Skip to content

Hosted evidence

Every claim on this site about hosted Reactor rests on a paid run listed here. Each run is one check of the hosted harness, an Effect program over the public API, run by the maintainer against hosted H3 Reference Turbo Realtime (reactor/h3-reference-to-video-turbo-realtime), the avatar, character and rejoin checks against Vidu S2-Avatar (reactor/vidu-s2-avatar), and fasth3 against FastH3 (reactor/fast-h3). Its evidence file and summary are committed under integration/hosted/evidence/, with the session ids, so Reactor can find its side of each run.

The harness refuses to run without --i-authorize-paid-sessions and a budget. Each run reserves its worst case in a ledger before any token exists, and a ledger’s total is capped at $10.

35 runs, 42 sessions, from 2026-09-24 to 2026-10-10. Every allocated session ended with its termination confirmed by an independent read. The third 0.11.0 candidate attempt passed.

Started (UTC) Code Check Sessions Estimate Result
2026-09-24 02:40 0.3.0-rc.0, from npm vertical 1 $0.750 Fail: acceptance seen only by metadata; no ICE pair named
2026-09-24 02:42 0.3.0-rc.0, from npm takeover 1 $0.750 Fail: no frames after attaching
2026-09-24 03:18 0.3.0-rc.0, from npm vertical 1 $0.750 Fail: as before; settled H3’s reply order
2026-09-24 05:10 checkout 0255871 vertical 1 $0.750 Pass
2026-09-24 05:10 checkout 0255871 takeover 1 $0.750 Pass
2026-09-25 00:21 0.3.1, from npm audio 1 $0.750 Pass
2026-09-27 19:58 0.6.0, from npm scheduler 1 $0.750 Fail: a position-zero clip went ahead of the running build
2026-09-27 19:59 0.6.0, from npm scheduler-renewal 2 $1.500 Pass
2026-09-28 00:16 0.7.0, from npm scheduler-edits 1 $0.750 Fail: a continued insert missed its place
2026-09-28 00:19 0.7.0, from npm scheduler-cut 1 $0.750 Fail: position zero again; a second stop cut the cutter
2026-09-28 02:31 checkout aabe52f scheduler-edits 1 $0.750 Pass
2026-09-28 04:51 checkout 8458d08 tokens 1 $0.275 Fail: an INACTIVE session taken as ended
2026-09-28 05:12 checkout f5daf55 tokens 1 $0.400 Fail: one criterion of the check’s own making
2026-09-28 05:14 checkout 59136a2 cut 1 $0.250 Pass, with a moderation probe
2026-09-28 12:03 checkout 69f34e9 vertical 1 $0.150 Fail: the check’s live-video rule met H3’s end-of-clip frame
2026-09-28 12:11 checkout 3539ed4 edits 1 $0.438 Pass
2026-09-28 15:51 checkout 799ed3d adoption 1 $0.388 Pass, all 17 criteria
2026-09-28 16:12 checkout c548077 tour 1 $0.463 Pass, all 30 criteria
2026-09-28 23:37 checkout 87d9f67 show 3 $1.538 Pass, all 25 criteria
2026-09-29 14:00 checkout c9f3f96 unconnected 2 $0.900 Pass
2026-09-29 14:03 checkout c9f3f96 show 3 $1.413 Pass, every criterion
2026-09-30 01:12 checkout b64aff33 dropped 1 $0.800 Pass, every criterion
2026-10-01 02:01 checkout d013d00 showreel 1 $1.680 Pass, every criterion; footage on the media release
2026-10-01 15:05 checkout becfca5 avatar 1 $1.155 Fail: no picture without a resume at live; end_call past 10 s
2026-10-02 12:07 checkout b7ce932 avatar 1 $0.406 Pass, every criterion
2026-10-02 12:49 checkout 819d45b character 1 $0.028 Fail: the native connection failed at the first command
2026-10-02 12:58 checkout 3239468 character 1 $0.196 Pass, every criterion
2026-10-03 12:26 0.10.0, from npm vertical 1 $0.455 Pass, all seven criteria
2026-10-03 12:27 0.10.0, from npm tour 1 $1.295 Pass, all 30 criteria
2026-10-03 12:28 0.10.0, from npm edits 1 $1.260 Pass, every criterion
2026-10-03 14:22 checkout ca391759 fasth3 1 $0.525 Pass, every criterion
2026-10-03 14:48 checkout ab4a792 rejoin 1 $0.245 Pass, all four criteria (run 5e507c87)
2026-10-09 13:03 checkout 7944b74 candidate 0 — Fail: first H3 session creation returned HTTP 402
2026-10-09 17:35 checkout e28d197 candidate 2 $5.460 Fail: FastH3’s phase closed past its 80 s deadline
2026-10-10 01:28 checkout 8b5958d candidate 2 $5.075 Pass, all 32 criteria (run 68595077)

The summaries, one per ledger: 0.3.0-rc.0, 0.3.1, 0.6.0, 0.7.0, 0.8.0-dev, 0.8.0-api, 0.8.0-probe, 0.8.0-rc, 0.8.0-dropped, 0.9.2-avatar 0.9.2-character 0.10.0, 0.10.0-fasth3, 0.10.0-rejoin, 0.11.0-dev-candidate, 0.11.0-dev-candidate-retry and 0.11.0-dev-candidate-corrected.

  • Estimates are the evidence’s own, at the rate Reactor’s pricing stated then: 125 credits a second, $0.0125. Up to 0.7.0 they count every started minute whole, so they are upper bounds; from 0.8.0 they count seconds. On 2026-09-30 the pricing lists 350 credits a second, so the same sessions would cost 2.8 times the per-second estimates today. The avatar, character and rejoin runs are at Vidu S2-Avatar’s 70 credits a second, $0.007, counted by the second. The published 0.10.0 H3 runs and the FastH3 probe use 350 credits a second, $0.035, counted by the second.
  • Failures. Each found cause was fixed before the check ran again to qualify. Most were the library’s: 0.3.0-rc.0’s acceptance, ICE pair and attach, 0.7.0’s misplaced continued insert and doubled stop, and an INACTIVE session read as ended. Two were criteria of the checks’ own making. The third 0.3.0-rc.0 vertical repeated the first on the same bytes, to settle billing and H3’s reply order. Position zero, listed below, is H3’s own ordering and is still unexplained. The first avatar run resumed no track at live and waited 10 s for an end_call that took 13.2 s; the reconnect it then tried for the picture hung 82 s, past its own deadline. The first character run’s connection failed with the native addon’s Protocol failure 21 ms after the provider’s first command, 6 ms after the connection was ready; its cause is not yet known, and the evidence now keeps the addon’s own sentence for that failure, so a recurrence names its check. The 0.11.0 candidate attempt received HTTP 402 at its first H3 creation, before any session was allocated. No candidate phase ran, and FastH3 was not attempted. The evidence records no charge estimate and retains the $6.30 reservation. A second attempt, authorized separately, used a ledger of its own with a $10 limit and reserved $6.30. H3 passed all 15 of its criteria. FastH3’s firm clip started 18.05 s after allocation, where H3’s started at 12.38 s, and every later step followed that much later. FastH3’s phase closed 80.05 s after allocation, past its fixed 80 s work deadline, so the run failed before its remaining criteria were judged. Both sessions’ ends were confirmed, and the public Ledger was left empty. The filler before the readyBy clip then asked for 10 s in place of the family’s longest clip. A third attempt at 8b5958d, authorized separately in a ledger of its own with a $10 limit, reserved $6.30 and passed all 32 criteria. FastH3’s firm clip was Ready 14.18 s after submission, and its session closed 75.87 s after allocation, inside its 80 s deadline; H3’s closed 68.99 s after allocation. Independent reads confirmed both ends, and the public Ledger was empty after each.
  • Commits. The 0.8.0-dev commits were renamed when the branch’s messages were reworded, with no change to any tree; the ledger’s summary maps each to its new name. The second show’s checkout, c9f3f96, matches the 0.8.0 release’s library source except for its version string.
  • The Reactor dashboard listed each session’s duration to the second, matching the time from allocation to the DELETE within a second (0.8.0-api), and listed a 10-second session at 10 s, so Reactor meters by the second (0.8.0-probe).
  • Whether the meter starts at creation or at ready stayed open: ready came 0.19–0.38 s after allocation, under the dashboard’s one-second resolution.
  • The balance readings fit neither the published rate nor whole minutes. 0.3.0-rc.0’s third run moved the balance by $0.03–0.05, against $0.11 at the published rate; 0.8.0-dev’s sessions had not moved it hours later; and the $1.09 between the readings around 0.8.0-probe fits neither.

Published 0.10.0 on Effect 4 (0.10.0, 2026-10-03): vertical, tour and edits passed on the npm client and native packages with Effect 4.0.0, with all three sessions confirmed ended. The tour reconnected in 2.014 s, kept its Ready clip and received fresh media. The edits aired in their planned places, with 71–143 ms seams and no dark frame. The runner’s one-second timer was at most 5.7 ms late. The three estimates total $3.010; dashboard charges have not yet been read.

Media and connection (0.3.0-rc.0, 0.3.1, 0.8.0-dev):

  • 1344×768 BGRA at 23.7–24.2 fps, all 124 frames of each 5-second clip, none lost; 48 kHz mono audio, none lost; a median 4.9–8.3 Mbps with 20–22 ms round trips over a direct path.
  • Allocation to ready in 2.5–2.8 s; an enqueue accepted by its correlated reply 0.03 s after submission; the clip started 2.06–2.18 s later and its first frame came 0.08–0.10 s after that.

Ownership and tokens (0.3.0-rc.0, 0.8.0-dev, 0.8.0-api):

  • A process attached 2.69 s after the owner was killed and received frames 0.21 s later.
  • An owner under Node on the isolated native host was killed 12.7 s into the run: a raw attach was ready 3.06 s after the kill, named the playing clip and closed without ending the session; H3Source.resume started 0.57 s after the creating token expired, adopted the session 2.71 s later, and enqueued a clip with a reference image and reference audio on a refreshed 12-second bound token.
  • An expired token was refused with 401 and one without the bind with 403. A 7-hour token was clamped to 6 hours.

Recovery (0.8.0-api, 0.8.0-rc):

  • A reconnect 2 s after the creating token expired moved to the next connection generation in 2.17 s, with the Ready clip kept.
  • A connection cut with 8.07 s of air secured read ready again 1.64 s later and showed a fresh frame 1.82 s after the drop, with no replacement session opened (1.84 s and 1.95 s in the second run).

Playout (0.8.0-api, 0.8.0-rc, 0.8.0-dev, 0.7.0):

  • One playout over three 75-second sessions: 20 clips and 2,372 frames at 22.2 fps, no frame lost, its reader never behind.
  • In the two show runs, seams of 46–169 ms with no dark frame, and a planned switch between sessions with 420 and 432 ms between clips on air.
  • An edit batch in effect 0.91–1.05 s before its boundary; its withdrawn clip never aired.
  • Cues 0–1 ms from due; an At item 545–559 ms late, never early.
  • A session ended by moderation was replaced: the clip it had Ready was rebuilt on a session opened 3.17 s after the verdict, and aired 5.3 s after the lost clip left the air. A cut lane stopped a clip with one stop and a 165–168 ms pause.
  • A continued clip built in 5.2–5.45 s against 2.1–2.2 s for an independent one (0.7.0). In the passing edits and show runs, the picture at a continued seam changed 2.3–5.8 times the clip’s own motion, and at their independent seams 11–20 times; a cut seam elsewhere measured 6.5 times.

Termination: every one of the 38 sessions confirmed ended. In the earlier 0.8.0 runs, where timed, CLOSED read 0.63–0.86 s after the DELETE.

The showreel (0.9.0-showreel, 2026-10-01): one Playout over one 70 s-capped session aired five 8-second scenes back to back, and the decoded frames and audio were recorded in-process under Bun as they arrived: 958 frames at 1344×768 and 24 fps, the reader never behind, seams of 89–120 ms with no dark frame, gaps on air of 2–67 ms between scenes, about $1.68. The footage is on the media release.

Vidu S2-Avatar (0.9.2-avatar, 2026-10-01 and 10-02): the avatar probe drove the model through the raw Session under Bun, in one session per run.

  • The second run’s picture came 0.20 s after both tracks were resumed at live, at 24.9 fps in frames of 928×1088 and of 640×360, none lost; its voice came as 48 kHz mono PCM.
  • An avatar was ready 26.7 s after create_avatar on the first run and 2.1 s on the second, from the same photo. A call was live 2.7–4.5 s after start_call.
  • The character’s first transcript came 2.5–2.6 s after live, and the sound of its answer 3.1–3.4 s after a say. An interrupt 1.5 s into an answer silenced it 0.40 s later.
  • update_call to another voice answered call_updated in 0.24 s.

The ViduS2Avatar provider (0.9.2-character, 2026-10-02): the character check ran one call through the provider alone, on Effect 4.0.0 under Bun, in one session per run.

  • The passing run’s provider made the avatar in 3.29 s and returned the call live 3.16 s after startCall; its own resume at live brought 251 frames, 640×360 before the call and 928×1088 in it, at a median interval of 40 ms, with 1,287 audio blocks and none lost. say was acknowledged in 0.03 s and answered, and endCall returned the call’s end in 8.65 s.
  • The failing run’s connection failed with a Protocol failure 21 ms after make’s first get_state; the same cluster and server version (1.20260930.52) served the passing runs.

These are behaviours of hosted Reactor and its models the runs recorded, worth knowing whatever client you use.

FastH3 (0.10.0-fasth3, 2026-10-03): five comparable state/queue pairs had equal generation counts, four with an unbuilt clip, and all 18 H3 state fields were present. Accepted length samples ranged from 5.167 to 14.375 s; off-grid samples rounded upward, while 5 and 15 s were refused. History kept at least two clips; its capacity and eviction rule were unmeasured. With flush off, clip_stopped came 130 ms after stop, but frozen-tail timing was unmeasured, so this run does not establish the documented drain time.

Dropped H3 session (0.8.0-dropped, 2026-09-30): run e87da9b1, from checkout b64aff33, read ACTIVE through 4.21 s after its owner was killed, then INACTIVE from 6.21 s through the last read at 60.21 s. Reactor had not ended it; the API key ended it and confirmed CLOSED.

  • H3 broadcasts the queue_update listing a new clip before the enqueue’s correlated clip_queued reply (0.3.0-rc.0).
  • An attached connection receives no media until it resumes its tracks, as Reactor’s sessions page says (0.3.0-rc.0).
  • The deployment reports version v0.0.0, where the documented schema is 0.5.5 (0.3.0-rc.0, 0.3.1).
  • A position-zero enqueue went ahead of the build in flight, where the schema says the running build is unaffected (0.6.0, 0.7.0). A queue read sent 1 ms behind an enqueue listed the new clip: commands applied in the order sent (0.7.0).
  • stop names no clip and lands after its reply; under autoplay it skips to the next clip, so a second stop cuts that one (0.7.0).
  • H3 flushes to black at each clip’s end by default (flush_on_clip_end), one dark frame per seam (0.6.0, 0.7.0); holding the last frame removes it.
  • A session whose last connection has gone reads INACTIVE and is still live (0.8.0-dev). One read ACTIVE for at least the first 2.1 s after its last connection closed, when the reads stopped (0.8.0-api); another, first read 8.4 s after its owner was killed, was INACTIVE (0.8.0-dev).
  • Content moderation’s terminate verdict came 0.40–1.01 s after the flagged enqueue, which had been answered normally, and named no category, input, command or request (0.8.0-dev, 0.8.0-api).
  • A 16,171-character prompt was accepted, then failed by clip_failed 2.39 s later: H3’s text budget is enforced at build (0.8.0-api).
  • A clip recording came back ClipReady in 2.6 s, but its download timed out at 15.1 s (0.8.0-api).
  • A session allocated and never connected read ACTIVE throughout and was ended by Reactor at its 60-second cap, 60.09 s after allocation; a second create on a spent single-session token was refused with 403 (0.8.0-probe).
  • Vidu S2-Avatar sent a call’s picture only once main_video was resumed after the call went live; on the same connection, the next call’s picture came with nothing done. Its voice came without that resume (0.9.2-avatar).
  • Vidu S2-Avatar acknowledged a refused command, and its command_error, which names the command but no request, came first. A command carrying an explicit null and a say with no text got an invalid_command error frame (0.9.2-avatar).
  • Vidu S2-Avatar answered end_call 6.9–13.2 s after it was sent, past the SDK’s 10-second reply deadline on the first run (0.9.2-avatar).
  • Vidu S2-Avatar declares clone_voice and refused it with CLONING_DISABLED; list_voices listed 56 voices, and call_max_seconds read 7200 (0.9.2-avatar).
  • A connection that replaced a Vidu S2-Avatar call’s own during the call carried the session’s messages, the state each second and the transcripts, but none of the call’s picture or voice, with both tracks resumed while the call was live (the first 0.9.2-avatar run).
  • Most of 0.9.0. One paid run has exercised 0.9.0: the one-session showreel check, from a checkout with the published linux-x64 addon. Playout.place and follows, 0.9.0’s renewal and its other checks have run only on ReactorTest, and its published client packages have not run on hosted Reactor.
  • The turn check, the vertical from a network that blocks outbound UDP so a relay must carry the media. Every paid run went over a direct path.
  • reactor-effect-browser. No paid run has used the browser host.
  • The in-process native host under Node. It ran under Bun, and Node ran only the isolated host. The quickstart and the terminal viewer use it when run with node.
  • macOS arm64. Every paid run was on Linux x64.
  • Long sessions. The longest paid run lasted under two minutes; no multi-hour renewal soak has run.
  • The avatar example, and more of Vidu S2-Avatar. The provider ran in the character and rejoin checks, not through the example program. A caller’s microphone or webcam and reference images have not run. Whether Reactor ends a Vidu S2-Avatar session at its cap is open: on the first run nothing read the session between its last message, at 81 s, and the key’s end at 164 s.
  • Restreaming and native publication. No run has sent native media on to a browser, to RTMP or to LiveKit.