Seed SC-364 · deterministic local simulation
Tune speed against wrong answers. Every cache decision is replayable, every similarity point has a receipt, and every tenant boundary is enforced before a candidate can be compared.
Policy changes restart the same seeded workload.
Higher rejects more paraphrases; lower admits more near-misses.
Entries remain inspectable after expiry but cannot serve.
Keyboard: → step, Space play/pause, R reset, 1–3 presets.
Step or play the seeded workload to inspect the first decision.
No response yet.
Ground truth is attached to every seeded request.
A miss calls the simulated origin and writes its known-good answer. A hit is judged against that same ground truth, so speed cannot hide an incorrect response.
Select any row to restore its complete explanation.
The deterministic replay is ready.
Expired and namespace-fenced entries remain visible.
Origin misses will populate the cache.
This is a lexical teaching model, not an embedding service.
score = clamp(0.60 × weighted-token-Jaccard + 0.25 × ordered-bigram-Dice + 0.15 × character-trigram-Dice − number / negation / contrast penalties)
Semantic normalization is deterministic: Unicode accent folding, lowercase, apostrophe/dash/punctuation folding, a fixed synonym/phrase map, stop-word removal, then fixed token weights. Numbers weigh 2.4×, negations 2×, contrast words 1.8×, domain terms 1.35×, and other tokens of at least nine characters 1.2×.
Exact mode is stricter: it uses Unicode accent folding, lowercase, apostrophe/punctuation folding, and whitespace collapse — but no synonym map or stop-word removal. Semantic scores are still shown as diagnostics when exact mode misses.