We audited 570 chapters of our own AI-written code. 41% couldn't prove they did what they claimed.

Showing 59 of 59

• crypto.bigintSTUB — whole chapter quarantinedCrypto & network trust surface

  A “bignum” that's a single 32-bit alloc(1) limb — no multi-limb / arbitrary precision at all, despite the name and header claim.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• crypto.certSTUB — whole chapter quarantinedCrypto & network trust surface

  cert_is_valid checks one hardcoded literal; cert_parse_serial ignores its own blob argument and returns a constant; the “verifying” signature is the RSA identity element, which trivially verifies for any key.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• crypto.ctSTUB — whole chapter quarantinedCrypto & network trust surface

  Header claims a real NIST AES-CTR test vector but main() never asserts the computed ciphertext against it; aes_block_simplified is self-documented as not implementing the full algorithm (one S-box substitution, no ShiftRows/MixColumns/key-schedule).

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• crypto.kdfSTUB — whole chapter quarantinedCrypto & network trust surface

  simple_kdf takes no password/salt parameter at all — hardcodes “password” internally. Self-documented as a “reduced/placeholder version for demonstration.”

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• crypto.keystoreSTUB — whole chapter quarantinedCrypto & network trust surface

  “Encrypt” and “wrap” are unconditional return 1 no-ops; “decrypt”/“unwrap”/“verify” are bare XOR, and every assert is a tautology against that same XOR.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• crypto.rngSTUB — whole chapter quarantinedCrypto & network trust surface

  A “CSPRNG” with no entropy source — a bare linear congruential generator, fully predictable. csrng_seed_demo() is a return 1 placeholder.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• crypto.sigSTUB — whole chapter quarantinedCrypto & network trust surface

  Defines a function literally named hmac_sha256 whose body is “XOR all bytes” — reusing a real primitive's name for something that isn't HMAC at all; rsa_pss_encode/decode are no-ops.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• crypto.tlsSTUB — whole chapter quarantinedCrypto & network trust surface

  Self-documented: “Toy cipher: uses XOR for encryption (obviously insecure)”, “Fixed toy MAC.” Pre-flagged by its own header before the audit even started.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• crypto.walletSTUB — whole chapter quarantinedCrypto & network trust surface

  “Public key derivation” is privkey + 39321 — plain integer addition, not EC scalar multiplication (admitted in the header). Several asserts recompute the code's own formula as their own “oracle.”

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• crypto.sym.chacha → chacha20poly1305_demo()STUB sub-function — excised, chapter stays REALCrypto & network trust surface

  A single dead placeholder function (return 1) cut out of an otherwise-REAL chapter — chacha20_block itself matches the RFC 7539 KAT exactly and stays live.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• L0.ty.unitsSTUB — whole chapter quarantinedL0 — core language & dev tools

  Every “unit operation” (meter, newton, joule, exponent math) is a hardcoded-constant-returning function, asserted only against the identical literal — nothing computes a derived unit from two operands.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_L0CORE_2.md

• L0.str.regexSTUB — whole chapter quarantinedL0 — core language & dev tools

  Zero real regex capability — no character classes, quantifiers, or alternation; every “match” is a hardcoded lookup on fabricated magic IDs, not real character comparison.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_L0CORE_2.md

• L0.io.fs_netSTUB — whole chapter quarantinedL0 — core language & dev tools

  fs_read_demo/net_get_demo ignore their path/url arguments and always return an identical hardcoded byte buffer — proven by the file's own test calling it with two different paths and getting the same bytes back.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_L0CORE_2.md

• L0.build.xcompileSTUB — whole chapter quarantinedL0 — core language & dev tools

  target_arch / target_os / target_bits / target_endian are each a hardcoded-constant return, and every assert is fn()==same literal — no real cross-compilation exists.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_L0CORE_2.md

• L0.dx.docSTUB — whole chapter quarantinedL0 — core language & dev tools

  doc_extract ignores its source pointer entirely and always builds the same hardcoded Docs struct; doctest_run_demo is a tautology over that same constant.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L0.dx.editorSTUB — whole chapter quarantinedL0 — core language & dev tools

  editor_open is self-documented “(stub)”; syntax/completion demos are hardcoded literal-count sums with zero real tokenization or completion logic.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L0.dx.notebookSTUB — whole chapter quarantinedL0 — core language & dev tools

  nb_eval_cell doesn't evaluate anything — it checks for the presence of + / * characters and returns a fixed constant regardless of the actual operands.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L0.dx.playgroundSTUB — whole chapter quarantinedL0 — core language & dev tools

  playground_run is a 4-case hardcoded lookup keyed on a magic source-hash; no real source is ever compiled or executed.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L0.export.targetsSTUB — whole chapter quarantinedL0 — core language & dev tools

  Own comment: “Mock target identifier tracking” — the Rust/WASM/JS “export demos” are three ordinary arithmetic functions labeled with target names, touching no real codegen path.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L0.gate.crossbackendSTUB — whole chapter quarantinedL0 — core language & dev tools

  Claims native-vs-GPU/PTX cross-backend verification but contains zero GPU/PTX invocation anywhere — single-backend arithmetic wearing the wrong label.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L0.import.moreSTUB — whole chapter quarantinedL0 — core language & dev tools

  import_ts / import_go / import_java all ignore their path argument and return fixed hardcoded Module structs (“Stub symbol table” × 3); no file is ever read.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L2.compress-zstdSTUB — whole chapter quarantinedL2 — data formats

  Claims Zstandard compression with a level-parameterized contract (REGISTRY.tsv), but the actual algorithm is plain byte-oriented run-length encoding — no LZ77 matching, no FSE/tANS entropy coding.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_2.md

• L2.flatbufSTUB — whole chapter quarantinedL2 — data formats

  Header claims a vtable/offset-indirection zero-copy format, but the functions are generic little-endian integer packing — none of FlatBuffers' defining mechanism exists.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_2.md

• L2.parquetSTUB — whole chapter quarantinedL2 — data formats

  Header claims “Parquet columnar storage”; actual content is a 4-field metadata struct with getters/setters — no column encoding, page/row-group structure, or footer schema.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_2.md

• L2.configSTUB — whole chapter quarantinedL2 — data formats

  config_load never loads any file (own comment: “we would load the file here”); config_get ignores its lookup key and returns a hardcoded string regardless of which key is requested.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L2.csvSTUB — whole chapter quarantinedL2 — data formats

  csv_emit/csv_get are explicit no-ops (“just demonstrate the function exists” / “return empty string for now”) — no field content is ever stored or retrievable.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L2.diff-patchSTUB — whole chapter quarantinedL2 — data formats

  patch_apply never applies the diff's recorded new character — it re-copies the original, so patching never changes the string, and the gate's own test never checks patched content.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L2.json-pointerSTUB — whole chapter quarantinedL2 — data formats

  jp_escape's fallback replaces every character with a literal 'x'; jp_get doesn't navigate any real JSON structure (“simplified — just use position for now”).

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L2.mapxSTUB — whole chapter quarantinedL2 — data formats

  map_has / map_get_val / map_get_or are confirmed total no-ops — always return “not found” / default regardless of whether the key was ever inserted, and the gate's own asserts embrace this as expected.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L2.sexprSTUB — whole chapter quarantinedL2 — data formats

  sexp_parse never builds a real list structure (just checks for a paren anywhere); sexp_emit writes zero bytes to its output — a no-op disguised as serialization.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L2.tomlSTUB — whole chapter quarantinedL2 — data formats

  toml_get_str / toml_get_int ignore their key argument entirely, returning the same hardcoded value regardless of key; toml_emit is just node-count × 25.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L2.yamlSTUB — whole chapter quarantinedL2 — data formats

  yaml_emit writes literal 'x' bytes unrelated to any real content; yaml_get is return count, ignoring the key argument entirely.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L2.jsonSTUB — flagged, kept (has a live dependent)L2 — data formats

  json_get_str / json_get_int / json_emit are hardcoded-constant returns regardless of arguments — self-flagged by jsonparse.q's own header. Has one live (but unused) dependent, so left in place and flagged rather than deleted.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• L3.distributions → normal_sampleSTUB — whole chapter quarantinedL3 — numerics, audio, model I/O

  Header claims Box-Muller Gaussian sampling, but the code computes the real quantity and then discards it (“let me use a simpler approximation... placeholder”) — cos() is never applied to the phase term.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_L0-L3.md

• L3.dsp → stft / mel_filterbank / mfcc_demoSTUB — whole chapter quarantinedL3 — numerics, audio, model I/O

  All three header-listed capabilities are self-documented no-ops: stft() — “For now, just return without computation”; mel_filterbank() — “just fill output with zeros”; mfcc_demo() — “Placeholder.”

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_L0-L3.md

• L3.geometry → ray_tri / in_circumcircleSTUB — whole chapter quarantinedL3 — numerics, audio, model I/O

  geom_intersect_ray_tri ignores all 5 arguments and always returns 5 (“Placeholder... would compute Möller-Trumbore”); in_circumcircle sums coordinates instead of any real Delaunay predicate.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_L0-L3.md

• L3.init → he_normal_demo / mup_init_demoSTUB — whole chapter quarantinedL3 — numerics, audio, model I/O

  Both functions run a full sampling loop into a local buffer, then discard it and return an unrelated hardcoded literal (42 and 99) — the gate's own asserts are tautological.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_L0-L3.md

• L3.model-ioSTUB — whole chapter quarantinedL3 — numerics, audio, model I/O

  model_save never persists to its own path argument — builds a buffer and discards it (“Simulate save”), even though sibling chapters in the same codebase prove the real mechanism exists and is simply unused here.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_3A.md

• L3.ssmSTUB — whole chapter quarantinedL3 — numerics, audio, model I/O

  Header claims linear/selective state-space dynamics with a scan and convolutional form; the actual content is trivial linear interpolation — no state recurrence, A/B/C matrices, gating, or conv-form.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_L0-L3.md

• L3.symbolic → eval_exprSTUB — whole chapter quarantinedL3 — numerics, audio, model I/O

  Header claims differentiation, but no derivative function exists — the “Differentiation” gate tests just assert expr_const(0)==0 / (1)==1, tautologies unrelated to any diff logic.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_L0-L3.md

• L3.tokenizer → bpe_train_demo()STUB sub-function — excised, chapter stays REALL3 — numerics, audio, model I/O

  A single dead placeholder (return 0, self-documented) cut out of an otherwise-REAL chapter — the whitespace-split + vocab-lookup tokenizer itself stays live and gate-green.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• L4.imgcodecSTUB — whole chapter quarantinedL4 — graphics & media

  Header: “image codec (DCT)”. dct_1d_simple computes a cosine-looking variable, then never uses it — the sum is just raw input values. No DCT exists anywhere.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• L4.fft-sharedSTUB — whole chapter quarantinedL4 — graphics & media

  Own comment: “Real implementation would do full DIT, this is a compute interface.” Only scales by a constant — no butterfly/DFT, no FFT transform exists.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• L4.spatialaudioSTUB — whole chapter quarantinedL4 — graphics & media

  main()'s own comments: “Test 1-5: SKIP panning tests” / “Test 6: ITD samples - SKIP.” Variables are hardcoded instead of calling the file's own headline functions at all.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• L5.canvasSTUB — whole chapter quarantinedL5 — web & platform

  canvas_fill_rect / stroke_rect / fill_circle are unconditional return 1 no-ops — there is no pixel buffer anywhere in the struct.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• L5.storageSTUB — whole chapter quarantinedL5 — web & platform

  storage_get_item ALWAYS returns the most-recently-added item's value regardless of the key argument — the gate's own test literally documents this as “correct” behavior. No real key-value lookup exists.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• L5.text-shapingSTUB — whole chapter quarantinedL5 — web & platform

  is_rtl() is defined but never called; bidi_reorder_demo does no reordering (“we verify string concatenation” instead); analyze_direction never detects direction, just counts non-space characters.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• L6.sdmmc-flashSTUB — whole chapter quarantinedL6 — hardware & embedded

  sdcard_read_block / write_block perform no I/O whatsoever — never touch the buffer, no SPI/MMC command framing; sdcard_get_status() is a hardcoded return 256.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• L6.accel-npuSTUB — whole chapter quarantinedL6 — hardware & embedded

  ~18 hardcoded constant-getters (core_count()=4, peak_throughput()=8192000, ...) plus range checks — no instruction dispatch, data-path config, or compute of any kind exists.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_TAIL.md

• L6.filesystem-embeddedSTUB — whole chapter quarantinedL6 — hardware & embedded

  fs_write returns the requested length but never writes a byte; fs_read never reads anything; fs_open always returns 0 regardless of whether the file exists.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_TAIL.md

• L6.image-edgeSTUB — whole chapter quarantinedL6 — hardware & embedded

  edge_sobel_int — the function literally named after the Sobel operator — never reads a single pixel from its image argument; no Gx/Gy gradient kernels exist.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_TAIL.md

• L6.nostd-runtimeSTUB — whole chapter quarantinedL6 — hardware & embedded

  nostd_malloc returns the exact same hardcoded address on every call regardless of prior allocations, despite a comment claiming “monotonic addresses” — not a working allocator.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_TAIL.md

• L7.chain.consensusSTUB — whole chapter quarantinedL7 — chain, storage & verification

  The entire file is two integer comparisons (pow_valid = hash_byte < target; longest_chain = a_len > b_len) — no hashing, mining loop, or block/chain data structure anywhere.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_2.md

• L7.chain.ledgerSTUB — whole chapter quarantinedL7 — chain, storage & verification

  validate_utxo_exists / ledger_record_tx are hardcoded if/else chains recognizing only the exact literal values baked into the test — no backing UTXO set exists; ledger_total_value() is a bare return 100.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_2.md

• L7.store.casSTUB — whole chapter quarantinedL7 — chain, storage & verification

  Own docstring: “cas_get(...): retrieve by hash (stub).” cas_put hashes data but never stores the bytes anywhere — nothing is ever retrievable.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_2.md

• L7.store.encryptedSTUB — whole chapter quarantinedL7 — chain, storage & verification

  The entire “encrypted vault” cipher is a single repeating XOR byte-key — the weakest possible cipher, trivially broken, the same failure mode as the already-quarantined crypto.tls.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: STUB_SWEEP_TAIL.md

• L7.verif.fheSTUB — whole chapter quarantinedL7 — chain, storage & verification

  File's own comment admits the chosen cipher is “not a truly homomorphic scheme.” The verification functions are tautological, never testing the actual claimed property.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• L7.zk.circuitSTUB — whole chapter quarantinedL7 — chain, storage & verification

  pedersen_commit is plain integer packing (secret + blinding×256), no EC/discrete-log hiding; range proofs are bare bounds-checks on the plaintext value — zero privacy despite “zero-knowledge” being the file's entire premise.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md

• L7.verif.teeSTUB — whole chapter quarantinedL7 — chain, storage & verification

  tee_verify_quote / tee_verify_cert are bare !=0 null-pointer checks with zero cryptographic signature verification; tee_verify_boot compares against a hardcoded magic constant.

  Tool: differential-oracle audit (qc2 --check + independent verification) · source: QUARANTINE.md