Prompts are just suggestions. We build tools that verify AI code before you save.

Five Computer-Science Ideas Explained Simply (With Published References)

Most AI coding plugins either wrap a chat prompt or search text paragraphs by keyword similarity. Resonant applies five proven software-engineering techniques directly to your code in under 2 milliseconds at zero extra LLM token cost:

• 1. Running Code on Tricky Inputs vs. Asking a Second AI to “Guess” (resonant_verify_cegar): Asking a second AI model to review code takes seconds, costs extra API tokens, and often misses the same edge cases (Peng et al., CWEval 2025; Vero et al., BaxBench 2025). Instead, Resonant tests your modified function in an isolated memory sandbox against boundary inputs (0, -1, None, empty lists []) and compares old vs. new behavior before saving—an application of Counterexample-Guided Abstraction Refinement (Clarke et al., CAV 2000).
• 2. Mapping Callers & Git History vs. Chopping Files into Paragraphs (resonant_architect_orient): Standard “RAG” search chops files into arbitrary text chunks and misses callers in other files. Resonant maps each function’s inputs, state changes, and return guarantees (Hoare, CACM 1969; Weiser, IEEE TSE 1984) alongside files that historically changed together in your Git commits (Zimmermann et al., IEEE TSE 2005)—cutting context token usage by ~87%.
• 3. Notes That Expire Automatically When Code Changes vs. Outdated Text Memory (resonant_memory_jtms): Standard AI memory tools suffer from “memory rot” when code changes but old notes remain (Wu et al., LongMemEval ICLR 2025). Resonant links every project note to a SHA-256 fingerprint of the actual code structure using a Justification Truth Maintenance System (Doyle, Artificial Intelligence 1979): when the function changes, stale notes automatically flag as STALE_AST_DRIFT.
• 4. Turning Bug Fixes into 0-Token Code Checks vs. Bloated Prompt Files (resonant_evolve_guardrail): Instead of adding more text warnings to a prompt file that the AI eventually ignores, Resonant generalizes the line that changed between the broken and fixed code into a reusable pattern (Plotkin, 1970; Rolim et al., ICSE 2017) and saves a <1ms local check that costs 0 prompt tokens on later edits.
• 5. Fact-Checking Completion Summaries Against Real Receipts (resonant_honesty_audit): Breaks the AI’s final summary into individual claims (Min et al., FActScore EMNLP 2023) and checks each one against real file paths, symbols, and test exit codes (Necula, Proof-Carrying Code, POPL 1997). See Published Literature & Architecture for full citations.