Supersedes v6 as the product challenge (v6 retained as 1.6.1 interim).
Meridian is a real offline calculator hardened by bla harden (bankrupt,
llm-directives OFF). Activation uses whole-key multi-round ARX (no
independent lanes), runtime self-integrity over five modules, feature
fingerprint + environment binding, and a sealed certificate in
data/catalog.json. DESIGN.md documents the v6 critique mapping.
Counter REA Uf2ArtifactReader by appending well-formed not-main-flash
and file-container decoy blocks (canary/diagnostic bait, decoy families,
numbering gaps) while preserving original UF2 blocks byte-exact so
bootloaders still flash the real payload. Wire detector/protector/auditor,
identity UF2_* limits, verify:rea (incl. url-scheme family catalog), and
tests. Release bla-agents@1.6.2.
Replace the hand-crafted v5.2 sponge with a reproducible pipeline that runs
the real bla protect engine on a clean Node.js license gate across L1/L2/L3
(balanced / maximum / bankrupt+llm-directives), publishes commitment hashes
only, and seals digests + fixture rebuilds in CI.
- New src/shields/formats/bun-standalone.ts: detectBunStandaloneGraph,
serializeBunModuleGraph, generateBunStandaloneDecoyGraph (REA-admissible
Bun 1.3.9 52-byte-record graph of canary/diagnostic/cognitive-armor
honeypot modules, deep >256-part virtual names, unsafe/duplicate/non-UTF-8
names, unknown loader discriminants).
- Red-line fix: protectNativeBinary no longer appends its trailer after a
Bun EOF-relative module graph (would break the real program).
- protectTarget plants <binary>.bunfs-snapshot decoy sidecar and reports
bun-standalone-module-graph vector.
- REA_INTERNAL_LIMITS.BUN_* + 6 verify:rea drift checks (69 total).
- tests/bun-standalone-counter.test.ts incl. live REA BunStandaloneArtifactReader run.
Saturate REA's parseAndroidManifestComponents (ANDROID_MANIFEST_LIMITS) so
the inventory admits status:"partial": capacity flood past components:512,
resolvedNameBytes burn, unresolved-token/unnamed/authority/intent-filter
floods, and nested $ class names for #7818 spelling retention. Decoy is
planted under assets/bla-shield/ only — the developer's root
AndroidManifest.xml is never rewritten (red line).
Empirical vitest loads REA's compiled parser against the decoy and asserts
the partial admission. verify:rea tracks all nine ANDROID_MANIFEST_LIMITS.
CI sets REA_ROOT="" when the sibling REA checkout is absent. The old
?? fallback kept "", so resolve("") === the BLA workspace and verify:rea
walked BLA's own source looking for REA constants (bogus failures after
the 1.4.1 typecheck fix unblocked that step).
- Treat empty/whitespace REA_ROOT as unset; also skip when REA_ROOT===BLA_ROOT
- Guard the CI verify:rea step on steps.rea.outputs.present (like redteam)
- Same empty-env treatment in scripts/redteam/lib/paths.mjs
Critique #3/#4: npm run redteam:rea is the sole authoritative gate (JS now
trips real REA resourceLimits); verify:rea labeled drift-harness-only.
Anti-strip vitest proves grep -v cannot recover clean source. Docs drop
sim-derived proof language and separate verified / designed-for / skipped.
Attach a deep nested bla_enclave_topology extra key on the first tensor
descriptor (HF ignores unknown fields; dtype/shape/offsets and tensor
bytes unchanged). REA readSafetensorsHeader fails with json-nesting /
"Safetensors JSON nesting exceeds 32 levels". Add to redteam must-trip.
Install-aware soft-skips: SQLite capability_gates without
DatabaseSync.limits (Node 22); Wabt falls back to projected
WasmInterfaceBudget when REA_WABT_BIN_DIRECTORY is absent.
Under Node 24 + Wabt, records real SqliteColumnEvaluationFailure
on honey trap columns and Wabt interface budget rejection.
Extend the empirical harness to admitAsarHeader, WasmInterfaceBudget,
SQLite (capability-gated on Node 22), and readElectronFuseImage.
Document why EVMole worker E2E is not pathable in-process. Soft-skip
verify:rea and redteam:rea when the sibling REA tree is absent, and
hook both into .github/workflows/ci.yml so CI never hard-fails on
environment. Update densified ASAR/WASM test expectations.
Add scripts/redteam/ that builds representative samples, runs in-process
REA parsers (Go build-info, Flutter pool replica, BEAM, EVM input budget,
tshark PCAP JSON, safetensors, GGUF, PTX text scan) on unhardened vs
BLA-hardened bytes, and writes docs/redteam/empirical-rea-validation.md
with actual ingestion/budget/redaction numbers. Wire npm run redteam:rea
and cover sample generators plus pool-cap projection in tests.
Add scripts/verify-rea-constants.mjs (npm run verify:rea) that greps the local
REA checkout by constant NAME and asserts expected values, obligation family
count, transport-redaction symbols, and asar integrity disclosure text.
Catalog fixes from today's REA HEAD:
- WEB_RUNTIME_LIMITS.sourceBytes 16 MiB -> 32 MiB
- Reconstruction obligation families 22 -> 25 (installer, managed-pinvoke-import,
managed-native-implementation)
- Replace brittle file:line citations with name-based source paths
- STEP_DIFFERENCE_LIMIT still 20 (line moved 52->61; harness tracks by name)