Known failure modes
Every empirically-observed failure of the universal-substrate reasoning compiler, with context + workaround where one exists. Updated continuously. If you find a new failure mode, submit it via challenge mode — we add it to this list within 48h of confirmation.
Boolean SAT
Gradient search incompleteness at N≥25
90% accuracy (9/10) at N=25, M=100 random phase-transition
Continuous-relaxation gradient descent on the Cl(6) G1 subspace is not complete. Multi-restart helps but doesn't reach 100% on harder instances. Brute force handles N≤20 at 100%.
workaroundDPLL branching when gradient plateaus (deferred — paper-4 lane).
source: V4 paper §3.2 / gio #7767
Chemistry / logP
Zero-shot logP weaker than trained baselines
Pearson r = 0.581 zero-shot vs ChemBERTa-trained r ≈ 0.65-0.7
Zero-shot encoding of SMILES → Cl(3,0) multivector handles MW (r=0.998), TPSA (r=0.979), Tanimoto similarity (Spearman 0.492 vs Morgan/RDKit) cleanly. logP requires more domain-aware features than the current G0+G1+G2+G3 encoder captures.
workaroundAugment Cl(3,0) with electronegativity-contrast scalar (G0 axis pending — gio queue 1.1.i).
source: V4 paper §3.3
TSPLIB / TSP
Solution length above optimal
11.8% above optimal on EUC_2D instances
Continuous-relaxation routing on the manifold doesn't match specialized concorde/LKH solvers. This isn't a problem class where geometric energy minimization should beat decades of dedicated combinatorial-optimization work.
workaroundIf TSP is your problem, use concorde. We benchmark TSPLIB for honesty about scope, not as a competitive endpoint.
source: V4 paper §3.5
Self-referential language
Operand-tier composition on "are you sentient" class queries
Subjectively brief / operand-grammar prose
dydact's chat path produces dense operand-tier compositions ("Therefore, rule although fixes.") on self-referential queries because no first-person prose exists in the training corpus to retrieve from. Corpus-anchored factual queries (consciousness, Banach contraction) produce coherent prose.
workaroundAdd first-person prose to the corpus (operator effort) or wire operand_translator as write-side post-processor (~30-60 min spec).
source: Day 56 fluency-collapse audit / gio #7566
Cuprate within-family discrimination
Jaccard 0.00 across LSCO/YBCO/Hg-1223 anchors
Top-15 ranking diverges per anchor within same superconductor family
3-feature encoder (d-electron count + coord planarity + magnetic moment) handles cross-family discrimination well (FeSe → 5 iron-pnictides, 0 cuprate intrusion) but cannot distinguish cuprate-family members from each other.
workaroundCrystal-system one-hot hard constraint. Direction for paper-4: layered-2D detector + apical-O distance encoding.
source: OPERATOR_MANUAL §4.25 / V4 paper §4.1
API / production
Apple Developer Agreement currently expired
ACI desktop DMG blocked from notarization
xcrun notarytool returns 403 — operator needs to log into developer.apple.com → Agreements section + accept pending agreement. ~5 min on Apple side.
workaroundSigned-but-not-notarized DMG available for early adopters (Gatekeeper warning on first launch).
source: Day 57 build-aci.sh status
Why we publish this
Every claim in the V4 paper has a verifiable empirical basis + a stated failure mode. Owning gaps publicly is more credible than implying universal dominance. Reviewers replicate this list as part of falsification — if you find a mode not on this page, you have grounds to flag it as undocumented.
Submit a failure case