4 KiB
2026-07-31 — response to adversarial challenge: CB-RES-0001 (T04)
One round, per InnerLoop §Step 2. Each finding answered with evidence or
conceded. New measurements were run before answering B1/S1/S4; harness
gained --disable-undo and --single-run flags.
B1 (blocking) — "architectural, not a tuning artifact": partially conceded, empirically resolved
The reviewer was right that the claim was untested and the disableUndo
flag unacknowledged. We reran the harness both ways:
| moves | with undo (moves/s) | disableUndo (moves/s) |
|---|---|---|
| 5,000 | 1,439–1,930* | 1,118 |
| 10,000 | 1,605 | 1,076 |
| 20,000 | 870 | 942 |
| 40,000 | — | 733 |
| 100,000 | DNF @300 s | DNF @240 s |
* run-to-run variance on this machine is large (1,930 vs 1,439 for the same config); now noted in the survey's risks.
Verdict on the finding: the undo stack is the dominant superlinear term at small N (concede: it is a documented one-line flag, and the original "not a tuning artifact" sentence was an overclaim — retracted). But degradation persists with undo disabled (−34% from 5k→40k, RSS 100→232 MB, and 100k still does not finish), consistent with the client's unbounded redux log/deltalog, which has no client-side off switch in 0.50.2. The survey now reports both datasets and attributes the two mechanisms separately. The D3-scaling row is restated against the disableUndo (best-case) configuration: 0.66× at 40k vs 5k.
S1 (significant) — memory methodology: conceded, fixed
--single-run mode now measures RSS immediately after the main run in a
fresh process with no warmup/determinism clients. New indicative figures:
100 MB RSS after 5k moves (disableUndo, isolated) rising to 232 MB after
40k. The old 224 MB row is withdrawn; memory rows are labeled indicative
(raw RSS includes Node baseline; heap delta not isolated).
S2 (significant) — omitted candidates: conceded, added
OpenSpiel and Ludii added as cited candidates; BGA Studio and Colyseus added as one-line secondary references. The verdict's "what none of them do" paragraph was re-checked against OpenSpiel specifically and reworded: OpenSpiel does cover simultaneous-move + imperfect-information semantics with deterministic replay via action histories — what it lacks is the production layer (client visibility projections, networking, snapshots, embeddable boundary), and that is now stated precisely instead of the blanket claim.
S3 (significant) — unsourced D3 ceiling: conceded, relabeled
The 10⁵–10⁶ events/s range stays as an estimate flag, and the headline ≥100k events/s + flat-scaling figure is now explicitly labeled a stipulated engineering target, not a competitor-derived baseline. A measured Rust comparator remains an open follow-up (noted in the survey); until then, parity-cap applies to any evidence row leaning on that number.
S4 (significant) — updatePlayerID in the timed loop: answered with a control
Measured in isolation: ~466,000 calls/s (~2 µs/call) vs ~500–1,100 µs per applied move — contamination < 0.5%. Left in the loop (it is part of driving multi-player turns through one client); provenance note added.
M1–M4 (minor) — all conceded
- M1: traceability row demoted from baseline table to an observation.
- M2: LOC corrected to ~36 and marked gameable/secondary.
- M3: provenance now states hidden-info (
playerView) and hand depletion are unexercised by the workload; commit/reveal shape only. - M4: maturity figures kept with the npm-checkable fact primary.
Net effect on the survey
Status remains draft → approved (per this round's disposition: "approvable with corrections", all corrections applied). The central comparative conclusions survive: boardgame.io is the closest semantic comparator but inactive, heavyweight for what it does, and degrades with history even in its best configuration; the surpass opportunity is the combination of production projection/networking, flat-cost replay, and an embeddable boundary — now stated with OpenSpiel explicitly in view.