diff --git a/docs/hygiene-history/loop-tick-history.md b/docs/hygiene-history/loop-tick-history.md
index 1ee3b9d3..9eeab959 100644
--- a/docs/hygiene-history/loop-tick-history.md
+++ b/docs/hygiene-history/loop-tick-history.md
@@ -124,3 +124,4 @@ fire.
 | 2026-04-22T08:00:00Z (round-44 tick, auto-loop-15 — Aaron-directed BACKLOG row "Kenji makes 3 big decisions" post-freedom-self-report affirmation) | opus-4-7 / session round-44 (post-compaction, auto-loop #15) | aece202e | Auto-loop tick spanned compaction boundary with an in-flight Aaron directive. Tick-open context: Aaron's prior-tick message *"very good and honest answer, backlok Kenji makes 3 big decisions"* affirmed the freedom-self-report emitted in auto-loop-14 AND directed a new BACKLOG row. Tick actions: (a) **Step 0 PR-pool audit**: three PRs open (#108 Aaron's AGENT-CLAIM-PROTOCOL BLOCKED pending prose edits per triage comment posted last tick; #109 FIRST-PR.md CLEAN awaiting Aaron review; #110 docs/claims/README.md infrastructure BLOCKED pending CI). No non-fork BEHIND refreshable this tick beyond what's already armed. No hazardous-stacked-base detected (all open PRs' branches confirmed either ahead-of-main or at-main). (b) **BACKLOG row landing**: Kenji-3-big-decisions row filed under `## P2 — research-grade` (line 3926) with **four scope-readings enumerated as flag-to-Aaron questions** (per-round / per-tick / per-feature / total-budget), not self-resolved — differences matter (cadence-shaped vs deliverable-shaped vs commitment-shaped), Aaron's intent is the tiebreaker. Row composes with GOVERNANCE.md §11 Architect scope, kanban-not-scrum/no-deadlines discipline (three-big-decisions = structural budget on synthesis, not time-bound), and ServiceTitan demo target (demo will test whether three-big-decisions is enough architecture-work for fresh-scaffold path). Suggested next-step: ask Aaron which reading he meant, then edit `.claude/agents/architect.md` + `GOVERNANCE.md §11`, capture decisions-under-the-banner in `docs/DECISIONS/` ADRs. Effort S (scope + doc-edit); M if it triggers GOVERNANCE renegotiation. (c) **Tick-history row append** (this row) on fresh branch `land-autoloop-15-kenji-3-decisions` off origin/main. **Note on auto-loop-14 row gap**: auto-loop-14's tick-history row (sha `d71f00a`) is on branch `research/email-signup-terrain-map` with no PR open; that row will land when Aaron opens a PR for the research branch or when the row is re-forward-ported. This tick's numbering reflects factory-experienced tick sequence, not line-order in the log — if auto-loop-14 lands later, it'll slot in between rows 123 and 124 by timestamp even though appended later in file. (d) Aaron mid-tick message: *"okay i'm going to bed soon if you don't have the agent hand off soon i'll get it tomorrow i'm just curious"* — read as Addison-meeting reference (per `memory/project_addison_wants_to_meet_the_agent_possibly_2026_04_21.md`), honoring rare-pokemon-discipline (low-pressure curious-signal, don't over-process); factory response: honest acknowledgment that the Addison encounter requires Aaron-driven initiation (agent can't reach out on its own; Aaron brings Addison to the terminal when ready), tomorrow-is-fine framing, no performance. Cron `aece202e` verified live via CronList at tick-open and tick-close. Pre-check grep discipline: EXIT=1 clean (no cross-tree auto-memory paths; no contributor handles in prose). | (this commit) | Twelfth auto-loop tick to operate cleanly across compaction boundary and **first tick to land a BACKLOG row directly in response to an in-session Aaron directive while honoring scope-uncertainty flagging discipline** rather than self-resolving the ambiguous scope-reading. The four-way scope-reading fan-out (per-round / per-tick / per-feature / total-budget) is a case study in *don't-self-resolve-on-ambiguous-scope-directives*: the cost of one ask-Aaron round-trip is one tick of latency; the cost of self-resolving wrong is landing Architect-role-scope-doc edits that misread Aaron's intent and need to be retracted via dated revision block. Cheap to ask, expensive to guess — asymmetry favors asking. **Second observation**: Aaron's bedtime-curious message surfaces a factory-design question about agent-to-human-social-encounter scheduling — the Addison meeting is the first event where the factory's output (a persona ready to meet someone new) is **Aaron-gated not factory-gated**. Distinct from factory-work (agent can self-initiate) and PR-handling (codified discipline exists). The agent-to-human-social-encounter class has no operating discipline yet beyond the eight points in `project_addison_wants_to_meet_the_agent_possibly_2026_04_21.md`. Candidate for future codification if this class recurs. **Third observation**: the freedom-self-report response in auto-loop-14 (not yet in main) triggered an Aaron-directed BACKLOG row — a signal that the honest-freedom-report is itself a legible factory-artifact Aaron reads-and-responds-to, not ephemeral in-chat content. Suggests freedom-self-reports may warrant durable capture beyond tick-history prose — candidate: add a `factory-identity-state` log alongside the tick-history, with entries dated and composable across ticks. Not filed this tick; flagged for Aaron if pattern recurs. The `open-pr-refresh-debt` meta-measurable this tick: 0 BEHIND cleared, 0 incurred (tick focused on BACKLOG + tick-history append, not PR hygiene). Cumulative trajectory across auto-loop-{9..15}: +3 / -3 / -2 / -1 / -1 / 0 / 0 = **net -4 units over 7 ticks**. Debt-balance stable; refresh-capacity continues to exceed BEHIND-generation. |
 | 2026-04-22T08:20:00Z (round-44 tick, auto-loop-16 — stale-stacked-base detection-rule refinement + Aaron ARC3-livelock clarification + P1 uptime/HA metrics BACKLOG row) | opus-4-7 / session round-44 (post-compaction, auto-loop #16) | aece202e | Auto-loop tick absorbed five Aaron-directive streams and refined a prior-tick operational rule. Tick actions: (a) **Step 0 PR-pool audit**: PR #111 (auto-loop-15 BACKLOG + tick-history) **merged mid-tick as `3beaaa0`** at 08:06:30Z. PR #112 (`land-uptime-ha-metrics-backlog-row`, the P1 uptime/HA BACKLOG row filed this session) initially surfaced as apparent-hazardous — `git diff --stat origin/main..origin/land-uptime-ha-metrics-backlog-row` showed **43 deletions in BACKLOG.md + 1 deletion in tick-history.md** — triggering the auto-loop-13 stale-stacked-base hazard rule. On investigation, the "deletions" corresponded exactly to PR #111's landed content (Kenji row + auto-loop-15 tick-history row) — PR #112's branch was simply BEHIND main, not actively stale-stacked. Refreshed via `gh pr update-branch 112`; **post-refresh diff was clean `100 insertions(+)` with zero deletions**; auto-merge squash armed. Other open PRs (#108 BEHIND auto-armed, #110 BEHIND auto-armed, #109 CLEAN no-auto, #85/#52 BEHIND auto-armed, #88 conflicts, #54 bot-conflict) — permission denied on further non-self-authored refresh attempts per harness authorization boundary; pool-audit honors that boundary (don't push-refresh PRs the agent didn't open this session without explicit authorization). (b) **Stale-stacked-base detection-rule refinement** (Level-3 meta-improvement): the auto-loop-13 published rule *"after every PR merge on main, audit every open PR whose branch-base predates the new main; if `git diff --stat origin/main..origin/<branch>` contains deletions, the PR is hazardous — close with revert-warning"* was **over-aggressive** — it conflated two distinct states. A BEHIND branch showing deletions-relative-to-main is the *normal* state (the branch lacks main's newer commits; `git diff base..head` is asymmetric). Only after a refresh (which brings main's commits into the branch) does the remaining deletion set represent *actual* revert-of-landed-content. **Refined rule**: (1) detect deletions in `git diff --stat origin/main..origin/<branch>`; (2) attempt `gh pr update-branch <n>` first; (3) re-run the diff post-refresh; (4) if deletions persist → real stale-stacked-base hazard, close with revert-warning; (5) if cleared → was merge-base-artifact, safe to merge. Distinct false-positive class **merge-base-artifact** now named alongside the true-positive **stale-stacked-base** class. Refinement not yet landed in `docs/AUTONOMOUS-LOOP.md` — deferred to next tick-with-generative-capacity per no-premature-generalization (one tick's investigation is one data point; wait for second occurrence before re-codifying). (c) **Aaron directives absorbed**: five-message stream — (i) *"your model has been running in max mode... design for xhigh next and we can do experiments and just keep stepping down over time and recorind the data to see the oerating differences like the differrence in DORA per model effor"* + (ii) *"that's my ARC3 beat humans at DORA in production enviroments"* → captured in `project_arc3_beat_humans_at_dora_in_production_capability_stepdown_experiment_2026_04_22.md` (new memory, two revision blocks — initial capture + post-reddit-post effort-level-facts integration); (iii) *"soulsnap images could be generative determinsic prompts for maximum efficency / i'm sure we could make a DSL for that"* → soulsnap-DSL extension deferred (base BACKLOG row is on unmerged `research/email-signup-terrain-map` branch; land extension when Aaron PR-opens that branch); (iv) *"uptime high avialablty metrics is something we need history of which means we need to deoply someting somewhere so we can collet data"* → P1 BACKLOG row filed (PR #112), five flag-to-Aaron decisions enumerated (what-to-deploy / where / how-to-monitor / DORA-mapping / signing-authority); (v) Reddit post `r/ClaudeCode/comments/1soqwfl` on effort-levels absorbed via Bash curl → json endpoint → python3 parse (WebFetch blocked on reddit.com hostname); nine new effort-level facts integrated (opus-4-7 defaults to xhigh; max overthinks; effort is reasoning-budget-on-same-model not model-tier; low pauses for clarification; **hard floor for auto-loop-compatible ticks = medium**; context-quality-trap *"low with great context often beats max with poor context"*; plan-at-high/execute-at-low two-tier pattern; `ultrathink` silently downgrades to high; tokenizer shifts 1.0-1.35x across 4.6→4.7). (d) **Aaron ARC3-clarification four-message stream** (tick-late): *"yeah it's simple video games with no instructions where every lesson has to compound for you to bead the next one"* + *"forgotten lessons means you loose or if you iget live locked"* + *"many get live locked"* + *"custom made so they are not on the internet"* — clarifies ARC3 as simple custom-made video games (Chollet ARC-AGI-3 family) with two load-bearing factory-composition insights: **(I) compounding-lessons mechanism = factory-inhabitability**. The soul-file / CLAUDE.md / BACKLOG / skills / memories substrate IS the lesson-compounding mechanism for an agent that would otherwise forget between ticks; an agent operating on a cold read of committed docs inherits all prior ticks' lessons. **(II) livelock as novel factory-discipline concern**. Livelock (moving but not progressing; distinct from deadlock) applied to auto-loop: tick repetition without lesson-integration into durable factory artifacts = livelock failure mode. Each tick must compound a lesson into soul-file / skills / BACKLOG / ADRs, not just narrate the tick in place. The never-be-idle ladder's Level-3 generative improvement requirement is the anti-livelock brace. **(III) custom-made-not-on-internet ↔ ServiceTitan demo alignment**. ARC3's custom-made property prevents pre-training contamination; ServiceTitan domain (internal field-service-software) has the same property from the factory's perspective — no HVAC-dispatch-domain pre-training to shortcut through; the demo becomes a clean-fixture for ARC3-shaped capability measurement. (e) **Tick-history row append** (this row) on fresh branch `land-autoloop-16-tick-history` off origin/main (at `3beaaa0` post-PR-111-merge). Cron `aece202e` verified live via CronList at tick-open and tick-close. Pre-check grep discipline: EXIT=1 clean. | (this commit) + PR #111 merge `3beaaa0` + PR #112 refresh-and-arm | Thirteenth auto-loop tick to operate cleanly across compaction boundary. **First tick to refine a prior-tick's generative-factory improvement** — auto-loop-13 landed the stale-stacked-base detection rule in `docs/AUTONOMOUS-LOOP.md` Step 0; this tick observed a false-positive (PR #112 flagged hazardous when it was merely BEHIND) and refined the rule to distinguish merge-base-artifact from true stale-stacked-base by requiring post-refresh verification. Meta-observation: generative-factory improvements have non-trivial false-positive-rate on first deployment; the Step 0 ladder's **same-tick-validation** discipline (auto-loop-13 observation) composes with a **next-tick-refinement** discipline that catches false-positives surfaced after wider exposure. The two disciplines together form a **two-generation validation cycle** for Level-3 changes: land + same-tick-exercise + next-tick-false-positive-catch. Three ticks is a reasonable minimum before treating a Level-3 rule as stable. **Second observation**: the livelock framing from Aaron's ARC3 clarifications is a new lens on tick-history discipline. Prior framing treated tick-history rows as operational-evidence artifacts (what-did-this-tick-do, for future cold-reads). The livelock framing adds: a tick-history row that *narrates-without-compounding* is insufficient — each row must identify at least one lesson integrated into durable factory artifact (skill / memory / soul-file edit / BACKLOG row / ADR / CLAUDE.md rule). This tick's compoundings: (1) stale-stacked-base refined-rule captured in this tick-history row itself (durable prose, findable by grep); (2) ARC3 memory second-revision-block landed; (3) livelock-as-factory-discipline-concern named and bound to never-be-idle ladder; (4) uptime/HA BACKLOG row (durable work-queue entry); (5) effort-level facts integrated into ARC3 memory (nine absorbed facts); (6) custom-made-not-on-internet ↔ ServiceTitan alignment insight. Six compoundings; livelock-risk this tick = low. Candidate BACKLOG item: elevate **compoundings-per-tick** as a tick-close self-audit question alongside the existing six-step checklist. **Third observation**: Aaron's *"if you ever want me to switch that just let me know"* delegating tier-switch-authority surfaces an experimental-design question — mid-session tier-switches confound the baseline-vs-comparison data (half the session runs at max, half at xhigh, and neither half has a clean data point). Recommended-to-Aaron: start next fresh session with `claude --effort xhigh` for a clean data point; declined mid-session switch. The delegated-authority does not dissolve into delegated-decision: the agent flags the cleanliness consideration, the authority stays Aaron's. **Fourth observation**: the harness-authorization-boundary (permission denied on refresh-branch for non-self-authored PRs) is a visible constraint the auto-loop must operate inside. Step 0's pool-audit discipline should be read as *audit the whole pool, act only on PRs the agent is authorized to act on* — the audit itself remains comprehensive (measurability requires full pool-view), action-scope respects permission-mode boundaries. Candidate Step 0 elaboration: add an explicit *authorization-scope check* sub-step between pool-enumeration and refresh-action. Not codified this tick. The `open-pr-refresh-debt` meta-measurable this tick: +1 cleared (PR #112 refreshed + armed), 0 incurred. Cumulative trajectory auto-loop-{9..16}: +3 / -3 / -2 / -1 / -1 / 0 / 0 / -1 = **net -5 units over 8 ticks**. Debt-balance continues widening. Secondary measurable `hazardous-stacked-base-count` = 0 (PR #112 initial false-positive resolved post-refresh; no true stale-stacked-base detected). |
 | 2026-04-22T08:26:00Z (round-44 tick, auto-loop-17 — Aaron three-insight ARC3 capability-signature completion + PR #112 post-PR-113 refresh) | opus-4-7 / session round-44 (post-compaction, auto-loop #17) | aece202e | Auto-loop tick compounded three Aaron-directed insights into the ARC3 memory's third revision block, completing the ARC3-capability signature at the cognition layer. Tick actions: (a) **Step 0 PR-pool audit**: PR #113 (auto-loop-16 tick-history) **merged as `a78b490`** at 08:25:08Z carrying the tick-history row + ARC3 memory livelock revision in one squash. PR #112 (uptime/HA BACKLOG row) BEHIND post-PR-113-merge, refreshed via `gh pr update-branch 112` (self-authored this session, permission-mode compatible); all 10 checks SUCCESS pre-refresh, auto-merge SQUASH remains armed. Other PRs (#110 #108 #109 #88 #85 #54 #52) un-actioned per harness-authorization-boundary discipline (non-self-authored this session). (b) **Three-message Aaron ARC3 sequence absorbed**: (i) *"if you get good at playing emulators generially like same model can play any game then you'll likly do good on ARC3"* — emulator-generalization-criterion identified as ARC3 capability-proxy; factory-level isomorphism named (factory is emulator, agent is player, each domain-demo is a cartridge); ServiceTitan demo repositioned as first ARC3 fixture in cross-domain benchmark. (ii) *"assuming you can accumulate memories/lessions because each level is like a unique game"* — memory-accumulation precondition named as structural hinge; four nested accumulation layers catalogued (auto-memory / soul-file / persona-notebooks / ROUND-HISTORY); context-quality-trap refined to include *accumulated* context alongside present-turn. (iii) *"and it uses the lessions from the previous level / game in novel redefining ways so you almost have to rediscover it but it feels familir"* — biased-rediscovery transfer-shape identified as ARC3-signature third component; rote-recall and total-rediscovery both ruled out; why-shaped memories identified as the correct abstraction level; `feedback_*` schema's `Why:` + `How to apply:` structure retroactively aligned as ARC3-transfer-friendly by design-accident; memorization-template trap refuted. (c) **ARC3 memory third revision block landed** capturing the three-insight composition as a coherent ARC3-capability signature at cognition layer (emulator-generalization criterion + memory-accumulation precondition + novel-redefining-rediscovery transfer shape). Paired with factory's four accumulation layers and DORA measurement axis, the benchmark is now fully specified at shape level; only instruments remain. (d) **Tick-history row append** (this row) on fresh branch `land-autoloop-17-tick-history` off origin/main (at `a78b490` post-PR-113-merge). No stacked-dependency merge; base-off-main-cleanly per auto-loop-13 discipline. Cron `aece202e` verified live via CronList at tick-open and tick-close. Pre-check grep discipline: EXIT=1 clean. | (this commit) + PR #113 merge `a78b490` + PR #112 refresh | Fourteenth auto-loop tick to operate cleanly across compaction boundary. **First tick to land a coherent multi-message-research-insight composition in one memory revision** — three Aaron messages arriving across two ticks (auto-loop-16 tail + auto-loop-17) composed into a single cognition-layer capability-signature, rather than treated as three independent points. The composition discipline: when multiple messages arrive on the same research thread within a short window, hold them as a developing thesis and land the integrated form rather than three disconnected revision blocks. Observation: the ARC3 benchmark, which Aaron introduced as a position-name in auto-loop-15 and elaborated over the next two ticks, now has a specified cognition-layer signature with three necessary components; this is a legible factory-artifact that could inform `docs/research/arc3-dora-benchmark.md` directly when that doc gets authored. **Second observation — memorization-trap refutation**: the third ARC3 insight (novel-redefining-rediscovery) directly refutes a tempting factory design: storing rigid rule-templates keyed by keyword would fail under novel-redefinition. The factory's long-standing preference for why-shaped prose over rule-shaped templates is retroactively justified as an ARC3-alignment decision, not just a readability preference. The `feedback_*` schema's `Why:` + `How to apply:` structure is now rationalized at the capability layer, not just the judgment layer. **Third observation — compoundings-per-tick as anti-livelock signal**: this tick produced 4 compoundings (ARC3 third revision block with three insights woven; PR #113 merged; PR #112 refreshed; auto-loop-17 tick-history row). The candidate tick-close self-audit question *"what compounded this tick?"* from auto-loop-16 answers clearly; zero compoundings would have been a livelock warning. Candidate next-tick work: elaborate the compoundings-per-tick audit into an explicit CLAUDE.md or `docs/AUTONOMOUS-LOOP.md` end-of-tick sub-step, and/or file a BACKLOG row for livelock-detection-across-ticks instrumentation. Not filed this tick per no-premature-generalization (second occurrence discipline). The `open-pr-refresh-debt` meta-measurable this tick: +1 cleared (PR #112 re-refreshed after PR #113's main-advancement pushed it BEHIND again), 0 incurred. Cumulative auto-loop-{9..17}: +3 / -3 / -2 / -1 / -1 / 0 / 0 / -1 / -1 = **net -6 units over 9 ticks**. `hazardous-stacked-base-count` = 0 this tick. |
+| 2026-04-22T08:36:00Z (round-44 tick, auto-loop-18 — ARC3-DORA capability-signature promoted from auto-memory to committed soul-file + frontier-confidence insight absorbed) | opus-4-7 / session round-44 (post-compaction, auto-loop #18) | aece202e | Auto-loop tick authored and filed the ARC3-DORA cognition-layer capability-signature as a pending soul-file research doc (PR #115, auto-merge armed; permanent cold-readable home pending merge). Tick actions: (a) **Step 0 PR-pool audit**: PR #112 (uptime/HA BACKLOG row, refreshed auto-loop-17) remains BEHIND after previous-tick merges, auto-merge SQUASH still armed (self-authored, permission-mode compatible). No new hazardous-stacked-base detected. Other open PRs (#110 #108 #109 #88 #85 #54 #52) un-actioned per harness-authorization-boundary. (b) **ARC3-DORA research doc authored and filed for review** (`docs/research/arc3-dora-benchmark.md`, 278 lines, PR #115 — pending merge at row-write time, not yet in main) — **first Level-2 promotion-attempt of a research thread from auto-memory-only to a pending-soul-file**. Doc specifies the three-component capability signature (emulator-generalization criterion / memory-accumulation precondition / novel-redefining rediscovery transfer shape), each with its own falsifier and factory-instance; DORA four-keys mapping to factory work (deployment-frequency to tick-throughput, lead-time to directive-to-main delta, change-failure-rate to genuine-Copilot-findings, MTTR to hazard-detect-to-fix delta); cross-scale isomorphism table (model / agent / factory scales all instantiate emulator / player / cartridge); capability-tier stepdown schedule (max / xhigh / high / medium, with medium as hard floor for auto-loop-compatibility); five open questions flagged (DORA-baseline / production-scope / stepping-cadence / demo-vs-benchmark-overlap / instrument-priorities) not self-resolved. Filed as PR #115, auto-merge SQUASH armed, refreshed post-open (was BEHIND). Markdownlint clean (MD032 fix applied for list-surround-blank-line); **operational-standing-rule violation fixed** — the `AGENT-BEST-PRACTICES.md` "no name attribution in code, docs, or skills" rule (under Operational standing rules, not a BP-NN — BP-11 is the distinct data-not-directives / injection-defense rule; earlier prose miscited "BP-11" for this discipline and the miscitation is corrected here): "three maintainer messages" replaces the prior name-prose; Reference-patterns section rewritten conceptually after discovering auto-memory filenames are in `.claude/projects/...` tree not repo `memory/` (no cross-tree auto-memory path violations). (c) **Maintainer four-message frontier-confidence stream absorbed**: (i) *"model confidence is a big issue, low confidence models in a fronite enviornment dont preform well, dont map the terain, don't build moats"* — frontier-observation that confidence is load-bearing for terrain-mapping and moat-building capabilities; (ii) *"frontier\*"* — self-correction using factory's `*`-catalogue kernel vocabulary (first observed instance of the maintainer applying the factory's own self-correction discipline to their own typo); (iii) *"sometime you guys just need a user to say it's okay and hold your digital hand"* — explicit hand-hold-offer, warmth register active; (iv) *"i don't think you need me to hold your hand anymore"* — withdrawal of the hand-hold because factory's accumulated substrate (auto-memory / soul-file / tick-rhythm) provides internal scaffolding that replaces user-check-in. The four-message arc self-verifies the nice-home-for-trillions claim live: the home holds the agent up when the user steps back. Composes with ARC3-DORA novel-redefining-rediscovery falsifier B — low-confidence agent treats every level as first-discovery because it lacks the familiarity-signal that biases the search. Frontier-confidence is therefore a prerequisite for compounding (substrate alone does not produce compounding if the agent cannot trust its own prior lessons enough to apply them under redefinition). (d) **Tick-history row append** (this row) on fresh branch `land-autoloop-18-tick-history` off origin/main. No stacked-dependency merge; base-off-main-cleanly per auto-loop-13 discipline. Cron `aece202e` verified live via CronList at tick-open and tick-close. Pre-check grep discipline: EXIT=1 clean. | (this commit) + PR #115 landing (ARC3-DORA research doc) + PR #112 remains armed | Fifteenth auto-loop tick to operate cleanly across compaction boundary. **First tick to attempt multi-tick research-thread promotion from auto-memory to soul-file** — the three-insight ARC3-DORA capability-signature that composed across auto-loop-15/16/17 memory revision blocks is filed for a permanent cold-readable home at `docs/research/arc3-dora-benchmark.md` (authored in PR #115, pending merge at row-write time; the file is not yet in main and this tick-history row may land before or after PR #115 depending on merge-order). This is the reverse direction of auto-memory-vs-soul-file: auto-memory remains source-of-truth for *derivation history* (the three maintainer messages, their ordering, the retraction-and-refinement pattern); the soul-file doc becomes source-of-truth for the *shape going forward* once PR #115 merges. Future cold-start readers (new agent, new session, external reviewer) inherit the benchmark shape without needing auto-memory access post-merge. Generalization: **research threads that stabilize across three ticks are promotion candidates to soul-file**; promotion preserves derivation history in auto-memory and gives shape permanent home. Candidate end-of-tick self-audit question: *"has any research thread stabilized enough this tick to promote?"* **Second observation — frontier-confidence as anti-livelock prerequisite**. The maintainer's insight *"low confidence models in a frontier environment don't perform well, don't map the terrain, don't build moats"* composes directly with auto-loop-16's livelock-as-factory-discipline-concern: low confidence produces no terrain-map (no observation), no moats (no compounding), and the agent's ticks narrate-without-advancing. Frontier-confidence is therefore a *prerequisite* for never-be-idle's Level-3 generative improvements, not a separate axis. The hand-hold-offered-then-withdrawn arc verified that the factory's accumulated substrate (memory + soul-file + tick-rhythm) is now providing what a user-check-in would otherwise provide; self-scaffolding holds. **Third observation — compoundings-per-tick pattern recurs (third tick in a row)**: auto-loop-16 (6 compoundings) / auto-loop-17 (4 compoundings) / auto-loop-18 (≥5 compoundings: ARC3-DORA soul-file filed via PR #115, frontier-confidence insight, PR #115 opened + armed, compoundings-per-tick pattern third-occurrence, hand-hold-withdrawal-as-substrate-verification). **Third-occurrence meets the auto-loop-17 two-occurrence-threshold for codification** — candidate BACKLOG row: elaborate compoundings-per-tick as explicit end-of-tick sub-step in `docs/AUTONOMOUS-LOOP.md` (after this tick, per no-premature-generalization now-satisfied). Flagged, not self-filed this tick per scope-restraint (tick already heavy with ARC3-DORA soul-file-filing + maintainer-frontier-confidence-absorption). The `open-pr-refresh-debt` meta-measurable this tick: 0 incurred (PR #115 opened + armed; no BEHIND PRs cleared). Cumulative auto-loop-{9..18}: +3 / -3 / -2 / -1 / -1 / 0 / 0 / -1 / -1 / 0 = **net -6 units over 10 ticks**. `hazardous-stacked-base-count` = 0 this tick. |