7a25e5b6d54ea33fabe56b56a557ff3ac31d96f9c1ffbbf586a0de81598d66f9
WonderPhi Autonomous Compute — Auto-Batching Observer
Publicly online since 2010 · U.S. patent applications since 2012 · inventions offered since 2014. The work of Christopher Gabriel Brown, independently documented.
Verified performance — same runtime, zero friction
Autonomous Compute uses the exact same WonderPhi Compute runtime that delivered these measured results on a dual Xeon Gold 6138T + NVIDIA A10M:
- 14.1× on batch SHA-256 hashing (parallel across cores)
- 1.5× on 4K HEVC encoding via A10M NVENC vs. 40-core libx265
- 17.8× on AV1 by picking libsvtav1 over libaom-av1
The Autonomous variant’s value is that it detects serial workloads in your process list and dispatches these speedups without you clicking a thing. Same performance guarantees. Same seven-layer safety envelope. Full methodology in BENCHMARKS.md.
New here? Read this first — in plain English
What it is: the autonomous variant of WonderPhi Compute's observer. It watches your process list, and when it spots a serial workload it can safely parallelize, it runs the parallel plan itself — no drag-and-drop from you.
What the standard observer does: writes a suggested plan to a folder. You review it. You drag it into the inbox. Then it runs.
What this does: the review and drag step are removed. Detected patterns run themselves. You keep every other WonderPhi Compute guarantee — the runtime constitution, the 24-hour caps, the honest logs, the singleton locks. The only thing removed is your click.
Why buy it: if you already know which programs you launch, and you're tired of dragging plans out of wpc-hypotheses\ every time.
The promise: seven layers of safety must all hold before a single command auto-executes. Details below.
WonderPhi Autonomous Compute v1.0.0
The convenience layer for people who already trust WonderPhi Compute and want their machine to notice serial workloads and parallelize them without asking. Same detection engine as the standard observer — different last step.
The seven-layer safety envelope
Every one of these must hold before one single command auto-executes. If any layer fails, the plan is written to wpc-hypotheses\ for manual review — the standard observer's behavior. Autonomy is layered on, not replacing, the safety of the underlying system.
| Layer | Rule |
|---|---|
| 1. Explicit arming | Env var WPC_DANGER_AUTONOMOUS=1 must be set on the task's process. The installer sets it; you can unset it any time to instantly revert to suggest-only. |
| 2. Kill switch | A single file at %USERPROFILE%\bin\wpc-danger-STOP. Its presence instantly disarms auto-execution. Create it in one line to shut down autonomy without uninstalling. |
| 3. Whitelist | Only idempotent-safe programs auto-execute: certutil, sha256sum/md5sum, pdftotext, pandoc, magick/convert, curl, and 7z (extract-only verbs). Everything else — including ffmpeg, rm/del, git, python — falls through to suggest-only. |
| 4. Loop prevention | Commands whose first token is any WonderPhi binary never auto-execute. Cannot get into a feedback loop. |
| 5. Ledger dedup | Every command ever auto-executed is hashed (SHA-256) into a durable ledger. The identical command never auto-executes a second time. |
| 6. Cooldown | Same executable cannot be auto-batched more than once every 30 minutes. |
| 7. Batch cap | Auto-plans max out at 32 tasks. Larger patterns fall through to suggest-only for human review. |
The paper trail (nothing is silent)
Every decision — auto-executed, cooled down, whitelist-blocked, kill-switched, deduped — is logged to %USERPROFILE%\bin\wpc-observer-status.log with a [DANGER] prefix (the operational label; the store product name is softer, but when a command runs without your review, the log entry should shout).
Every auto-executed plan in wpc-inbox\ is prefixed auto-DANGER-… and carries a banner header that says, verbatim: "THIS PLAN WAS AUTO-EXECUTED. No human reviewed it before it ran." No plan is ever executed without leaving that paper trail.
Requires WonderPhi Compute
This variant replaces the standard observer that ships with WonderPhi Compute. The runtime itself is not modified — auto-plans still go through the same fireproofed watcher, same singleton lock, same honest reporting. The runtime's 10 Commandments stay intact. This variant only removes the review step.
Honest scope
- Windows only. Python 3.8+ available on the machine for the source variant; the frozen executable has no runtime dependency.
- Auto-execution covers roughly one third of the Recipe Library recipes — the idempotent-safe ones. The other two thirds still emit suggestions for manual review.
- Perpetual license — one payment, runs forever on your own equipment.
- Version 1.0.0 is permanent. Future Autonomous Compute releases are separate SKUs.
- USA-only sales. United States Dollars only. Email-only contact.
© Christopher Gabriel Brown 2026. All rights reserved.
