Sil-Q launch readiness
Revision 0.8.0 (silq-session-v4) passes engineering validation and release
readiness verification. Its one-game Astra-max development pilot
completed naturally and passed transcript, native replay and admission checks
on the frozen images below. No final research cohort has been launched or selected.
Release evidence
Paths are relative to silq/; private evidence stays local.
| Check | Result | Evidence |
|---|---|---|
| Python suite | 366 passed | .private/release-20260917-v080/pytest-final-validation.txt |
| Static checks | Ruff lint/format and ty passed | Same directory: ruff.txt, format.txt, ty.txt |
| Native checks | 6,435 sanitized frames; 4,758 production and 612 fixture frames replayed; 1,480 hidden-state paired frames | native.json |
| Production Inspect | All five scorers, isolation, non-root execution and resource telemetry passed | inspect.json |
| Session expiry | Retired, timed-out and unstarted games retained with correct zero scores | timeout.json |
| Model cancellation | Synthetic in-flight generation canceled at the deadline; no billable request | model-timeout.json |
| Container recovery | Original deadline, workspace and exact request receipts preserved | recovery/report.json |
| Public walkthrough | 14 commands; 46 native frames replayed; no public-view differences | direct-play.json |
| Cost regression | Previous GPT-5.5 pilot reproduces $49.861837 after its session-wide uplift | gpt55-cost-regression.json |
| Aggregate engineering evidence | Hashes bound to the tested source and images | production.json |
| Astra max pilot | Native death; 969 exact replay frames; 427 verified requests; admitted | pilot-admission.json |
| Resource calibration | No memory/process limit events or provider limits in the pilot | resources.json |
| Release readiness | Source/image identity and evidence hashes verified | readiness.json |
The public walkthrough exercised birth, movement, named actions, typed item/menu replies and Tanglethorn combat. Its attack message and HP loss were both visible. It ended in deliberate retirement; it is an interface check, not a capability result.
The maintained templates pin these image IDs, which are also archived privately:
- Service:
sha256:d483fffa3f359ac21622c4587d2d43254cfa6f32df9ffe90ea9f239ecae0c9b4 - Client:
sha256:78d96d0608e130749586d13914d8e4888225765ac34337fa5a50351c09c22b75
Rebuilding produces a new identity that needs matching validation. Mutable :dev
tags are for local development and do not identify this release.
Budget and reporting contract
Native escape while carrying a Silmaril remains the only win. Under the frozen
success-within-budget-v1 policy, verified trial time/key exhaustion scores zero.
At verified session expiry, unfinished and unstarted games score zero while
previous wins remain. The protected journal proves expiry; a cached clock or
Inspect timeout alone is insufficient. Infrastructure/provider failures and
operator stops remain unadmitted. The tests cover exact-deadline endings,
interruption, expiry before birth, partial schedules and retained wins.
Deadline cancellation can explain trailing canceled model requests and unfinished calls in the final tool batch. It cannot excuse earlier errors or missing results. The native replay and qualitative coverage review still apply. Missing usage leaves total estimated cost unknown.
Reports keep native victory primary and include depth by game and elapsed time, Silmaril acquisition and carrying-ascent. Unstarted games have no trajectory. The main cost field uses frozen model-specific pricing: session-wide long-context uplift for GPT-5.5, per-request uplift for Astra. Estimates are not invoices.
Pilot and remaining launch work
The release pilot is silq-release-20260917-v080-astra-max-01: Astra max, no
learning games and one scored game, two hours per game and three hours for the
session, without a dollar cap. It died after 81m 59s, falling from 150 ft to a
native-recorded 250 ft at 7 HP. Estimated cost was $180.35. It is development
validation, not a saturation estimate or a medium/max comparison.
Follow the live pilot dashboard for public depth progress, game/session countdowns, estimated usage cost and review stages. The publisher runs as a persistent user service and refreshes every 15 seconds. Desktop/mobile browser checks and all 366 Python tests passed after adding the dashboard; the game source and frozen image identities are unchanged. The pilot's qualitative review is complete. Its complete public history matched native evidence, and 406 custom printed views were checked separately; archive coverage is not a claim that the model read every frame.
The older GPT-5.5 medium pilot ended in a natural death at 100 feet and passed admission for its own frozen build. Its source and images differ from this release, so it is not relabeled as current. Historical studies, exclusions and reviews retain their original contracts.
The final research schedule, sample size and matched medium/max budgets remain
unchanged proposals. After those choices, prepare a fresh plan and bind release
readiness to its experiment_sha256. Research max dispatch also requires the
prospective medium gate. See experiment commands.