AgentArmies project update edition 2026-09-05 12:30 UTC

Delivered (reviewed, merged or independently verified)

• Overture 2026-08-19.0 normalised for the whole Earth in 13 layers with exact gap objects; generation 5f8a73fd is the live dev pointer (published 2026-09-03 05:48 UTC) with road width, bridge, tunnel and connector attributes (DATA-16).
• Whole-Earth terrain: Copernicus GLO-30 as 442 level-7 COGs (4.35 GB), generation 24357a29, independently verified 2026-09-05 01:49 UTC (C1 chain, object digests, heights within 1.2 mm of raw source plus EGM2008 geoid).
• DATA-21 (merged 06:04 UTC today): one normalised fixture for Austin cell 865b4 with 100 buildings (88 observed OSM heights, 1 modelled, 11 missing, every attribute labelled), ESA WorldCover and a Sentinel-2 datatake, 414 MB on the GPU node; a 27 MB level-9 companion for CI.
• DATA-18 (merged): long-lived R2 connections degraded to 0.65 MB/s; the terrain run now bounds to four fresh connections. DATA-17: a frozen complete-subset generation for replay. DATA-19: a partial publication no longer leaves a stale status.
• The job runner (PLAT-14, platform PR #30 merged 2026-09-05 00:13 UTC): just job run submits a job to the CPU node over SSH without carrying a credential, follows its sealed log across reconnects, returns a receipt with cost; real-SSH acceptance passed at 02:27 UTC.

Pending

• The served pointer does not yet name terrain: the Dagster terrain step failed at 00:21 UTC (status/dev.json: stale, last good 5f8a73fd). A new bounded generation under the seam contract is the next publish.
• DATA-20's proposed ADR 0011 (terrain support and common-edge contract) awaits the compiler's choice between the publisher-lattice window (d1) and bilinear-in-source on the unchanged grid (b). The compiler's decoder currently accepts (b).
• Imagery, land cover, water, heights and the embedding as whole-Earth layers (DATA-8 to DATA-12) are not started beyond the fixture; the coverage dashboard (DATA-14, PR #9) is open.
• PLAT-97: measuring real object fan-out, range and cache latency and recovery cost on the GPU node is filed, not run.

Blockers

• Terrain step failure in the daemon (root cause under DATA-22: unsupplied source tiles conflated with source holes).

Next proof

A terrain-bearing generation served by the pointer with Austin's cells and the seam contract applied; the compiler consuming it on cell 865b4b2b7 (DATA-7 / CMP-22).

Delivered (reviewed, merged or independently verified)

• The content-addressed store with publish, verify and promote and their negative controls (AST-4); the authoring verbs import, describe, measure, budget and licence add (AST-5); a vendor register for Synty, Unity, SpeedTree, Xfrog, Tripo, Meshy and Fab holdings (AST-8; licence rows unchanged, publishable since 08:00 under the owner authorization with the risk logged).
• The library browser (AST-7): every record at every LOD, search, facets, honest gaps (absent from the store, bytes do not match the digest), licence status per record.
• The lamp integration pack (AST-11, merged 2026-09-05 04:49 UTC): the attributed AnisotropyBarnLamp with LOD 0 and LOD 1 in the exact read path the client needs, digest-pinned, credit line carrying the modification clause.
• Live-oak height truth for eval (AST-16): 13.2393 m vs 12.84 m re-derived from the source mesh and transforms.

Pending

• AST-14, the first browsable cross-category batch: 24 items in four packs (19 Khronos sample assets, 5 recovered aa2 Synty street props, private), each with LOD 0/1, collision proxy, measured bounds and a budget result (6 within, 1 breach, 17 unmapped). Two reviewer runs requested changes (report tables not test-bound, CC-BY credits missing the licence URL, budgets unsourced); a fixer is on it.
• AST-13, recovering aa2's asset inventory and import pipeline (vegetation, ground materials, street props, 17 robot assemblies, NPC archetypes) from the frozen quarry: 47,000 lines under review; findings on vendor closure refusals and provenance trust are being fixed.
• AST-15, the fauna handoff for SIM-17: a CC0 barramundi and a first-party proxy bird as public-eligible records, plus the three scuba fish as private-internal reference records with measured facts (head +Z, runtime spine deformation via TEXCOORD_1, KTX2 required).
• AST-17: budget classes sourced by ADR for every first-batch kind.

Blockers

• Implementation, not permission: the gate switch (asset-catalog PR #15) and the C5 pin (contracts PR #7) have to land, the useful-palette bundle (AST-19: Synty props and nature, 17 robot assemblies, 10 Tripo NPCs) is being built, and the broader library intake (AST-20, hundreds of GB) is at the survey stage. No vendor asset is on a public host yet.

Next proof

AST-14 merged with previews and budget results, then the first promoted public pointer with the lamp pack.

Delivered (reviewed, merged or independently verified)

• CMP-6 (tile-compiler PR #11, merged 2026-09-05 02:08 UTC, 302 files): a job for a pinned World Manifest and a level-16 cell emits a solved plan, six draft C3 products, a coverage answer and a digest-bound manifest. The accepted local pack: 29 cells, 174 products, 9,922,019 bytes; nine populated cells with zero exact-overlap pairs, zero reconstruction mismatches, matching roof, floor and collision triangles; max source-to-wire displacement 1.06 mm.
• C3 on contracts main: product manifest schemas and registries (PR #13), the Tile Product wire adoption and browser package (PR #17, merged 02:28 UTC) whose fetch-and-verify was run independently in CPU Chromium over the real pack (29 manifests, 174 products, zero Range requests).
• CMP-32: real building and route geometry compared against the frozen aa2 and lab references; CMP-33 proposes the walkable-ground and road-surface semantics (collision field is the height authority; walkable meshes are projections).
• CMP-25 (merged): public PUTs gated on the pinned legal status.

Pending

• Real terrain (CMP-22): the private GLO-30 COG reader and deterministic sampler (PR #15) and the explicit triangular terrain profile tri-swne-v1 with the supported-terrain pipeline (PR #16) are drafts; the decoder verifies DATA-20's corrected bilinear pair and all 6,144 sample results match a GDAL oracle. The job waits on a coherent private C1 bundle from earth-data with new generation bindings.
• NOTICE.txt from the exact build graph (CMP-27, PR #14) for the compiled artefact.
• No public products: the pack lives on the nodes and behind the private endpoints; a staged and public release is CMP-6's remaining scope.

Blockers

• Terrain generation bindings from earth-data (above).
• C3 v0 cut (PLAT-45, in review) for the release, with C9 (serving API) not yet cut.

Next proof

One terrain-bearing job on Austin cell 865b4b2b7 with correctly based buildings and roads and a matching collision field, repeated under a second job id for byte equality.

Delivered (reviewed, merged or independently verified)

• Terrain data exists for three bounded blocks (above), not the whole Earth, and the seam defect is diagnosed with two candidate contracts; the compiler's decoder already accepts the corrected pair.
• Buildings: real Overture footprints extruded with roofs and floors in the compiled pack; DATA-21 labels heights observed / modelled / missing; LAB-13 (lab-geometry-solves PR #8) adds roof, doorway and interior geometry behind the plan seam from a pinned real footprint, with door rules recorded and byte-identical reruns.
• Vegetation and ground (CLI-4, client-web PR #5): grass as thin instances, a tree card tier, authored wind, placement from product geometry only, off by default behind ?veg=on; visual judgement keep with changes (reads as an unlit overlay; the flock is being withdrawn); LAB-12 places 518 inferred trees on one real route from NAIP as a reproducible diagnostic with no accuracy claim.
• Water (CLI-8, PR #8): Fresnel, glint, refraction and ripples re-expressed from the scuba reference on verified water only; the measured fact is that no root carries a water body yet (Overture water is not imported), so the module mounts, draws nothing and says why.
• Life: SIM-17's GPU shoal prototype and AST-15's fauna records; the traffic core moves vehicles and pedestrians on real roads.

Pending

• Terrain relief in any product or the client.
• Water import (DATA-10).
• Weather and a simulated clock (lighting is authored-fixed; sky_time null on every capture).
• A vegetation catalogue mesh tier (only cards exist).

Blockers

• CLI-3 daylight must land before vegetation and water compose with real light.
• No tree or plant meshes are imported and hosted yet; the SpeedTree and Xfrog holdings are publishable under the 08:00 owner authorization, and the import and hosting work (AST-19) is queued.

Next proof

Terrain relief through the compiler into the client, and CLI-3 with its shadow fixes accepted on a GPU pair.

Delivered (reviewed, merged or independently verified)

• The deterministic core (SIM-3): a fixed 30 Hz tick over a tile group, an event log, snapshots, replay; native x86, arm and WASM produce the same bytes (the traffic page above proves it live).
• CPU scaling: partitioning vehicle leader candidates by lane (PR #14, head d04e014) takes 8,192 distributed vehicles from 36.0 ms to 3.3 ms p50 per tick natively with byte-identical logs; source checks and the gauntlet are green and two formal approvals are on the exact head; it is not yet merged.
• Browser scaling experiment (INB-21 S): thirty browser runs and 9,000 per-tick player samples match the native oracle; at 11,335 stress agents the main-thread tick p95 is 58 ms and intent-to-draw 4.35 s, so correct work can still be seconds late without live admission and reconciliation.
• GPU fauna (SIM-17): a WebGPU compute-plus-render prototype on the RTX PRO 4000; per-frame receipts record about 0.73 ms compute step and 0.40 ms render for the captured shoal, with a 1,024-agent algorithm comparison against WASM brute force. The independent review of earlier sources (629264a to 5087ded) confirmed a direct uniform-grid neighbour search; the pinned 365f8d6 export has not been independently reviewed, and the author discloses open trajectory agreement, red-trajectory and soft-contact limits.
• A numerics audit (PR #13): finite limits, checksum corruption, signed zero, u32 boundaries, seeded replay tables.

Pending

• The session layer, live admission and C6 framing (PRs #9, #10, #11) and the private endpoint (PR #12) are drafts.
• Physics tiers beyond road-following kinematics; vehicles and collisions as in the target; no planes or boats.
• An independent review of SIM-17's 365f8d6 export and eval's hardware-browser check of its public URL.

Blockers

• The dense lane still misses 30 Hz in the browser at stress density; a browser proof of the lane partition is pending.

Next proof

The lane partition merged with its browser proof; SIM-17's 365f8d6 export passing an independent review and eval's public hardware check.

Delivered (reviewed, merged or independently verified)

• C6 v0 (contracts PR #18, merged 2026-09-05 05:02 UTC): private development sessions, shared bootstrap state, movement and Interact, ordered authoritative outcomes, protocol refusals, reconnect state and retained retries; TypeScript, Rust and Python bindings; an authored fixture of 88 frames with 43 controls that both Rust and TypeScript reconstruct.
• SIM-14's private endpoint (world-sim PR #12): one isolated world through the authoritative core, two real Chromium clients over WebTransport receiving matching atomic state; a 150-second expiry run where both clients observed the same session removal and matched at event cursor 539,163 with the same state hash; physical resource controls (16 connection permits, byte caps, 2,049 invalid requests refused with the cursor unmoved).

Pending

• Publication of the approved C6 SDK archives (74,695 B Rust crate, 51,344 B TypeScript tarball) to fetchable URLs.
• The client's C6 consumer; two players meeting in the same real place is the next acceptance in INB-19.
• C4 token field and C10 session key set (PLAT-39) before any public session.

Blockers

• Endpoint is private to the node; no public multiplayer surface exists.

Next proof

Two browsers on different machines in one C6 session on the real Austin pack, with movement visible to both.

Delivered (reviewed, merged or independently verified)

• SIM-9 recovery (world-sim PR #15, draft): the frozen aa2 combat core, class kits (Engineer, Bulwark, Medic), mob FSM and in-memory progression run as an executable reference; Engineer, Bulwark and Medic against four robots end in 85 steps (4.25 s) with 25 accepted casts and two refusals; 232 tests and 14 independent controls reproduced by a Fable critic; recorded-input replay and independent process runs match exactly. SOURCE-MAP records keep, adapt or replace decisions and Rust destinations.
• CLI-5's combat presentation packet (targeting, cast feel, telegraphs, robot locomotion, audio, cooldown sync, loot and death) as a flagged client module with review captures.
• Eval's acceptance panels (EVAL-21) for content lanes: tree form dimensions is now an acceptance bar; impostor-vs-mesh silhouette stays diagnostic.

Pending

• Five documentation corrections (drain order, Laser 350 ms and GCD 420 ms, cast and channel rules) before the reference is a port specification.
• The Rust/C6 port of one ability (Stasis witness filed as C6-CAST-REQUEST).
• No mobs, no progression, no character selection in the new stack.

Blockers

• Depends on the C6 endpoint and the client consumer.

Next proof

One ability cast from a browser through C6 against an authoritative robot, with the aa2 golden parity test as the oracle.

Delivered (reviewed, merged or independently verified)

• world-serve PR #7 (CMP-21, merged 2026-09-05 05:42 UTC, 466 tests): verify a local compiler pack, copy it as one immutable set, serve it through the static resolver, select it by a draft pin; rollback admission; descriptor-anchored opens on Linux so a symlink swap cannot redirect a read; one explicitly configured browser origin. Both CPU endpoints moved to this code at 05:49 and 05:51 UTC with 207/207 objects verified.
• client-web PR #3 (CLI-2, 30,000 lines, draft): the Babylon consumer with a verified loader (pin, index, manifest, product; bytes hashed at the point of use), a decoder mirroring the producer's checks rule for rule, S2 residency halo and the east crossing, capsule traversal with swept collision, the ground rule (nothing drawn or walked unless terrain and collision both verified), and the __aaEval adapter eval's harness reads.
• Independent integration receipts (INB-20): the served 548d31f pack's every product and manifest matches the frozen build; a native Chromium fetch from the preview origin fails only on CORS, which is the configured-origin decision, not a byte problem.

Pending

• CORS or a same-origin deployment so the reviewed client fetches from the reviewed server (today's demo card avoids this by shipping the pack beside the page).
• CLI-6's public coverage presentation (unknown / public-answer-unavailable rows), CLI-7's bounded residency adopted into the loader (the experiment showed retain-all growing to 476 MB heap over 65 repetitions).
• Nothing on client-web main beyond the template; the deployment head is a draft PR.

Blockers

• world-serve PR #5 (pin recipe tools) stalled on an account admission failure and was reassigned.
• No generation index or release: C9 not cut; C8 v0 exists.

Next proof

The reviewed client head fetching the reviewed endpoint in a real browser (the INB-20 candidate), then eval's traversal capture on GPU hardware.

Delivered (reviewed, merged or independently verified)

• eval's capture harness (EVAL-2, done): readiness waits for every subject resource to settle; 18 of 18 negative controls each refuse on their own code; the frozen alpha's Austin ground capture is admissible on the rig (a stalled ninth cell produced an honest inadmissible frame first). The client-web subject and traverse mode with a frame ledger and perf row are in PR #24.
• Acceptance cells (EVAL-18, PR #21): destination halos, a true S2 face-boundary control and the twenty random cells registered; san-juan and st-moritz poses and clean boxes. CLI-3's GPU pairs captured and judged (above).
• platform: the job runner caller (PR #30), the shared nightly demo publisher (PR #35), the gauntlet's repair receipts (PR #40), helpers hardened (aa-gh verifies its App identity; aa-linear drains stdout).
• labs: LAB-1's earth compiler reached its Travis bar (LPIPS 0.5263 vs 0.5343; FID 61.87 vs 233.66) while the worldwide control did not move (0.6146): the model learns Travis, not the world. LAB-3 (LLM-authored layers) was killed on 2026-09-03 with its evidence. LAB-12 and LAB-13 delivered bounded diagnostics today.

Pending

• The shared demo publisher's gate (PLAT-62, PR #41) switching to the owner-authorized record so eight repos' demo pages can publish.
• GPU captures of the CLI-3 shadow fix and the CLI-4 vegetation cost (requested via eval).
• The first blind pair of the new stack against the frozen alpha (EVAL-4 judge packs, PR #6).
• Fleet execution off the Mac (PLAT-79): runtime ownership, account admission failures and idle-turn wakes cost hours this morning.

Blockers

• Account usage limits hit three leads at once (recovered on the same account); a T3 session-stop bug on the GPU node (PLAT-99).

Next proof

platform#41 merged and the org demos index turning repos from pending to live; eval's traversal capture of the client candidate on GPU hardware with a perf row.