Adversarial review · 12 September 2026 · Tomorrow: Sunday 13 September
Pursue Minecraft tomorrow, with the early gates enforced
Minecraft revision 2 is the stronger five-hour choice on the evidence available. Agent Arena offers concrete character, console, camera and provider reuse candidates, and compatible files exist on the Laptop. That gives more time to finish the new power choices, rewind and comparison. Unreal revision 8 has greater visual freedom and a cleaner eventual Windows handoff, but its working toolchain, representative package and compatible art are not established.
Independent final review. A fresh reviewer read both complete revised pages and their full execution briefs, the readiness inventory and arithmetic check. I formed this comparison independently of the specialist recommendation sections, then checked the final consistency report. I did not inspect the private source again or execute either game.
8Specialist audits
84Findings, including overlaps
2Plans rewritten
1Fresh comparison agent
Open revised Minecraft plan · Open revised Unreal plan
This is a planning audit. Source and documentation checks, arithmetic analysis, read-only machine inventory and webpage validation were performed. Neither game was built or launched; no provider latency, frame-rate or delivery benchmark is claimed.
The choice for tomorrow
Choose Minecraft revision 2 for Sunday, subject to permitted reuse, a working pinned source build and an accepted legitimate judge route. Its existing visual and interaction foundation gives the strongest evidenced starting point for a finished playable loop. Keep the independent court and capped scenery attempt. A missed startup gate does not make an untested Unreal project a safe replacement.
| Criterion | Minecraft | Unreal |
|---|---|---|
| Existing work and new integration | Bodies, console, lifecycle, cameras and provider transport are inspected reuse candidates. The reducer, FORK adapter, branch restoration and comparison remain new. The installed JAR has not been matched to the reviewed source or launched. | The same game rules and provider behavior must be built, alongside the scene, controls, actor projection and packaged runtime integration. No working Unreal project or named compatible asset set is evidenced. |
| Visual quality within five hours | An existing coherent block style, bodies and camera controls provide a concrete starting point. A small readable court can be finished without new shaders. It will look like stylised Minecraft Singapore. | Offers more control over materials, lighting and presentation, but the revision allows simple bodies and existing assets only. Its higher visual ceiling does not establish a better finished scene by tomorrow's deadline. |
| Playable choices and inhabitants | The six-round allocation and reroute loop is real player agency if implemented. Existing transport helps, but useful live FORK actions still need proof. Three restricted roles are a small simulation, not open-ended autonomous city life. | Uses the same arithmetic-checked dilemma and equally restricted roles. Engine choice adds no gameplay depth by itself. Live provider wiring is additional work, and full-run latency remains unknown. |
| Rewind correctness and destruction | Harder restore boundary: exact cuboid, actors, server jobs and durable role memory. Reused upstream behavior must be fenced. Demolition is a fixed decorative mask outside scored play, not general destruction. | The simpler restore design is an advantage: fixed logical state and actor positions, with no block world or destruction to restore. Provider memory, delayed callbacks and immutable A results still need verification. |
| Singapore data dependency | Arnis adds conversion and generation uncertainty, but a timed attempt cannot block the independent court. Neither an accepted Arnis world nor its compatibility is proven. The fallback must carry an honest stylised label. | Fewer dependencies: static locator and a small street reference, with no required regional importer or streaming. This is the stronger data plan, although both now protect the core game from map failure. |
| Packaging and judge access | More recipient friction: legitimate client access, matching Java/Fabric/mods, one coordinator and runtime access. The actual judge route and public screening conditions need resolution. A mod ZIP alone is not delivery. | An accepted standalone Windows package would remove the Minecraft client requirement. It still needs a clean download and launch without the editor, a usable runtime endpoint and host acceptance. None has passed. |
| Machines, film and the delivery deadline | The Laptop has an observed installation head start; Desktop readiness is unknown. Generation, runtime changes and capture share one lease. Client-account switching costs operator time. Reused cameras may help protect the full A/B film and edit reservation. | Editor, imports, builds and capture share the unproven Desktop route. More agents cannot parallelise that queue or replace the human operator. Both plans budget the full delivery chain correctly, but no build, gameplay, capture or upload duration is measured. |
What would change the decision
- Prefer Unreal before committing to the build if permitted readiness checks reveal a pinned working Unreal project, compatible art, representative recipient package and capture route, while Minecraft source reproduction or coordinator integration fails. An installed engine or stronger GPU alone is insufficient.
- Prefer a proven Unreal route if the accepted evaluation or screening arrangement cannot support Minecraft's legitimate client and applicable usage conditions, while Windows delivery is accepted. Host acceptance alone does not establish third-party permission.
- If hosts disallow the Agent Arena reuse on which this recommendation depends, reassess before execution. Minecraft loses its main advantage; this does not establish that a fresh Unreal build will fit.
- If neither route passes readiness, neither full five-hour promise is justified. Missing the 11:10 startup gate allows only the stated identified repair through 11:25. Then agree a smaller deliverable and preserve submission time. Do not start an untested alternative engine.
- If live inhabitants are mandatory for Lucas, either route must pass useful live actions and the complete live A/rewind/B within measured budgets. A fixture reduction requires his explicit choice; it does not silently satisfy that preference.
The proof needed before calling the build ready
At the first host-permitted readiness check
Pin the actual source and Java 25/Gradle/Node/Fabric/Carpet versions. Reproduce a dedicated install without altering the original. Verify one coordinator, actual bodies, console, camera controls and approved provider access. Compare the built artifact with its source manifest. Record the legitimate judge route and allowed reuse; do not treat files on disk as a passing launch.
11:10, with only an identified repair until 11:25
A recipient launches or joins through the selected delivery topology and performs one new fixture FORK action through the actual adapter, observing the committed scenario change. Existing character movement alone fails this gate. Probe a useful live FORK proposal separately and record real total request time.
Noon to 12:30
Complete canonical A/rewind/B, then the entire chosen runtime mode. Verify three exact cuboid and actor restores, outside sentinels, resource conservation, delayed-work fencing and absence of an A-only fact/receipt in B after restore and restart. A human explains choice and consequence after the short introduction. Record full-run waits, cancellation and recipient performance.
Before protecting the release baseline
Measure the complete recipient and capture route, including client switching, one retake and verified media transfer. Reserve the Laptop edit/export time. Keep a named clean-start package/world/mode with matching real footage; a launch shell is not a fallback. Recompute B/C/E/U deadlines from evidence and retain the accepted release if a later chain will not fit.
This is an evidence-based preference, not a success probability. No game was built or launched, no provider call or full-run timing passed, and the proposed balance was checked only arithmetically. Default Java 21 does not prove Java 25 is unavailable elsewhere. Desktop SSH timeouts establish neither its power state nor installed software. Minecraft source reuse has not been reproduced; Unreal readiness has not been disproven. Host reuse, recipient access, screening conditions and runtime availability through judging remain open. Neither revised plan delivers realistic island-scale Singapore, unrestricted inhabitants or general destruction.
What changed in both plans
| Area | Problem found | Revised approach |
|---|---|---|
| The central choice | The old two-plus-two charge budget could cover all three repair rounds without clinic downtime. | Both plans now use two one-charge clinic-only batteries, fixed phase order and explicit reference runs. |
| Unreal scope | Streaming Singapore, data conversion, character/material work and a new runtime service competed for one production queue. | A fixed local court and an attributed static Singapore locator are the baseline. Existing compatible art and simple lighting replace new asset pipelines. |
| Minecraft scope | Generic editing, TNT containment and survival-state rewind created new physics and persistence work. | One initial checkpoint, an exact small solid/air cuboid and a controlled facade-demolition effect. General world-history rewind and native TNT are outside tomorrow's baseline. |
| Erased knowledge | A fresh provider conversation could still read old local memories, receipts or queued actions. | Branch context/storage is isolated; authoritative epoch checks fence delayed writes. Comparison archives stay outside role context. |
| Agent organisation | Nine opening sessions fed a single overloaded Main and could share a Git index. | Start with four hands-on owners. Expand only for independent packets, one committing writer per worktree, explicit outboxes and real task IDs. |
| Data claims | Generator defaults, coordinate conventions and private-only credits hid extra work and dependencies. | One explicitly pinned source path, an external wall-clock limit, a generated-world reopen check and public-safe release credits. |
| Delivery | A three-second capture and a fixed upload window did not measure the real demonstration. | Time a full accepted-mode A/restore/B rehearsal, process switches, retake and transfers. Video upload, processing and full hosted playback now affect all earlier deadlines. |
| Fallback honesty | A bootable shell or fixture build could be called a protected release after live scope changed. | Only a named accepted package/world/mode with matching footage qualifies. An engine switch is not an automatic response to a late failure. |
The power choice now has a cost
These are arithmetic-checked reference runs, not playtest results. Both start with one charge in the clinic and one spare charge, with three powered workshop repairs needed. Grid power begins the round after repair completes.
| After six rounds | A: keep clinic powered | B: power workshop first |
|---|---|---|
| Clinic downtime | 0 rounds | 1 round, round 3 |
| Repair progress | 0 of 3 | 3 of 3 |
| Grid active | Not during this episode | From round 4 |
| Remaining charge | 2 | 0 |
These outcomes assume the courier delivers by round 2 and the engineer performs the permitted repairs. Live model choices may differ; the game must show what actually happened.
The delivery budget includes the upload
Use measured durations. B covers the package and actual judge-style access check; C covers the whole capture reservation, process switches and retake; E covers remaining edit/export/local playback; U covers video upload, processing and complete hosted playback with margin.
X = min(14:20, 14:40 - U)M = min(14:05, X - E)S = min(13:35, 14:00 - B, M - B - C)
The export finishes by X, final footage reaches the editor by M, and the serial release path starts by S. Earlier work cutoffs move forward if those durations demand it. A short codec sample is not a full-run measurement. Submission acknowledgement remains targeted for 15:10.
What we actually know
Observed
- The eight reviewers examined both baseline plans; the Minecraft reviewer also inspected the existing private source.
- The Laptop accepted read-only SSH queries. Existing Agent Arena and matching Minecraft mod files were found.
- Java, Node, Git and FFmpeg were present on the Laptop's command path.
- Both configured Desktop SSH routes timed out during this audit.
- The original charge balance and proposed replacement were analysed across twelve allocation schedules.
Still unverified
- A clean Agent Arena launch from the reviewed source, live provider access and Java 25 runtime readiness. Java on the inspected Laptop path was 21.
- The Desktop's current toolchain, actual GPU model and capture session. An SSH timeout does not prove it is off.
- A representative Unreal package running on Laptop. No engine was found in the limited common paths inspected there; other paths were not ruled out.
- Generation/import, complete live A/B, rendering, restoration and release timings.
- Host acceptance of reused code and the intended mod/executable delivery and screening route.
Lucas will ask the hosts later. The planning default is disclosed reuse, new FORK implementation during the event window, and a real runnable prototype with clear access requirements. Those assumptions have not been presented as confirmed rules.
The public event listing requires a working prototype and 90-second video. Arnis v3.2.0 has more optional sources/features than the old brief described. Natural Earth supplies a public-domain locator alternative. The revised plans also link the applicable Minecraft guidelines, Cesium terms and engine/tool documentation.
Who tried to break the plans
| Review role | Focus |
|---|---|
| Unreal production skeptic | Cold toolchain, map imports, shader/build queues, binary assets and runtime API work. |
| Minecraft production skeptic | Actual reuse boundaries, physical actions, durable memory, reset scope and package contents. |
| Gameplay critic | Twelve allocation schedules, action semantics, visible consequences and the ninety-second story. |
| Geodata fact checker | Current Arnis sources/settings, Singapore coordinates, static locator and public data credits. |
| Rules and provider fact checker | Event claims, model/account assumptions, distribution, provider terms and judging access. |
| Agent operations reviewer | Actual task interfaces, writer/worktree isolation, Main bottlenecks and reconnect behaviour. |
| Cinematic and delivery producer | Full-take timing, human reservations, package access and video processing deadlines. |
| Adversarial QA | Contradictory prompts, stale writes, atomic rounds, false fallback assumptions and focused acceptance. |
A separate agent compared the revised versions after the edits and final cross-check. Its recommendation is shown above. This is expert judgment from the available evidence, not a measured completion probability.
Every finding has a disposition
All 84 reported findings, including overlaps, were checked against both rewritten plans and their execution prompts. Relevant changes are recorded below. Host decisions and runtime proof remain open where stated; a written correction is not a passed implementation test.
Open all 84 findings and their outcomes
| ID | Finding | Status | How it was handled |
|---|---|---|---|
U01 | Required island streaming overloaded the build | Plan revised | Unreal now defaults to a local court and a static locator; streamed Singapore is removed from required work. |
U02 | Cold Unreal readiness was assumed | Runtime evidence still required | Selection depends on a real pinned-toolchain and packaged-launch check. Installed hardware alone is not readiness. |
U03 | Geographic origin changes could move saved actors | Plan revised | The baseline uses fixed local coordinates. A future georeferenced variant needs an explicit origin/anchor strategy. |
U04 | A new runtime backend was left on Main's critical path | Plan revised | Use an approved tested route and a narrow Gameplay-owned adapter; settle live mode before baseline capture. |
U05 | Runtime language and binary-authoring choices were open | Plan revised | Pin a small C++ value-state core and a simple map with minimal existing presentation assets. |
U06 | Rigs, generated materials and extra polish could trigger late imports | Plan revised | Use existing compatible characters/materials and remove new texture, rigging and optional-feature work from tomorrow's baseline. |
U07 | The cinematic depended on unproved actor bindings | Plan revised | Use actual local gameplay with a short opening; any camera sequence must already work in the accepted package. |
U08 | Cooked assets and build configuration were not fixed | Plan revised | Pin one map/configuration from the first package and verify required runtime references in the actual release. |
U09 | Rendering choices preceded Laptop evidence | Plan revised | Choose conservative rendering from package one; record measured settings instead of inferring FPS from hardware. |
U10 | Three seconds did not measure a full demonstration | Plan revised | A full accepted-mode A/restore/B rehearsal feeds capture and editing budgets. |
U11 | Opening staffing created an integration queue | Plan revised | Start Main and three hands-on leads; expand only for ready, independent work. |
MC-S01 | Existing reset did not prove a bounded cuboid restore | Plan revised | Use an exact small cuboid including air, with outside sentinel checks and one known initial checkpoint. |
MC-S02 | Fresh conversations could retrieve erased durable memory | Plan revised | Isolate branch-owned sessions, notebooks, observations, receipts and context; retain comparison archives outside role retrieval. |
MC-S03 | Physical action failure could not roll back logical resources | Plan revised | Commit one authoritative logical round; movement and props present committed state and cannot duplicate resource effects. |
MC-S04 | Existing movement did not prove the new FORK adapter | Plan revised | The first integration gate exercises a real new scenario action through validation, commitment and display. |
MC-S05 | Contained TNT was a custom physics subsystem | Plan revised | Use a predetermined facade-removal effect with no live explosive entity or entity damage. |
MC-S06 | General survival restoration inflated rewind scope | Plan revised | Use a fixed roster, plain-block checkpoint and documented role anchors; arbitrary survival/world-history rewind is excluded. |
MC-S07 | Gameplay lacked ownership of actual integration files | Plan revised | Assign a concrete upstream adapter/provider file lease to the Gameplay integration owner before dispatch. |
MC-S08 | Generator defaults and world conversion were under-specified | Plan revised | Pin a minimal generator preset and verify a disposable sample through save, stop and reopen on the chosen stack. |
MC-S09 | Existing installers and coordinators could conflict | Plan revised | Use a dedicated FORK profile and one explicit launch topology with exactly one active runtime authority. |
MC-S10 | Map generation blocked scenario integration | Plan revised | Fix the court coordinates early and make geography a replaceable surround, not a prerequisite for gameplay. |
MC-S11 | One live action did not prove usable episode latency | Runtime evidence still required | Measure the full chosen-mode episode and bound provider waits/retries before accepting a capture baseline. |
MC-S12 | Actual package contents differed from assumed minimal reuse | Plan revised | Inspect the release allowlist and notices; exclude optional voice dependencies, unrelated maps and account files. |
G01 | The original battery budget erased the central dilemma | Plan revised | Use two one-charge clinic-only batteries and explicit six-round A/B reference outcomes. |
G02 | Final repair count hid when grid power became useful | Plan revised | Show the clinic's six-round service timeline and grid activation round alongside repair and charge totals. |
G03 | Battery sockets, ownership and source precedence were ambiguous | Plan revised | Freeze immutable battery IDs, clinic-only sockets and grid/generator/battery priority with consumption only on use. |
G04 | General actions and rotating order made outcomes ambiguous | Plan revised | Use three tiny role action sets, fixed phases and one atomic logical round with bounded visual presentation. |
G05 | Waits and failures could stall the demonstration indefinitely | Plan revised | Committed waits advance rounds; infrastructure failure does not. Provider attempts and retries are visibly bounded. |
G06 | The Minecraft rewind promise covered too much state | Plan revised | Limit to one known initial checkpoint and owned scenario/plain-block/role state with isolated branch context. |
G07 | The separate TNT demonstration consumed core development time | Plan revised | Use a short paused controlled-demolition demonstration, followed by restore before scored play. |
G08 | Three avatars did not establish meaningful live AI | Plan revised | Record actual role proposals, accepted effects and mode; explain each role's function without claiming unproved autonomy. |
G09 | The film gave too little time to the main comparison | Plan revised | Capture complete runs, then clearly time-compress genuine waiting while retaining a legible comparison. |
G10 | Singapore scope exceeded its demonstrated gameplay role | Plan revised | Promise one small Singapore-referenced scene and a fictional crisis, with a static locator; real mapped surroundings remain bounded and conditional. |
G11 | The Minecraft pitch could mostly showcase prior work | Plan revised | Pin the upstream base and old/new ledger; identify new scenario, restoration and comparison work in the film. |
G12 | No early check established that the choice was understandable | Plan revised | At noon, a short human comprehension check follows the reference-run validation and controls remaining polish. |
GEO-01 | Unresolved Cesium eligibility remained on the default path | Plan revised | Use the local open-data/static-locator baseline; an unverified hosted-data route does not count as ready. |
GEO-02 | Current Arnis defaults added unbudgeted sources and assets | Plan revised | Freeze explicit minimal settings, actual command/options and every remaining source in the output manifest. |
GEO-03 | Arnis's timeout did not bound the full job | Plan revised | Enforce an external acquisition-through-acceptance deadline with reserved conversion/check time. |
GEO-04 | Broad Java support was not a conversion test | Plan revised | Generate a tiny sample, upgrade only a disposable copy, save/reopen and freeze the accepted output/dependencies. |
GEO-05 | Coordinate conventions could disagree across workers | Plan revised | Use named geographic fields, explicit units/axes/origin and a documented round-trip tolerance. |
GEO-06 | The island locator fallback had no concrete source | Plan revised | Use a pinned generalized Natural Earth Singapore locator or another already verified permitted source. |
GEO-07 | Private source manifests were insufficient for public recipients | Plan revised | Release public-safe credits and fulfil any applicable data offer without publishing private gameplay source. |
GEO-08 | The Unreal source catalogue disguised an extra import programme | Plan revised | Select one local source and one locator; remove mandatory regional downloads and secondary-data exploration. |
GEO-09 | Cesium video branding requirements were incomplete | Plan revised | The default no longer uses Cesium; any separately accepted future variant must satisfy its actual logo and source-credit rules. |
GEO-10 | Higgsfield's Minecraft integration mismatched the pinned loader | Plan revised | No Higgsfield Minecraft mod or loader migration is included; tomorrow's baseline uses existing assets. |
RF01 | Host reuse and delivery acceptance were unresolved | Open input; plan revised | Retain conservative planning assumptions and obtain actual host wording later; no approval is asserted. |
RF02 | Live runtime AI was conflated with an event requirement | Plan revised | Distinguish the user's desired live inhabitants from the listing's requirement to use Astra during development. |
RF03 | Runtime access after submission was not established | Open input; plan revised | Record the judging-access period and approved runtime route; keep only an explicitly accepted labelled fallback. |
RF04 | A Minecraft profile ZIP could bundle impermissible content | Plan revised | Ship permitted authored mods/world and notices with legitimate client requirements, excluding game binaries and account state. |
RF05 | Minecraft promotion and public-screening applicability were unresolved | Open input; plan revised | Use separate title/credit cards and neutral roles; keep actual event-use permissions as an open question, not an automatic ban or approval. |
RF06 | Keeping tiles out of prompts did not settle Cesium terms | Plan revised | Remove hosted Cesium from the default build unless the actual applicable agreement is already confirmed. |
RF07 | Development and runtime were assumed to share one limit pool | Plan revised | Record each actual authentication/billing route, supported settings and limits; reduce shared load only when the pool is shared. |
RF08 | An optional integration could force a Minecraft version change | Plan revised | Keep the pinned loader/game stack and exclude Higgsfield's incompatible in-game integration. |
RF09 | Data attribution was not an explicit release artifact | Plan revised | Include visible map credits plus a public-safe notices/data-information artifact with the actual release. |
OPS-01 | Peer Codex task creation was assumed available | Plan revised | Use a callable peer-task API only when available; otherwise report real named lead-agent IDs or human-launch briefs. |
OPS-02 | Main owned too many implementation dependencies | Plan revised | Gameplay and World receive complete scoped deliverables; Main retains contracts, short shared setup, acceptance and release authority. |
OPS-03 | Concurrent leaf commits could share an index and HEAD | Plan revised | Every committing writer gets its own branch/worktree and one Git mutation owner. |
OPS-04 | Nine opening sessions were treated as a useful default | Plan revised | Begin with four hands-on owners; the initial useful expansion has two leaves and one reviewer, within actual caps. |
OPS-05 | Workers lacked clear status paths and escalation | Plan revised | Assign per-team outboxes, five-minute/two-attempt blocker reporting and one pending immutable delivery per team. |
OPS-06 | SSH process persistence was assumed | Runtime evidence still required | Use a proved local runner or a bounded disconnect/reconnect check; otherwise humans own long runtime/capture processes. |
OPS-07 | Capture omitted stop/switch/start overhead | Plan revised | C includes process transitions, profile/world readiness, full takes, transfer and any required restart. |
OPS-08 | A result hash did not guarantee artifact receipt | Plan revised | Main acknowledges only locally resolvable commits and verified artifacts, using the named transport or a verified bundle fallback. |
OPS-09 | Long duplicated briefs could drift during execution | Plan revised | Shorten obsolete material and pin plan/brief and contract hashes; first worker packets contain only relevant concrete work. |
DP-01 | Video upload and processing were outside the deadline formula | Plan revised | Add U and export deadline X so hosted playback time moves capture and build cutoffs earlier. |
DP-02 | Short sample timing understated real capture work | Plan revised | Keep a codec smoke check but budget from full gameplay, reset, setup and retake measurements. |
DP-03 | Fallback mode could fall below the chosen product minimum | Plan revised | Qualify a baseline only in an explicitly accepted mode, with a completed core route and matching footage. |
DP-04 | Package timing omitted the actual judge setup route | Plan revised | B includes download, clean install/prerequisites, the chosen-mode A/B route and recipient access. |
DP-05 | Narration was left until final export time | Plan revised | Freeze a short timed script and record an accepted early voice take with small alternate lines. |
DP-06 | The distinctive fork appeared too late in the film | Plan revised | Shorten the opening and give the main A/rewind/B comparison the largest share of the ninety seconds. |
DP-07 | Camera source identity did not prove installed playback | Plan revised | Test the camera bundle in the exact accepted package/profile and identify staging honestly. |
DP-08 | The emergency repair cap could be misread as extra time | Plan revised | Accept a repair only when the entire recheck/repackage/recapture/audio chain still meets recalculated media deadlines. |
DP-09 | Two humans were scheduled as if always available | Plan revised | Reserve player/capture and narration/edit roles; remove secondary presentation work if only one person is available. |
DP-10 | A failed launch could trigger an untested engine switch | Plan revised | Select a route early; use a smaller approved version of it unless the alternate independently passed readiness. |
QA-01 | Independent implementations could choose different power rules | Plan revised | Use one canonical initial-state/rule table and two exact reference traces with conservation checks. |
QA-02 | Role actions and movement costs were not fixed | Plan revised | Pin role anchors and tiny action sets; logical delivery is one committed action with clearly labelled physical presentation. |
QA-03 | Per-action revisions could reject later actors in the same round | Plan revised | Collect against one frozen round state, stage all intents and commit one new revision. |
QA-04 | Lost completion acknowledgements could duplicate effects | Plan revised | Use stable action/round IDs and authoritative exactly-once logical effects; presentation cannot spend resources. |
QA-05 | Unreal provider history could retain the erased future | Plan revised | Use stateless requests or fresh branch-scoped sessions and inspect actual serialized context. |
QA-06 | Invalid intent and infrastructure failure had conflicting effects | Plan revised | Specify an outcome table for wait, illegal intent, malformed batch and transport timeout. |
QA-07 | Cancelled model work could still mutate the restored world | Plan revised | Fence delayed mutations at the server/controller authority and retire all old-epoch writers before resume. |
QA-08 | A world directory alone was not a complete checkpoint | Plan revised | Bundle world, scenario and context seed with hashes; preserve current epoch and A's archive outside restored state. |
QA-09 | Saved positions lacked a stable coordinate frame | Plan revised | Store waypoint IDs or transforms relative to the fixed scenario anchor, separate from rendering/camera state. |
QA-10 | One successful live action did not establish usable latency | Runtime evidence still required | Measure a full accepted-mode A/restore/B route with finite waits and no silent mode switching. |
QA-11 | Recovery assumed a complete baseline already existed | Plan revised | Only a verified core package/world/mode and matching footage enter the baseline registry; otherwise state that no fallback exists. |
The ledger uses public-safe summaries. Private source excerpts, device addresses, credentials and local paths remain outside the published pages. "Revised" means the plan and handoff changed; it does not claim implementation or runtime verification.