TicTactics: one reviewed registered original, 2026-11-14

The registered tictac route in lib/apps.js selects the existing local test/binaries/entertainment-pack/tictac.exe (28,432 bytes, SHA-256 0c42754ee5441d74ded4e8389817ec7c7b834984a2c81c4807ab22c0afed3e37), with runSlice cooperative 25000 / compat 100 and no declared DLL or file closure. Exact media sizes and SHA-256 are recorded in scratch/runs/20261114-tictactics-reviewed-outcome/. No download, alternate executable or configuration change is authorized.

Rules established from the native executable

This is not assumed to be ordinary 3×3 tic-tac-toe. Resource menu entries offer 3×3, 3×3×3 and 4×4×4. The preference reader at 0x01002040 calls the clamping profile helper 0x01001f70: key index 4 (Game) defaults to 2, range 0–2 (0x010020cd–0x010020da). The shape initializer 0x01003560 maps 0 to width 3 / one layer, 1 to width 3 / three layers, and 2 to width 4 / four layers. It computes cells per layer as width squared and total cells as that product times layers (0x01003631–0x0100364e). Thus the absent-profile default is a 64-cell cube. The original's complete initial pixels confirmed all four 4×4 layers empty before gameplay.

Native pixel-to-cell mapping is 0x010014c0; 0x010015c0 accepts a mapped cell only when it is not -1 and its native occupancy is zero. Mouse-down previews a valid cell; mouse-up commits it (0x01001ab7 onward). The native commit routine 0x01002d50 paints the current piece, checks a winning line, checks full-board exhaustion, switches players with XOR 3, and invokes the computer only on player value 1 (0x01002de4–0x01002e00). This establishes separate own placement and computer reply; receipt consumption alone proves neither persistent board change nor a natural outcome.

The rule initializer builds native winning lines from the static direction table at 0x01005050. For the width-4 variant, the outcome checker at 0x01002b70 requires four equal nonzero pieces on a generated line. The width-3 path compares three. Winning pieces are retained for native flashing; full-board exhaustion ends play. Startup 0x01003760 chooses the starting player from the untouched Start preference, randomly when it is zero, and can immediately place a computer opening. Human color and actual next turn must therefore be established from complete original pixels and this native mapping, not from the title or a presumed X/O convention. No live board memory, seed, solver, minimax or native AI move computation is read or reproduced.

The native menu IDs further establish color identity without assuming X/O: “Random player starts” is ID 0x6f, “Red always starts” is 0x70, and “Blue always starts” is 0x71. The unchanged start-option handler at 0x01001827 subtracts 0x6f, so native player 1 is Red and player 2 Blue. The commit path invokes the computer on player 1, making the human Blue. The complete initial original pixels showed all 64 cells empty and Blue to move; these static menu labels are not a reason to change the starting preference.

Guarded controller

tools/tictactics-controller.py derives finite single-click transport from the released turn-based controllers. It requires a permanent PID/start/window binding, a personally reviewed complete board, rules and player/turn evidence, an empty target and a settled opponent before each placement. Placement scenes expire after 5 seconds; menu scenes after 30. There is at most one ordinary bootstrap Start and an irreversible final 180-second Quit reserve.

Each action prearms an independent 1750ms fallback UP, validates a unique nonce and owner/epoch receipt, and joins the release child before capturing. UNKNOWN, mixed identity, cancellation, late ACK and child death close input without replay or rebinding. Policy and transport fixtures establish controls, never original-game progress. Evidence retains compact logs, selected PNGs, Git SHA and binary patches, without source or media snapshots.

Validation: 19 policy tests, nine loopback HTTP controls and ten actual CLI controls pass. The actual CLI cases cover a finite pointer pulse, late ACK, UNKNOWN after consumption, cancellation, window replacement, release-child death, parent death, fallback UP and capture ordering. They use a synthetic Solitaire fixture and establish no TicTactics game progress. One HTTP fixture initially read the terminal record before the release child finished writing it; waiting for that terminal record fixes the test-side race.

GLtron ticket-lifetime and Golf actual waits, both EOFs, heavy release and absent exact PIDs have been personally checked before starting heavy work. The sole original followed the full locked build and explicit publication.

Original result: partial game, native Close

Code was published on main as b2210743410396469c7b6cbd30f2bedecd2b82e6 before the sole original. Its launch Git SHA was b06f21ddba0086afb4dc0a878138d1ad1e61799a; relevant dirty source hunks are retained as relevant-uncommitted.patch, and the initial owned new files as owned-initial.patch. No registry or runtime source change was warranted. The locked build passed in 38.410 seconds. It overlapped the final synthetic CLI controls briefly; the built regular and compat module hashes exactly matched the released Golf modules, and the original began after both waits.

The native process was permanently bound as PID 4122226, start token 274486002, main window 0x10001, screen 800×600. The fixed launch was 1200 seconds, real ticks and default batch size 1000, with an independent 1220-second watchdog and an irreversible final 180-second Quit reserve. No ordinary Start was needed: the initial native window already contained the untouched default game. No second game, restart, option change or rebind occurred.

Layer numbers run top to bottom; rows run back to front, columns left to right. Every unlisted cell remained empty. The model personally chose each placement after reviewing its complete current scene; there was no board reader, solver, tracking process or queued action list.

Input Own persistent effect Separate computer reply Evidence
Blue L2 row 2 column 2, pixel (266,175) Blue appears at that cell Red L4 row 4 column 4 choice8 review, move1.png, consumed receipt and UP ACK
Blue L2 row 2 column 3, pixel (287,175) Both Blue pieces persist Red L4 row 1 column 4; earlier Red persists choice12 review, move2.png, consumed receipt and UP ACK

Final complete board, B Blue/human and R Red/computer:

L1: .... / .... / .... / ....
L2: .... / .BB. / .... / ....
L3: .... / .... / .... / ....
L4: ...R / .... / .... / ...R

Blue was next. No native win, loss or draw was visible. Further placement proposals expired before intent creation or input, so a third move was never sent. Complete-scene presentation and tool turnaround repeatedly consumed the five-second freshness allowance. Some image presentations appeared blank despite byte identity with the complete retained PNG; this was not claimed as a blank native board or a new board state. Rather than relax freshness, the run ended as partial.

The complete native caption Close was personally reviewed from move2.png, which cmp proved byte-identical to the fresh quit2.png scene. The one Close click at (367,41) was consumed; its independently prearmed physical UP received ACK. The original then returned [Exit] code=0. Native PID 4122226, watchdog 4122225 and supervisor 4122223 have actual waits, both EOFs, and absent exact PIDs. The supervisor recorded heavy release, port 8627 was closed, and the independent deadline watchdog did not fire. This establishes native Close separately from host teardown. The outer interactive tool session returned 143; that ancillary value is preserved, not substituted for the actual native/supervisor waits. Native exit reported one held main critical section, so no all-thread-clean claim is made.

Evidence limits and retained artifacts

The 339 launch pins matched at final readback, including the registered 28,432-byte executable and unchanged SHA-256 above. The registered route has no explicit DLL closure. The runner nevertheless auto-loaded test/binaries/dlls/msvcrt.dll (278,581 bytes, SHA-256 887eb5ce93edb7192ca3e9220f07f9ca0f94db02af5862ebcbdfcb852db99fd1). It was omitted from the before-run pins: its post-run identity is recorded, but its before hash remains UNKNOWN. Thus the complete dynamically loaded media closure is not claimed to have a proved before/after match.

The synthetic parent-death orphan's unavailable wait remains UNKNOWN in the control evidence. It is not original-game outcome evidence. Redundant passive PNGs are removed with their hashes and selected-image aliases kept; only the initial scene, first reply and final partial board PNGs remain. The run contains compact logs, actual wait records, result JSON and Git patches; no source tree, media, WASM copy or frame bundle. No box was created (box ID/expiry null, state NOT_CREATED, stop receipt NOT_APPLICABLE), and /home retained over 1.5 GB free. The next heavy worker was explicitly released; this bounded task makes no further original input or launch.