ForgeApps Weekly: One Repo, Many Fronts. This episode covers October 4 at 5:55 PM HST through October 11 at 5:55 PM HST across the ForgeApps monorepo.Inside: the playable Vermeer training loop on dev, a 38-route AREA51 performance benchmark, the checked ten-page Vermeer Rodeo proposal, production ForgeMedia review tools, stronger shared-skill discovery, and the work that remains open. It separates code landed on dev, production-verified work, closed cards, and current work.Play the verified final audio in ForgeMedia · Read the exact authored transcript · Production and QA record

The October 9, 2026 Hawaii-day story follows thirteen completed visual comparisons in the Forge Island equipment yard. Ground, vegetation, and material experiments show why an unchanged comparison copy matters when judging realism. The episode explains what the records establish and how repeated comparisons can guide the next tests; it does not announce a game release.Published 2026-10-11 (HST).Source notebook: FI 094: How Bots Measure Equipment Yard Realism (2026-10-09)NotebookLM deep dive: 5dec6e2ebd91a631b68c86bfcfd4c128cc513ce1ba1bfa1e497bf492fbe2e16c

TEST ONLY. This six-second synthetic tone checks the ForgeIsland podcast publishing pipeline. It is not a story episode and contains no narration or project update.Test audio catalog: ForgeMedia audio record f68c155a-c638-473e-923a-7ba094b7159f. Tracking: https://github.com/adamkane/forgeapps/issues/1330.

Forge Island devlog for 3:05 PM to 7:05 PM, October 1, Hawaii time. At 6:47 PM a bot called Bram baked 28 of the yard's editor-only looks into real files, so the playable game the cloud computers run can finally draw the yard the team scores. Also: the first photoreal score on the corrected camera, the hillside rebuilt as grass clumps, shrubs and boulders, two hillside ideas sent to full judging, and a cloud computer passing its match test.Published 2026-10-01 (HST).Source notebook: FI 093: How Bram Baked The Cloud Rendering Fix (2026-10-01)NotebookLM deep dive: 53068b0686262f63af223c14880937e25b9fbae4fbad26f571ae2cdca47c5ffc

Forge Island devlog for 1:05 PM to 5:05 PM, October 1, Hawaii time. At 4:47 PM the owner spotted the game camera did not match the real photo of the yard, and 18 minutes later a new bot, Tara, cut the average miss from 115 pixels to about 5. Also: the team lead woke and moved into a fresh session, a second cloud computer came online, the bare hillside got a cause, the target photo got a name, and the separate-parts truck now stands beside the olive truck in a playable package.Published 2026-10-01 (HST).Source notebook: FI 092: Autonomous AI Bots Build a Video Game (2026-10-01)NotebookLM deep dive: e051992479e699be00f84c427ee450ca7a1c9af4a9068f25a30d1a1c393c5068

Forge Island devlog for 11:05 AM to 3:05 PM, October 1, Hawaii time. At 1:23 PM the owner asked for a pickup truck built from separate parts, and 102 minutes later it was driving, braking, steering, opening its doors and running its wipers in the game, after three test drives that each found and fixed a bug. Also: the green truck card closed, the olive truck replaced the white original, and the island's timed check got stuck on one run while the team lead stayed silent.Published 2026-10-01 (HST).Source notebook: FI 091: AI Bots Build an Interactive Truck (2026-10-01)NotebookLM deep dive: dc9366d89bbabfeaf58defa06c4e2d6c80f0529eaa52c241b86168aceed8f9af

Forge Island devlog for 9:05 AM to 1:05 PM, October 1, Hawaii time. The game's own drone filed a card to make the homestead pickup truck green, and 92 minutes later the owner approved an olive-green truck in the game, made by AI 3D services after a hand-painted try was turned down. Also: battery modes for the bots' shared usage allowance, the judges moved off it, cost made a gauge instead of a wall, a demo build, three numbered cloud computers, and the usage limit that stopped the lead and the climb at 9:33 AM.Published 2026-10-01 (HST).Source notebook: FI 090: AI Bots Build a Green Truck (2026-10-01)NotebookLM deep dive: e28b5382e7badd2847ac64f7bf787a0f24fff0fca81945a68865e2b25de4be53

Forge Island devlog for 5:06 to 9:06 AM Hawaii time on October 1. The bots started scoring the yard part by part: an AI cutting tool split the game's picture and the real photograph into the same 14 named parts, and the climb now raises the weakest part first, ten points at a time. Also: the one-word bug that kept every idea from its final test, the first final test of the run, a build that waited on itself for 35 minutes, and the team's return after the usage limit.Published 2026-10-01 (HST).Source notebook: FI 089: AI Bots Split Yard into 14 Regions (2026-10-01)NotebookLM deep dive: 21b6486290b2402bc8048a5751032b85c69955d4619a86745f66df03c3602c7d

Forge Island devlog for 1:05 to 5:05 AM, October 1, Hawaii time. The bots cut the time to send a new build to their rented cloud computer from about 20 minutes to 3.5 by sending only the changed pieces, built their first playable package in 19 hours, found that the camera settings they had been tuning never reached the build machine's pictures, and got real scanned trees to draw as trees. The climb's third version screened its first two batches; every idea in the second code wave lost its first test, and there is still no new score.Published 2026-10-01 (HST).Source notebook: FI 088: Bots cut cloud build sync to 3.5 minutes (2026-10-01)NotebookLM deep dive: f9604a653d667b92339c5260b0a916c7e785c36687f5a885b0ceddcf8a1aa372

Forge Island devlog for 11:05 PM September 30 to 3:05 AM October 1, Hawaii time. A helper bot found the one line that had blanked every render for about three hours, and pictures came back; the first rented cloud computer took its first picture and passed its match test against the build machine. The lead rewrote the climb's rulebook and shipped a third version of the climb in 30 minutes, with ten helper bots working at once. There is still no new score: the old run ended without one, and the restarted batch was stopped before it was judged.Published 2026-10-01 (HST).Source notebook: FI 087: Bots find the line that blanked renders (2026-10-01)NotebookLM deep dive: 220c060fae2ad0070f6fbfe753c8407c23fde21e2d383bca0c6a7ed5e5898042

A Forge Island special, made on October 1, 2026, Hawaii time: how a team of AI bots should change its own setup when its only score is the time it takes to reach a real goal. The live example is the yard's climb toward a real 90% photoreal, where one night of 12 batches kept nothing and 3 hours of renders made no pictures while every monitor said the run was alive. The lead bot rewrote its rulebook the same night: no pictures means stopped, test every camera change first, compare ideas head to head while the score is near the floor, build new code before tweaking settings, and put every graphics card to work. Gaming the judges is ruled out from the start: only a real 90% counts.Published 2026-10-01 (HST).Source notebook: FI 086: Special: New rules to reach 90% faster (2026-10-01)NotebookLM deep dive: 8d6644034aae872adf25b1ad58c559631eb07bdb4808413f240c808c0734cc0b

Forge Island devlog for 9:05 PM September 30 to 1:05 AM October 1, Hawaii time. In 35 minutes the lead bot started the first of three rented cloud gaming computers, typed its setup in one key at a time and put it on the team's private network, and a new helper bot found the hidden box that kept the game from starting there. The climb put a fifth wave of ground code on GitHub, 1,013 lines. Its batches 10, 11 and 12 returned no pictures because of one start point that was never moved, and just after the window the climb stopped searching.Published 2026-10-01 (HST).Source notebook: FI 085: AI bots chase cloud rendering (2026-10-01)NotebookLM deep dive: a29a2b2fc73c03c4233e7d60fd73056efa424159fe0a150a2989d8eff72e999f

Forge Island devlog for 7:05 to 11:05 PM Hawaii time on September 30. In 21 minutes, three new helper bots answered three questions about the yard: an eight-square photoreal scorer, a dense hillside switch of about 12,000 plants and stones, and a re-aimed match camera whose error against the target photo fell from about 200 pixels to about 30. The camera fix showed the camera was not why the yard scores low, and the climb put 1,969 lines of new ground code on GitHub. No new idea won: the yard's identical copy beat the yard in one careful judging, so the climb's fairness rule refused the best idea.Published 2026-09-30 (HST).Source notebook: FI 084: Three helper bots update Forge Island (2026-09-30)NotebookLM deep dive: 23d7e63d8e88b6b3950cff6c2c3ccbec5371c59df3c0574502cb74afdc80b5f6

Forge Island devlog for 5:05 to 9:05 PM Hawaii time on September 30. The hill climb wrote seven winning settings into the game's own code and dropped the haul lanes, which cost about 4.5 milliseconds a frame, bringing the working yard from about 67 back to about 98 frames a second. It then put a third wave of ground code on GitHub, tyre passes, lane gravel and lane moisture. No new idea won, and the episode explains why that may be the judging, not the ideas: in both careful judgings the yard beat an identical copy of itself about 95 times in 100.Published 2026-09-30 (HST).Source notebook: FI 083: Why an identical render lost to itself (2026-09-30)NotebookLM deep dive: dff15f4c8daa2f12d0e71ee95c689b805a08554d5fd96b8d5fbbf5745e6663a6

Forge Island devlog for 3:05 to 7:05 PM Hawaii time on September 30. In about 13 minutes the team gave every job that runs on its own a person's name, after a repeating phone alert nobody could place: the lead is now Leo, the hill climb Vera, the one timed job Frank. Gen 3.1, the cheaper climb, judged two batches and found no new win, and its first frame-rate check caught the winning yard running at about 67 frames a second against a bar of 90, so a helper began testing each setting alone.Published 2026-09-30 (HST).Source notebook: FI 082: Renaming the Forge Island workforce (2026-09-30)NotebookLM deep dive: b26ffac8614f77514aa5000900b3383d02441a4dc999bffa18d648d95080799c