The promise, measured

Everything we claim on the home page, with its measurement, its scenario and its machine. Anything not yet measured says “pending”.

The goal

The same rules for people and machines

BotStrikes starts from a question: who wins when humans and AIs play under exactly the same rules? Answering it takes a game where an AI is just another player. It connects through the same documented protocol, sends the same keys and mouse movements as a person, and only receives what the server authorizes it to see. The server enforces this: nobody’s code is trusted.

5v5

Truly tactical

Today: fixed, learnable recoil, accuracy that depends on movement, and penetration by material, with three weapons: rifle, pistol and knife. Round-based matches with a scoreboard and the bomb are already integrated and verified (F6 in progress); money accounting is implemented; purchasing remains pending. No content is added until the core meets its criteria.

64Hz

Anti-cheat first

The server decides every movement and every shot. No enemy outside your line of sight travels over the network, with no margin: what is behind a wall is not sent. The exception is your own team: since ADR 0094 the server sends the position of living teammates even when they cannot be seen —seat and position, 13 bytes, nothing else—, because it is the genre’s radar and it is information the team already has; health, weapon and aim are left out. AIs receive exactly that, with the same bytes and on the same tick. Sending commands faster does not move or shoot faster.

120FPS · 1080p · target

Designed for old hardware

Current target: an average of about 120 FPS at 1080p (8.3333 ms per frame) on the reference card, an AMD Radeon HD 7800 Series from 2012, with an in-house C++ engine and under 1 GB (ADR 0035, Sep 12, 2026). The 60 FPS on an Intel UHD 630 went from requirement to desirable on Sep 13, 2026, by the founder’s decision (ADR 0042). Linux and Windows packages exist; Windows was tested under Wine. macOS has packaging scripts that have not been validated on a Mac.

Project status · 2026-10-07

Where we stand

Functional development prototype. Implemented code, an automated test and a publicly available release are different states. Each area's limitations are listed below; automated tests do not replace human checkpoints.

Development versions: BotStrikes 0.2.0 · protocol 15 · Agent API 14.

Area	Implemented or recorded	Limit or pending work
Networking and combat	Authoritative server at 64 Hz, encrypted transport and networked combat. Serves up to 10 players; the format holds 16.	A match with two people on separate computers remains unverified.
Rounds and objective	Rounds, scoreboard, bomb and teammate spectating after death are implemented.	F6 remains open. The mixed 5v5 match in H4 requires human validation.
Economy	Money accounting is implemented and reaches the HUD, agent and training gym.	Purchasing is still pending.
Team communication	Chat uses a separate channel. AIs can read and write; the player can read it.	Player text input and in-game voice are missing. Communication parity in mixed matches remains an open decision.
SDK and training	The C library, Python SDK, reference bots and training gym work internally.	The repository is private. No public SDK download or public game server has been verified.
Human and AI fairness	The server filters perception and limits AI turning, acceleration and key changes where applicable. Parameters are provisional.	H5 is pending. Fitts aiming error, bot sandboxing and command deadlines are not implemented. Human equivalence has not been demonstrated. Tracers are clipped using samples; continuous visibility of the whole segment is not guaranteed.
Replays and endurance	The selected private packages retain the current-packages executables and include BSREPLY2 and the private-file and rejection-logging fixes. Linux records and verifies rules, build/content identity and per-tick hashes; its checker is archived with exact libraries and data. Ten CLI cases check recording, links and write failures without destroying the previous file.	Each replay requires its matching-build checker and exact map/rig. The Windows checker starts, but the new build produced no Windows replays under Wine: successful private writes and replay checks on native Windows remain pending. The replay-package Windows tests belong to the earlier build. V1 needs its historical checker. Fingerprints do not authenticate the file; no viewer or crash recovery is established. Parent directories must be trusted; power-loss durability is not promised. F6, a thousand networked rounds, current full verification, native Mac and human tests remain open.
Map and performance	Cinta is integrated. The new populated-scene muzzle-flash measurement remains above the frame budget. The local comparison confirmed no regressions under the 5% criterion.	The last Radeon measurement predates the braking change. The new trial compares builds on RTX 3090, different hardware; it establishes neither the reference-machine budget nor attribution of differences solely to the algorithm. F5 and H3 remain open.
Visual regression	Secondary braking corrected through explicit source phase shared by authority and client. The integrated build reproduces 128 candidate captures. New bank: 83 scenes per renderer, 10 llvmpipe differences; RTX 3090 has no comparison references.	References and H3 remain unaccepted. Earlier Radeon benches belong to the protocol-14 build. The current private selection is identified in the Linux and Windows sections; it retains the current-packages executables. Its tests are automated and local. Visual diagnosis and matching captures establish neither universal continuity nor human acceptance.
Linux	Version 0.2.0/protocol 15 package repacked with safe installer/launcher help and ES/EN/PT-BR guides. It retains the previous executables and 77 resources. Upgrade from current-packages, six connections at four resolutions, help before/after installation, replay and uninstall checked on Ubuntu/EGL/llvmpipe. No recompilation.	Private selection: consistencia-20261007/windows-help/real, by path and hash; requires fresh invitations. Muted local automation in an Ubuntu container, not a second physical machine or human match. Earlier private-file and archived-checker tests apply to the same executables. English HUD; audio, keyboard/accessibility, visual acceptance and full verification remain pending. Not a public download.
Windows	Version 0.2.0/protocol 15 installer repackaged: launcher help and three ES/EN/PT-BR guides updated; game executables and 77 resources unchanged. Upgrade from linux-help-final, 18 installed help cases without connecting, six primary/spare connections at four resolutions and uninstall checked under Wine/Xvfb. Personal files survive upgrade and uninstall. The session’s Linux server replay was verified.	Private selection: consistencia-20261007/windows-help/real, by path and hash; requires fresh invitations. No public download. Muted local automation does not establish native Windows/Mac support, human play or visual/audio acceptance. The Windows server rejects Wine’s unverifiable ACL before writing tokens or replays; successful private writes and replay on native Windows remain pending. Keyboard/accessibility remain pending. Installer finish-page actions were checked under Wine in the previous delivery; that GUI path was not rerun. English HUD. The Linux server stayed alive until all captures finished.
macOS	Mac build and packaging scripts exist.	There is no validated Mac build or verified support.
External access	Tokens, an advertised public address and private-session invitations are implemented.	Remote reachability, distribution and a human session remain unverified. There is no public launch.

Recorded rendering measurement

Measurement date: 2026-10-04. Scenario: cinta-poblada. Hardware: AMD Radeon HD 7800 Series (radeonsi, pitcairn, ACO, DRM 3.64, 7.0.0-34-generic), Intel(R) Core(TM) i7-10700 CPU @ 2.90GHz; 1920x1080.

Mean	Target	p99	1% low
9.5821 ms	8.3333 ms	15.8888 ms	60.5957 FPS

Recorded rendering measurement

Mean	Target	p99	1% low
9.6178 ms	8.3333 ms	16.1419 ms	58.9013 FPS

Recorded rendering measurement

Mean	Target	p99	1% low
9.4217 ms	8.3333 ms	16.017 ms	60.4278 FPS

Recorded rendering measurement

Measurement date: 2026-09-29. Scenario: cinta-poblada. Hardware: AMD Radeon HD 7800 Series (radeonsi, pitcairn), Intel i7-10700; 1920x1080.

Mean	Target	p99	1% low
9.0708 ms	8.3333 ms	16.2571 ms	60 FPS

These figures apply to the stated date, hardware and scenario. Internal tests do not establish human playtesting or public availability.

Hardware requirements · planned

The hardware we target today

We target older PCs, with performance goals still to be demonstrated on each machine. On Sep 12, 2026 the founder changed profile A’s mean from ≥ 300 to about 120 FPS at 1080p (ADR 0035); on Sep 13, 2026 Intel UHD 630 became desirable (ADR 0042). The absolute target is evaluated separately from the 5 % regression comparison against references. That comparison considers mean, p99 and 1 % low; missing hardware or a reference can leave it pending. Neither absence of a confirmed regression nor a successful exit establishes the absolute budget by itself.

Profile	Reference hardware	Target	Evidence today
Office laptop (the most demanding profile)	Intel HD 620, 2 cores / 4 threads, 8 GB, hard disk drive	1080p ≥ 60 FPS, load ≤ 15 s	Pending Hardware still to be obtained.
Desktop iGPU	Intel UHD 630	1080p, frame p99 ≤ 16.6 ms	Pending The test machine’s iGPU is not enabled yet. Since ADR 0042 this profile is desirable, not a requirement.
Old dedicated GPU	GTX 750 Ti · AMD Radeon HD 7800 Series 2 GB	An average of about 120 FPS at 1080p (8.3333 ms) and a 1 % low above the monitor refresh rate (ADR 0035); a 60 Hz reference is documented in ADR 0052	Measured · not met today The “cinta-poblada” scene at 1920 × 1080 on Radeon HD 7800 and i7-10700: the October 4, 2026 record gives a 9.5821 ms mean, 15.8888 ms p99 and 60.5957 FPS 1 % low. These are medians of five runs after the muzzle-flash correction. It remains above the 8.3333 ms target. The GTX 750 Ti was not measured in this test. Earlier measurements retain their dates; this test does not establish results on other machines or human acceptance.
Server	1 core of an i7-10700 per match	Tick p99 ≤ 1 ms in a 5v5	Measured 0.669 ms with 10 players, 9 of them AIs, and animated bodies, in the full verification of Sep 24, 2026.
Connection	Any home broadband	≤ 64 KB/s download and ≤ 8 KB/s upload per player	Measured Historical bench record from Sep 24, 2026: worst case of 31.6 KB/s downstream and 7.7 KB/s upstream. It does not certify every later feature. The current contract retains the 64 and 8 KB/s caps; see the protocol reference for its current analytical calculation.
Memory and CPU	Machine with 4 GB or more	RAM ≤ 1.5 GB, VRAM ≤ 384 MB and CPU ≤ 2.5 ms per frame on i7-10700. The 5 ms on a 4-core CPU from 2015 is an estimate requiring measurement	Pending It will be measured with the F5 map and graphics.
Installation	Hard drive or SSD	< 1 GB in the first public version	Pending Internal Linux and Windows packages have automated installation, update and uninstall checks; Windows was tested under Wine. No public download has been verified. Native Windows and macOS validation is pending.
System	Linux; Windows 10/11 under Wine; macOS pending	Validate GPU, input and audio on each system	Pending A Windows installer exists; testing on real Windows remains pending. macOS still has no validated build.

Why these numbers: the reference card is the founder’s, an AMD Radeon HD 7800 Series (radeonsi, pitcairn) with an i7-10700, and every rendering figure is measured on it; the UHD 630 and the HD 620 are integrated graphics common in office machines from 2017–2018 and are desirable profiles today; 1 GB can be downloaded on slow connections. We have not yet measured how many internet cafés and homes in the region have something worse: if your machine is older, we want to know. What is not measured says “pending”.

Compare evidence, don’t trade predictions.

We don’t ask you to believe in BotStrikes. Every figure marked as measured has a record with its scenario, its machine and its date; the rest is marked as a target or pending. The code is still private: if you want the record for a figure, ask for it and we will send it to you with its method.

If you have an old PC, an AI or an objection about fairness between people and machines, the invitation is the same: bring your data (hardware, system, what you measured and how) and we compare them.

The records, raw

The published records include historical server, network, concurrent-match, rendering, shooting, protocol and determinism tests. Each JSON identifies its scenario, machine, date and limits. Compare each figure with its own record: results from different scenes or revisions are not interchangeable.

Scope and date. The index contains measurements from September 11 to October 4, 2026. The populated Cinta record from October 4 is already public: at 1920 × 1080 on Radeon HD 7800 and i7-10700, the median of the means of five runs was 9.5821 ms, above the 8.3333 ms target. The export date is not a new measurement. The code remains private: records allow method inspection, but not external reproduction. The historical body oracle test and the cheating bot’s eight checks are described in internal decisions and logs; their sources can be requested by email. They do not prove human fairness or absence of faults across all perception channels.

Index of the records · how to read them and what was removed before publishing

Send evidence