security(adversarial-review): ADVERSARIAL_REVIEW_MASTER.md

This commit is contained in:
Allaun Silverfox 2026-06-23 04:56:57 -05:00
parent a2215a3100
commit 4a67a05e0b

View file

@ -0,0 +1,217 @@
# ADVERSARIAL REVIEW — SilverSight Full Project Audit
**Date**: 2026-06-23
**Doctrine**: Guilty Until Proven Innocent
**Reviewers**: 5 specialized adversarial agents
**Scope**: ~130 source files across Lean 4, Python, WGSL, HTML5/JS
---
## EXECUTIVE SUMMARY
The adversarial review found **156 issues** across 5 domains:
| Domain | Critical | High | Medium | Low | Total |
|--------|----------|------|--------|-----|-------|
| Formal Verification | 5 | 8 | 10 | 5 | 28 |
| Code Implementation | 11 | 18 | 27 | 15 | 71 |
| Mathematical Foundation | 5 | 7 | 7 | 5 | 24 |
| Systems Integration | 4 | 9 | 7 | 2 | 22 |
| Cryptographic Security | 5 | 8 | 10 | 5 | 28 |
| **UNIQUE (deduplicated)** | **18** | **28** | **42** | **25** | **113** |
**Verdict**: The project is a promising research prototype with genuine insights
(particularly in Sidon sets, codec consistency, AVM core, and the eigensolid
conditional theorem) but is **NOT production-ready**. The flagship claims
(Chentsov's theorem, NP-hard solving, Finsler metric) are not substantiated.
---
## TOP 10 CRITICAL ISSUES (Project-Threatening)
### CR-1: Chentsov's Theorem is NOT Proven
**Files**: `ChentsovFinite.lean`
**Finding**: The flagship claim proves only that a functional equation H(t)=c/t
has unique solution, and that Fisher metric is invariant under splittings. It
**never connects** Chentsov invariance to the metric structure. The final steps
use `rfl` (reflexivity) where genuine mathematical arguments are required.
**Impact**: The project's unique selling point is unproven.
**Fix**: Write the connecting argument from Chentsov invariance to the
functional equation, or restate the theorem with its actual scope.
### CR-2: 7 Axioms Assert Deep Unproven Number Theory
**Files**: `PVGS_DQ_Bridge/` section files
**Finding**: Baker's theorem, BMS bounds, Goormaghtigh conjecture are stated
as axioms. If any axiom is inconsistent or misstated, the entire bridge collapses.
**Impact**: Potential inconsistency in the formal system.
**Fix**: Replace axioms with proven bounds or clearly mark as conjectures.
### CR-3: SQL Injection in Citation Module
**File**: `python/hachimoji_citation.py`
**Finding**: User-provided equation strings are concatenated directly into SQL
queries via naive `replace("'", "''")`. PoC: `"1=1'; DROP TABLE arxiv_papers; --"`
**Impact**: Arbitrary SQL execution on the citation database.
**Fix**: Use parameterized queries (psycopg2 `%s` placeholders).
### CR-4: Receipt System Trivially Forgeable
**File**: `formal/RRCLib/ReceiptCore.lean`, `python/integration_sprint.py`
**Finding**: Receipt `valid: Bool` field has NO cryptographic validation. An
adversary sets `valid := true` and all checks pass. SHA256 truncated to 24 hex
chars has birthday bound of ~2^48, not 2^128.
**Impact**: Entire trust system is forgeable.
**Fix**: Add HMAC-SHA256 signature with secret key, use full 64-char SHA256.
### CR-5: Memory Exhaustion via Unbounded Enumeration
**Files**: `python/dna_gpu.py`, `dna_lut.py`, `dna_qubo_sort.py`
**Finding**: All enumerate `2**n_vars` without bounds checking. At n=30, allocates
256 GB.
**Impact**: Denial of service / OOM kill.
**Fix**: Add max_n_vars guard (e.g., 20), streaming computation for large n.
### CR-6: Q16_16 Arithmetic Precision Loss
**File**: `python/pist_braid_bridge.py`, `python/q16_canonical.py`
**Finding**: `q16_mul()` in bridge uses `//` (floor division), canonical uses
`round()`. Lean uses `floor`. Three variants produce different results by ±1 LSB.
At boundary values, `float_to_q16(32767.5)` differs by 32,767 LSBs.
**Impact**: Cross-module numerical drift.
**Fix**: Single canonical Q16_16 library, exact integer arithmetic throughout.
### CR-7: Finsler Metric Fails Finsler Axioms
**File**: `qubo/finsler_metric.py`
**Finding**: Claimed Randers metric F = α + β fails homogeneity (depends on
endpoints, not tangent vectors), requires post-hoc `max(F, 1e-10)` for positivity,
and does not enforce |β| < α.
**Impact**: Mathematical foundation of QUBO solver is invalid.
**Fix**: Implement proper Randers metric with tangent vector arguments and
enforce |β| < α.
### CR-8: Spectral Profile Has Trivial Collisions
**File**: `python/spectral_profile.py`
**Finding**: Profiles computed from byte-level statistics only. Confirmed
collisions: `"a+b=c"`, `"x+y=z"`, `"x*y=z"`, `"p/q=r"` all produce identical
8D profiles.
**Impact**: Classifier can be poisoned by adversarial input crafting.
**Fix**: Add semantic features or cryptographic hash to break collisions.
### CR-9: Eigensolid Convergence is Enumeration, Not Proof
**File**: `formal/CoreFormalism/BraidEigensolid.lean`
**Finding**: The convergence theorem checks 8 cases via `native_decide`, not a
mathematical proof. The `receipt_invertible` theorem is a tautology:
`(A==B) ↔ (valid A ↔ valid B)` proves nothing about bijection.
**Impact**: The compressor correctness claim is unsubstantiated.
**Fix**: Prove convergence by induction, or clearly mark as verified-by-computation.
### CR-10: DNA Encoding Fully Reversible (No Encryption)
**Files**: `python/dna_codec.py`, `formal/CoreFormalism/HachimojiCodec.lean`
**Finding**: Perfect bijection with no key, no salt, no encryption. Any
intercepted DNA sequence leaks its entire payload via simple lookup table.
**Impact**: Zero confidentiality for encoded data.
**Fix**: Add keyed permutation layer or encrypt-then-encode for sensitive data.
---
## DOMAIN-SPECIFIC SUMMARIES
### Formal Verification (28 findings)
- **23 sorries** across the project (proof gaps)
- **7 axioms** (potential inconsistency risks)
- **30+ native_decide** (computation, not proof)
- **293 theorem declarations** but many are tautologies
- **What IS proven well**: SidonSets (Lindstrom bound), HachimojiCodec
consistency, SilverSightCore AVM theorems, InteractionGraphSidon CRT theorems,
BraidEigensolid.eigensolid_trivial (conditional)
- **What's NOT proven**: Chentsov uniqueness, eigensolid convergence,
receipt bijection, PVGS bridge theorems
### Code Implementation (71 findings)
- **8% test coverage** overall, 23% of tested Python
- **19 modules with ZERO tests** (integration pipeline, chaos game, all WGSL,
all HTML5/JS, citation layer)
- **SQL injection** in citation module
- **Command injection** via subprocess
- **WGSL race conditions** in all shaders
- **WebGPU O(n²) dispatches** for large datasets
- **Q16_16 precision loss** in Python→float conversion
- **Massive code duplication** (qubo_energy 5x, BASES 6x, int_to_dna 3x)
### Mathematical Foundation (24 findings)
- **Chentsov theorem is vacuous** — rfl where proof needed
- **Finsler metric fails axioms** — not actually a Finsler metric
- **NP-hard claims unsubstantiated** — total work exceeds brute force
- **Eigensolid_trivial is genuinely correct** — a real conditional theorem
- **Phi-corkscrew bijection is valid** — ψ/2π irrational gives injectivity
- **Fisher-Rao distance formula is correct** — arccos(Σ√(pᵢqᵢ)) is standard
### Systems Integration (22 findings)
- **Lean↔Python Q16_16 divergence**: `//` vs `round()` vs `floor`
- **"Spectral profile" means 3 different things** across 2 languages
- **Python receipts and Lean ReceiptCore have ZERO common fields**
- **No end-to-end integration test exists**
- **DNA alphabet ordering is actually CONSISTENT** (ABCGPSTZ everywhere) —
the "known bug" was previously fixed
- **GPU buffer formats differ** between JS and Python paths
### Cryptographic Security (28 findings)
- **Receipt forgery**: trivial (valid: Bool with no crypto)
- **DNA encoding**: fully reversible with public LUT
- **Spectral collisions**: confirmed with 4 distinct equations
- **Chaos game LCG**: fully predictable from hash
- **Q16_16 div-by-zero**: returns MAX_RAW silently
- **WGSL buffer overflow**: unbounded `sequences[seq_idx]`
- **Cross-mode agreement bypass**: 50% difference still "agrees"
---
## WHAT IS GENUINELY GOOD
Despite the issues, several components are solid:
1. **SidonSets.lean** — Genuine formal mathematics (Lindstrom bound, Johnson
bound, Singer theorem infrastructure)
2. **HachimojiCodec.lean** — Honestly proven consistency theorems
3. **SilverSightCore.lean** — All TIC/AVM theorems correctly proven by induction
4. **BraidEigensolid.eigensolid_trivial** — Real conditional theorem with
complete proof
5. **Integration sprint** — All 4 modes agree (CV=0.0), all checks pass
6. **Φ-corkscrew bijection** — Mathematically valid (irrational rotation)
7. **Golden spiral recovery** — Bijective encoding with exact roundtrip
8. **BraidSpherionBridge** — The braidCross ↔ Mountain.merge correspondence
lemma is well-structured
---
## RECOMMENDATIONS
### Immediate (Block Production)
1. Fix SQL injection in `hachimoji_citation.py`
2. Add bounds checking to all `2**n` enumerations
3. Unify Q16_16 arithmetic to single canonical implementation
4. Add HMAC to receipt system
### Short-Term (Before Public Release)
5. Write the missing Chentsov connecting argument or restate theorem
6. Replace 7 axioms with proven bounds or conjecture markers
7. Add keyed encryption layer to DNA encoding for sensitive data
8. Write end-to-end integration test
9. Achieve 80%+ test coverage for critical paths
10. Fix Finsler metric to satisfy axioms
### Long-Term (Research Direction)
11. Replace `native_decide` proofs with actual proofs
12. Prove eigensolid convergence by induction
13. Close the 23 sorries
14. Add semantic features to spectral profile to break collisions
15. Formalize the Φ-corkscrew bijection in Lean
---
## REVIEWER FILES
| Reviewer | File | Lines |
|----------|------|-------|
| Formal Verification | `ADVERSARIAL_REVIEW_FORMAL.md` | 696 |
| Code Implementation | `ADVERSARIAL_REVIEW_CODE.md` | 1,003 |
| Mathematical Foundation | `ADVERSARIAL_REVIEW_MATH.md` | 723 |
| Systems Integration | `ADVERSARIAL_REVIEW_SYSTEMS.md` | ~800 |
| Cryptographic Security | `ADVERSARIAL_REVIEW_CRYPTO.md` | 785 |
| **Master Synthesis** | **ADVERSARIAL_REVIEW_MASTER.md** | **This file** |