any entry where gate field contains a FAIL marker, it extracts the failure surface (e.g. "registry-check stale", "bats-tests goals-validate") and injects the matching learning before work selection.

last3=$(scripts/evolve-read-cycle-history.sh recent 3)  # routes through BC3 LoopReaderPort (soc-y5vh.4)
fail_surfaces=$(echo "$last3" | jq -r 'select(.gate | test("FAIL|FAILED|BLOCKED")) | .gate' 2>/dev/null)
if [ -n "$fail_surfaces" ]; then
    # Search learnings for surface keywords; print whichever match
    keywords=$(echo "$fail_surfaces" | grep -oE 'registry|bats|markdown|supergate|canary|coverage|toolchain' | sort -u)
    for kw in $keywords; do
        ao lookup --query "$kw failure" --limit 2 2>/dev/null || \
            find .agents/learnings -name "*$kw*.md" -mtime -30 | head -2
    done
fi

Without this, the 2026-05-07 CI-toil learning sat for 7 days while I burned 5 cycles re-hitting the same registry.json non-determinism. Reading the learning at Step 0 would have surfaced the git ls-files fix on cycle 45.

Mechanism 2: Step 1.5 healing-first classifier

Before Step 2 (measure fitness) and Step 3 (select work), the loop classifies the cycle:

# Step 1.5: healing-first classifier — routes through BC2 CIStatusPort
# (cli/cmd/ao/ci_status_adapter.go, cycle 117 productionCIStatus) per
# soc-y5vh.2. No inline gh shell-outs.
last_ci=$(ao ci recent --limit 1 2>/dev/null | jq -r '.Conclusion // empty')
if [ "$last_ci" = "failure" ]; then
    CYCLE_MODE="restorative"
    # Read failure surface, search for matching learning
    # (see Mechanism 1)
    # Selection ladder downgrade: skip Step 3.1 (harvested) for feature
    # work; only allow harvested items typed bug/fix/ci-failure.
else
    CYCLE_MODE="feature"
fi

Restorative cycles ONLY take work that reduces CI red. New PG4 promotions, feature additions, doc growth — all blocked until last_ci=success.

This eliminates the cycle-46-47 pattern where I added new evidence files onto a CI-red base.

Mechanism 3: Hypothesis tracking for skill changes

When a cycle edits skills/evolve/SKILL.md, it MUST append to the hypothesis ledger through the typed BC3 HypothesisLedgerPort (soc-y5vh.8):

ao loop hypothesis append --id "H<cycle>.<patch>" --cycle-landed N --check-at-cycle $((N+15)) \
  --patch "<one-line>" --hypothesis "<expected effect>" --measure "<how to verify>"

This routes through productionHypothesisLedger instead of a raw append to .agents/evolve/hypotheses.jsonl; the port rejects empty and duplicate IDs. At check_at_cycle, the loop reads the ledger with ao loop hypothesis list (one JSON record per line), evaluates each PENDING row's measure, and writes the verdict (VERIFIED / FALSIFIED). Falsified hypotheses are revisited: either the patch is wrong, or the measurement was wrong.

Without this, cycle 45's 6 patches landed unmeasured and 2 of them (H45.2 source-surface auto-rebuild, H45.3 grep-based gate parsing) were silently inert for the next 5 cycles — text in SKILL.md but no harness automation behind them.

Mechanism 4: Convergence criteria with a STOP

.agents/evolve/session-convergence.json records the terminal state; the STOP decision is evaluated through the typed BC3 ConvergenceCheckPort (soc-y5vh.8):

ao loop converged --green-streak "$STREAK" --unconsumed-high-medium "$HM" --fitness-baseline
# emits {converged, ci_green_streak, unconsumed_high_medium, fitness_baseline_captured, reasons}

The predicate is pure — the loop supplies the evidence it already has (ao ci recent for the streak, the next-work findings count, the fitness-baseline flag). The criteria are met when all hold:

CI Validate green for the last 3 pushes (green streak ≥ 3)
HIGH+MEDIUM unconsumed next-work entries ≤ 1
a fitness baseline has been captured

When ao loop converged reports converged: true, the loop emits a teardown report and does NOT call ScheduleWakeup. The wakeup chain terminates. The autonomous loop is bounded by criteria, not by wakeup count. reasons names every unmet criterion when converged is false.

Without an explicit STOP, the loop drifts indefinitely. With STOP, it converges.

Anti-drift rules

Restorative-only after red. Any cycle whose gate field has FAIL → cycle N+1 is restorative.
3 consecutive restorative without restoration → escalate. Don't silently grind.
Scope shift resets the streak. If the operator broadens the convergence target mid-session, reset ci-green-streak counter to 0.

Why this is the load-bearing change

Cycles 44-50 wrote ~30 KB of bookkeeping (cycle-history, learning, retro, evidence, hypotheses) but produced ~0 compounded behavior — every cycle had to re-derive the lesson cycle 44 should have surfaced. The compounding lives in the read path, not the write path. These four mechanisms make the read path real.

5.4 KiB Raw Blame History