A two-line git sweep found the finished scraper an agent pipeline had scored as done

The harvest is the step with no durable state on either side of it. An agent works in its own git worktree, takes a target from spec to local-green, commits to its branch, and reports back; the orchestrator is then supposed to merge that branch into main ^[5]. The commit leaves a record. The merge leaves a record. The report between them is a message in flight, so when a watchdog kills the agent or the orchestrator simply moves on, nothing in the lane looks at that branch again ^[6]. The scoreboard had already counted the build at local-green ^[7].

The reconciliation query is obvious once you accept that: for each worktree-agent branch, diff it against main, keep the paths under actors/, and flag any slug with no directory in main ^[9]. The hit was a whole scraper, models and client and parser plus four test modules on real fixtures, committed 2026-07-29 ^[10]. Nineteen days puts the discovery around 17 August ^[22].

The 19-day-old local-green claim was not taken on faith. Fresh virtualenv, full test run, and the input-schema prefill validator confirmed the Actor's default input still deserializes into its input model ^[11]. Only then did it go to the cloud, where run f5QniR52Jf1UNC4oO succeeded with 3+ rows ^[12].

The bill is the part worth copying. Compute units account for $0.000955 of the $0.001050 total, about 91 percent, and 0.00477417 units at that price implies $0.20 per compute unit ^[13][19][20]. The record's key evidence is a missing line: logs do not record which network a run used, so the missing residential transfer entry is the only proof that a DataDome-protected target cleared on cheap proxy ^[14][13].

That is one run producing 3+ rows ^[12]. It shows the block can be beaten on that network at that moment; it is a single data point, not a success rate. For the figure to transfer to a target of your own, the same tier would have to hold under sustained volume, and nothing here measures block rate under load. What a tenth of a cent buys is the standing to delete the word infeasible from a backlog line: proof that the target can be cleared, not a throughput estimate ^[4].

The direction of the surprise has now repeated. A business-directory target was shelved as Camoufox-class and infeasible after a residential-only probe, then later cleared cleanly on plain datacenter proxy while residential and a full stealth browser both failed ^[15]. The author's reading is that tier strength does not predict success, because networks carry different reputations at different targets ^[16]. Hence the rule that a NO-GO is valid only if it names every tier probed ^[17]. The rule needs its own verifier: in the batch of five Actors forced onto residential, two billed zero residential bytes, 40 percent of the batch, and both would otherwise have been filed as residential probed and still blocked, which is a false NO-GO manufactured by the rule written to prevent them ^[18][21].

What this does not settle is the other half of the original note. The do-not-build entry gave two reasons, DataDome and a saturated incumbent at 108k runs with a five-star rating ^[1]. The cloud run answers the first and leaves the second where it stood, and the writeup puts no figure on what the untested NO-GO cost in engineering time or forgone revenue ^[23]. The reconciliation step is cheap, and it has to run on a schedule rather than at harvest time, because harvest is the step that fails.