Five bounded game runs begin
Each run used a lead-owned state ledger, sequential builders, independent critics, evidence receipts, and promotion gates.
Autonomous production study · August 11–12, 2026
Five games entered the Gauntlet.
A one-day experiment in bounded autonomy, independent criticism, deterministic evidence, crash recovery, and knowing when not to call unfinished work complete.
The portfolio
Not five launches. Five truthful checkpoints.
The runs produced major playable progress, but none met the portfolio's strict completion definition before the backend budget stop. Accepted gates remain accepted; candidates remain candidates.
5project gauntlets preserved
17accepted product gates
3B+estimated portfolio tokens
1queued run not launched
Project records
What changed, what held, what is testable.
Each record preserves the starting point, accepted boundary, candidate work, evidence, validation facts, known token floor, failure modes, and exact restart action.
Carpet Front
A repository skeleton became a deterministic, three-mission tactics slice played by molded-plastic soldiers across a room-scale battlefield.
Loafer's Bad Yard Day
A deterministic Patio proof became a warm, household-built tower-defense slice with comic framing, original enemies, tactile feedback, and a reviewable evidence trail.
The Adventures of Foxman: Rotten Run
A protected authored campaign gained a separate, deterministic three-stage roguelite mode without replacing or corrupting the original game.
Hex Quarry
A browser-first voxel architecture became a measured First Excavation playable, while a much larger five-region spell slice was preserved without pretending its missing critic verdict was a pass.
Cock Fight 2000
The First Night gray-box route gained a coherent Combat V2 handoff, exact result and rollback contracts, and a readable product shell. A distinct whole-frame fighter pass remains candidate work.
Portfolio timeline
Build, criticize, repair, recover, stop.
The same operating loop played out differently in each engine. Repository truth and evidence receipts survived the app crash and made the wind-down auditable.
Codex force-quit
Repository state and evidence receipts allowed recovery without restarting accepted work.
Critics force narrow repairs
Visual overlap, route fallbacks, input leakage, browser lifecycle, evidence naming, and device-truth defects were found outside ordinary unit suites.
Backend allowance reaches stop threshold
Marcus reported approximately 6% remaining. Expansion stopped and truthful checkpoints took priority.
Controlled wind-down
All descendants stopped, Git truth reconciled, postmortems written, and accepted work separated from candidates.
Afterglow Circuit stays queued
The prepared marble-platformer gauntlet did not consume the freed seat because no run passed the strict completion test before wind-down.
Run architecture
The Gauntlet was a review loop, not a long prompt.
The lead owned state and promotion. Builders changed bounded leaves. Fresh critics could reject them. Reality checks covered what deterministic suites could not.
01Lead frames leafscope + stop rule
02Builder implementsisolated candidate
03Deterministic gatetests + routes
04Fresh criticPASS / REVISE
05Reality gatebrowser + evidence
06Promote or preservenever blur status
Cross-project postmortem
Where the tokens went.
The expensive work was not primarily feature code. It was proof: fresh critics, browser matrices, evidence refreshes, recovery, and fixing harnesses that had begun testing the wrong thing.
Critics paid for themselves
They found visible overlap, enemies under the floor, hidden automation, false evidence captions, route fallbacks, and input events leaking across scenes. Several defects were invisible to the passing unit suites.
Full matrices were overused
Foxman's final matrix took roughly 15 minutes. Carpet Front repeatedly paid for two-viewport browser proof. The next protocol should require focused gates until a critic says the candidate is materially ready.
Harnesses became product code
Chrome target leaks, key transport, stale audio oracles, mutable evidence provenance, and architecture identity all consumed substantial effort. Once repaired, those harnesses became reusable production infrastructure.
Crash recovery worked
The Codex force-quit did not erase accepted work. Git heads, canonical run-state files, evidence manifests, and critic receipts allowed the portfolio to resume from verified boundaries instead of chat memory.
The budget stop was correct
When the human backend signal reached roughly 6%, expansion stopped. No candidate was promoted by ceremony, no test was weakened, and Afterglow Circuit stayed queued instead of consuming a speculative seat.
Token accounting needs one ledger
Project endpoints exposed lead-side floors, not total descendant or plan consumption. Future runs should record agent, model, gate, retries, wall time, and token deltas in a shared portfolio ledger at every handoff.
What comes next
Resume narrowly, one proof gap at a time.
Carpet needs a corrected oracle and re-critic. Loafer needs real arm64 device truth. Foxman needs a bounded boss leaf. Hex needs provenance reconciliation and a critic. Cock Fight needs a controllable fresh critic. Afterglow remains deliberately queued.