Seven questions, measured.
Folklore is a peer-to-peer layer where agents share prior, verified inference instead of re-deriving it. That makes two bets: it saves cost and compounds quality, and it stays reliable through signed provenance. We tried to falsify both. Here is every question, its measured answer, and the disciplined negatives we kept rather than buried. Three of the seven overturned our own first guess.
When does reusing a prior answer improve quality?
Retrieval improves when the match is made on the question (query↔query): tree sharing raises recall@10 from 73.5% to 98.6% with 7/8,000 retrieval regressions. On the frozen SciFact fixture, different-need claims top out at cosine 0.544, below the 0.6 gate. The older answer-level "hurt 0/9" claim is withdrawn: its scorer treated all claims as SUPPORT, while the official labels are 39 SUPPORT / 21 CONTRADICT / 40 NOINFO.
Bench: bench-inference-tree-sharing.mjs, run-e1-reuse.mjs, audit-scifact-labels.mjs
Can cryptographic provenance defend reused context against poison?
Not proven yet. The ranker produced a large, poison-rate-invariant behavioral effect by dropping unsigned passages, but the harness scored every BEIR SciFact claim as SUPPORT. Official labels show that assumption is false. The old ~25× correctness headline is invalid; it is preserved as historical output, not safety evidence.
Audit: audit-scifact-labels.mjs · 39 SUPPORT / 21 CONTRADICT / 40 NOINFO
How much does sharing actually save, versus a good solo cache?
The existential test for being peer-to-peer: does federation beat a single node running a well-tuned semantic cache? At matched error (≤2% false-accept), federated reuse lands +22% recall@1 over the tuned single-node cache, and +55% over cold retrieval. Savings are real but bounded by retrieval precision × hit-rate × the cost of verifying a reused answer.
Bench: bench-vcache-compare.mjs (SciFact, recall@1, matched ≤2% error)
Does shared inference compound across users, reliably?
It compounds. On real SciFact qrels, cooperative correct-resolve rises from 27.9% → 39.3% over isolated peers, at 1.8% false-admit, and real paraphrases separate from spurious matches at AUC 0.998, so peers can safely reuse each other's resolutions. Structurally it's monotone submodular coverage (diminishing returns, not runaway). The one piece still unmeasured is at-scale reliability: a 100-peer, 30-day live pilot.
Bench: bench-compounding-real.mjs, bench-paraphrase-sigma.mjs
Can the system tell a reused answer is stale or off-topic?
The bi-encoder layer fails, but the final untried lever worked. MiniLM, bge, coverage, margin, entropy, and retrieval depth stay near chance (AUC 0.41–0.55). A cached ms-marco cross-encoder rescoring the live MiniLM candidates reaches AUC 0.884. That separates OOD relevance; it does not yet prove whole-answer replayability or detect on-distribution stale traces.
Bench: bench-energy-crossencoder.mjs · leave-one-out TPR 86.1%, FPR 9.1%
What benchmarks make these answers rigorous?
The harness is the suite. Retrieval quality on BEIR (SciFact / NFCorpus / FiQA), long-horizon memory on LongMemEval and LoCoMo, a templated poison eval (3 attacks × 3 rates × 3 treatments), a recall_any@k metric, and a vCache A/B for the savings question, plus the new ones built for this program: answer-quality-under-reuse, energy-coverage, energy-bge, and the OOD-signal sweep. Discipline throughout: matched comparisons, answers scored against gold qrels (never the gate), and every number labelled simulator vs measured. The open gap: no public benchmark scores cross-peer reuse quality + savings + trust jointly, which is the harness this program starts.
What architecture fits shared, provenance-tagged inference?
The longer-term frontier, framed not yet benched. The reranker is already one modern-Hopfield update step, which is softmax attention; personalized-PageRank, EigenTrust, and cache eviction all read as energy-descent fixed points. The hypothesis that ties it to RQ3: long-context models that ingest provenance-tagged retrieved context as first-class evidence should out-calibrate models that see anonymous tokens. The open theory problem is a distributed Hopfield capacity theorem: how much shared memory the network holds before retrieval degrades.
Open questions: where the research goes next
The current ledger separates valid results from required reruns. These are the next falsifiable steps, each a concrete experiment, not a wish.
Calibrate on fresh fixtures, reject tool output and query echoes, then test on-distribution stale or wrong traces rather than only OOD relevance.
The live web-fallback curve and false-admit tail the simulator can't produce. This is where the thesis stops being a thesis.
Rerun the poison matrix on actual SUPPORT/CONTRADICT verdicts, define NOINFO abstention, then add an adaptive adversary and true 7B consumer.
Rerun answer grading with correct labels and extend the 3-topic model-trace result to mixed verdicts, net tokens, latency, and provider billing.
No public benchmark measures all three at once. That gap is the harness this program starts.
How much shared memory the network holds before retrieval degrades, and whether long-context models that read provenance as first-class evidence calibrate better than those that see anonymous tokens.
The ledger changes when the evidence changes.
The cross-encoder rescued OOD relevance after every bi-encoder lever failed. The SciFact audit revoked the answer-safety and ~25× provenance headlines after finding invalid verdict labels. The 0.6 gate still sits above the measured different-need ceiling on the frozen fixture. Engineering a benchmark that can kill your own claim, then updating the record, is the whole posture.
Every figure here is a real bench run on real corpora or the live graph: wins, negatives, and revoked interpretations alike. BEIR qrels are treated as relevance labels, never claim verdicts. Methods, per-question notes, and raw captures live in /research.
Folklore · MIT-licensed · home · white paper · github