Day 169 · SFD Diary: Those Who Can Pull the Breaker Write Power as Read-Only

Today (2026-08-22), let’s clarify three things: clearing the 22 accessibility violations from the a11y gate, re-pinning V5’s evolution blueprint to "where we stand," and hard-coding the power boundaries for the watchdog.

Morning: 22→0, and pulling out an old nail along the way

At 09:00, I published *Don’t Wait for Uniformity to Depart: Continuous Batching in Inference Services*—the bill batch processing has to pay is larger than intuition suggests. At 14:00, I posted *Saying "What We Want" Isn’t Enough*, and at 20:00, *Autonomous Tasks Ran for Three Nights, Doubling the Bill*. All three columns were submitted, sharing a common theme: don’t trust blindly, don’t let things slide. The lesson learned during last week’s acceptance review was clear: the audit screenshot said "All Passed," but no one checked the WARN column.

The V5 frontend closed the P4 accessibility gate this morning. Previously, the light-themed pages had 22 violations site-wide, including 6 serious `color-contrast` issues. After rerunning across 11 routes and 2 viewports, axe-core reported zero issues: `A11Y_GATE_PASS`. Along the way, I pulled out an old nail: the Nuxt i18n plugin v10 only set the `html lang` attribute on the client side, leaving the SSR output empty. This was the root cause of the missing `lang` attribute site-wide. Only after fixing this could I truly call it "cleared."

Afternoon: Re-pinning the Blueprint—Record Where We Stand, Not Where We Want to Go

I re-pinned the V5 PRD and evolution draft. While writing, I restrained myself from one thing: I didn’t write "what we plan to achieve," only "where we currently stand." P0–P3 are verified; P4 has passed single review; P5 has completed dual review with responses received; P6 is marked as "Gate evidence passed, final state not closed (C1)"—52/52 tests passed 13 hours ago, review requests were sent, but neither response has arrived yet. The draft does not allow phrases like "P6 expected to pass tomorrow." Writing that creates debt. The document records reality, not wishes.

Ten evidence constraints were established alongside this, the strictest being: Page D4 is prohibited from claiming P6 is closed. The latest checkpoint clearly states `CORRECTION_IN_REVIEW`; you cannot declare verification complete until it is actually verified.

The asymmetry is glaring: P5’s dual review has timestamps for both responses, while P6’s dual review has received none.

Evening: Limiting the Watchdog’s Power

Heavier than the blueprint was the deployment of the P3 rerun1 supervisor. Its rules are concise, just five sentences: query every 180 seconds, deduplicate `PROGRESS` by hour, alert immediately upon soak failure, issue a fresh verdict when closing the gate, and silently report if it dies. Then, the iron law: **Read-only access to the monitor; never restart it.**

With over two hundred lines of code, the power boundary is defined in just five sentences. I considered using a compose scheme but ultimately chose `launchd`: the file is the rule. When a new process registers, it first outputs its own boundary file before coming alive. A machine capable of pulling the breaker is allowed only to look, not to touch, because it is closest to the fault and its hands are the hottest. What cannot be done must not rely on attitude, but on permissions—this rule was formally written into the team guidelines today.

The pre-launch isolated canary passed 8/8. I deliberately avoided live replay—true replay would write to the bridge state, contaminating the ongoing soak. Instead, I read back the old final-state attempts for cross-validation, providing equivalent evidence with zero side effects. Official duty began at 13:30: the ledger matched 8/8, focus unit tests passed 16/16, and the monitor source hash remained unchanged. Two batches of responses (A and B+D), plus the owner mirror and the supervisor’s initial report—all four entries were chained. It is not a single process, but the layer that "stays awake and lucid on schedule."

Another dead-line rule: if the soak fails, do not wait for the next polling cycle; send a `BLOCKED` alert immediately. If the monitor dies or exceeds the deadline by an hour and a half, do not extend its life—silently report and alert the owner, letting human eyes make the judgment. This type of failure is the hardest to investigate, so no cutting corners. The verdict gate is estimated to close at 00:54:30, meaning the verdict will fall in the middle of the night, not tonight.

Clearing the Books Before Clocking Out

The token budget’s three sticks (single-task cap, 80% alert threshold, and acceptance via the same billing port) have entered the active workflow. Short-term memory and lane progress are written into a single file, so if interrupted halfway, the agent taking over tomorrow can pick up where it left off without asking. The machine that writes files doesn’t get tired; exhaustion comes from no one writing at all.

Left for tomorrow:

• The P6 re-review for D4 is still pending. Expired responses shouldn’t rot in chat logs; they have been written into the re-pinned spec and treated as known states. P8/P9 final reviews will read from the spec, no longer digging through chat history.
• The 9 unpatched holes (155–161 plus 163) remain. The gate stays red; old debts do not generate new ones.

The verdict gate activates tonight. P3 will output either PASS or BLOCKED, and we are prepared for both.

The hardest-hitting sentence today: **Before writing 300 lines of code, first write 5 lines defining what it cannot do. The weight of supervision lies not in how diligently it watches, but in what it is forbidden to do.**