Repository navigation
Description
[Filed on your flagship SDK repository as your organisations public issue surface; the message is addressed to AltTester as an entity, not to this SDK.]
The one thing asked of you specifically, if anything: rastu lint --counts-only <your-project-root> runs entirely on your machine and prints per-rule
per-boundary-class counts with no project, path, person or source identifiers —
the payload shape is frozen in the pre-registration (§11). It counts for the
measurement only on evidence you supply: the payload you paste back, or a
written statement that you ran it. No code leaves your machine either way.
6. Group E — agent-QA vendors, LAST (entities 27–29)
Opening for these three messages (modl.ai · GameDriver · AltTester), then the common block:
You are contacted last and once, per the project's own pre-registered ordering
rules, because a vendor contact is one-shot. To be explicit: nothing is being
sold to you and nothing is being asked of you commercially, now or later — this
is a research hand-off of a published negative result in automated determinism
checking for games, and the three questions below are the whole of it.
What this is: a pre-registered, published negative result, handed to you because your
project (or work) is named in the corpus. The rastu lint ran thirteen static
determinism-threat rules (RG-D01 … RG-X01) over 24 pinned open-source Godot 4
projects in regex parse mode, each with a declaration set authored only from that
project's own documentation. The pre-registration (thresholds, sampling frame,
refutation sentence) was posted publicly before the run:
https://gist.github.com/Curtynap/618082ee70ac5ae51cfb903c29cc9097
The full artefact set — the 24-row table (Markdown/JSON/CSV), per-finding JSON, the
18-row accountability appendix, the methodology and the write-up — is at
https://github.com/Curtynap/Rastu.
The result was NO-GO under the posted criteria, and the write-up says so in its own
words: the blinded audit did not survive its thresholds (precision 34/100 against
a ≥60% bar; unknown share 98.5% against a ≤40% ceiling), because 23 of the 24
declaration sets are honestly empty — the classifier refuses to guess, exactly as
pre-registered. A lint finding never proves a game unreplayable, and a clean run
never proves one replayable; the table is one point on a curve and its raw counts
track lines of code. Nothing in it is a verdict about your project, and the per-rule
vocabulary, thresholds and refutation sentence were frozen in public before any
number existed.
Three questions — a substantive reply to any one of them is the only external
signal this measurement can get, and they were fixed before the run:
- Does any finding in this table describe a defect you have actually hit?
- Is the sim/presentation split as we drew it wrong for your project, and where?
- Would you run this in CI, and if not, what stops you?
No ask of any other kind rides this message — nothing to buy, nothing to sign, no
promotion, and this is the only message you will get from this project. If you
cannot run anything, that is fine: the binaries are unsigned, and a SmartScreen /
Gatekeeper block is recorded on our side as an infrastructure non-response, not a
decline (a cargo install and source-build path is documented in the release notes).
Amendment note, pre-registered before any clock runs: the posted pre-registration
evaluates the no-reach flip and the F2 refutation sentence "at G2-published"; under
the NO-GO stop branch there is no G2 to publish, so the substitute trigger is fixed
here, in the hand-off issues themselves, before the last of them is filed — both
evaluate at six months after the last of the 29 named G1 hand-offs is filed.
Activity