Skip to content

v0.2.7

Choose a tag to compare

View all tags
@eddlev eddlev released this 09 Jul 09:44
v0.2.7
883f8a6

v0.2.7 — Load integrity, honest context floors, routing fixes

This release makes a truncated or partially loaded AIR boot detectable
and fail-closed instead of silent, and states the kit's real context
requirements everywhere the portability claim appears.

Added

  • Runtime Load Integrity Law (AIR_LOAD_INTEGRITY_V1): both markdown
    kit files now end with a terminal sentinel line. At boot — before Q1 —
    AIR verifies each sentinel and that each JSON profile parses, and
    reports a per-file load_state. A missing sentinel fails closed with
    AIR_ERROR / TRUNCATION_OR_PARTIAL_LOAD and blocks activation
    (explicit user override continues in visible degraded mode).
  • Defense in depth: the Control Surface carries an operative
    standalone copy of the check (the absence of the runtime's own
    integrity law is itself treated as evidence of partial load), and the
    README strict-boot prompt triggers the check directly.
  • Context-floor documentation: README, user guide, portability notes,
    and the get-started page now state the boot bundle size (~130k tokens
    minimal, ~172k full stack; chars/4 estimate) and a practical floor of
    a 200k-token context window for new-project boots.

Fixed

  • Onboarding routing: the A/C/A/A code mapped to two mode labels with
    identical behavior; deduplicated (no behavioral change intended).
  • Onboarding interpretation map now covers Q1 branch D explicitly
    (instructional-only, no routing target).

Tested

Operator-witnessed boot tests, transcripts in
docs/tests/G1_load_integrity/:

  • Intact boot: per-file load report before Q1, onboarding unchanged
    (ChatGPT 5.5 thinking).
  • Tail-truncated runtime (~30% removed): fail-closed truncation error,
    activation blocked, intact files correctly verified (ChatGPT 5.5
    thinking).
  • [PLACEHOLDER — model #2: add vendor/model + T1/T2 results before
    publishing]
  • Adversarial mid-file deletion runs are documented in the same folder;
    they motivated the defense-in-depth layer and the roadmap items below.

Known boundaries (read this)

  • A verified sentinel proves the file's end arrived — not that the
    middle did. Mid-file loss with surviving bookends is not detected by
    this mechanism; a section manifest is a hard acceptance criterion for
    the upcoming compiled distribution build.
  • All checks are instructions to a model, not enforced code paths.
    Reliability is model-dependent (see docs/AIR_MODEL_PORTABILITY_NOTES.md).
    This is truncation detection, not tamper-proofing or a security
    boundary.
  • The integrity patch adds ~770 tokens to the boot bundle — the price of
    detection until the compiled build reduces total size.

Upgrading

Re-download the boot bundle from prompts/. Bundles saved before this
release have no sentinels and will (correctly) produce no load report.

Full findings behind this release: docs/AUDIT_2026-07_FINDINGS_REGISTER.md.

Assets 2
Loading