You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$15,000 left to reach the funding goal.
I built Yankı (github.com/mevlanayalcin/yanki, MIT license) after running into the same failure mode over and over in retrieval pipelines: even strong reasoning models routinely invent plausible quotes, stitch two separate sentences together as if they were a verbatim passage, or cite the wrong document chunk when passages contradict each other. Most teams try to catch this by calling a second LLM as a judge, which adds latency, costs another model call, and often hallucinates right alongside the generator.
Yankı takes a purely deterministic route in standard-library Python. In src/yanki/normalize.py and src/yanki/verify.py, we maintain a strict 1:1 character-index map across Unicode casefolding, whitespace collapsing, and diacritic stripping. That matters in practice because naive Python .lower() turns characters like Turkish dotted capital İ into two codepoints (i + U+0307), shifting every downstream character offset in a document. Yankı resolves exact character spans, checks multi-source citation attribution, and triggers structured retry or abstention in under 0.3 ms per response. Right now the repository ships v0.2.0 with zero external runtime dependencies and a 24-case offline benchmark (python3 -m yanki.eval --offline --strict-thresholds). I also serve as Principal Investigator on EuroHPC LUMI project EHPC-AIF-2026PG01-1954 based in Ankara.
What I want to do over the next six months is scale our small 24-case testbed into an open 2,500-case adversarial grounding and citation-integrity benchmark across current frontier reasoning models (GPT-6 Astra, Claude, and Gemini Pro). Specifically, we will test three failure regimes where frontier agents still break down:
1. Contradictory multi-hop retrieval chunks where timestamps or numbers conflict across sources.
2. Indirect prompt-injection strings embedded inside retrieved documents that try to hijack citations or suppress abstention.
3. Long-context (100k+ token) multi-turn self-correction loops, measuring whether feeding deterministic span-verification errors back to the model actually fixes hallucinations or just makes the model rephrase ungrounded claims without quotes.
Every benchmark case, evaluation script, raw model trace, and failure scorecard will live in the public MIT repo at github.com/mevlanayalcin/yanki (project overview at mevlanayalcin.com.tr/work/yanki/).
Budget breakdown ($15,000 target, $5,000 minimum):
$12,000 goes directly to frontier API tokens (OpenAI GPT-6 Astra, Anthropic Claude, and Google Gemini Pro) to run multi-turn long-context sweeps across all 2,500 cases and publish reproducible baselines.
$3,000 covers adversarial corpus construction (English and morphologically rich Turkish legal/technical texts), human ground-truth span annotation checks, and hosting the public trace viewer.
If funded at the $5,000 minimum, I will scope the sweep to 1,000 core adversarial cases across the top three frontier models and release the full code and traces under MIT.