EviCortex

Prototype A benchmark result

Status: last verified local result, 2026-07-21. This is an engineering checkpoint, not a publication-quality comparison or a general claim that EviCortex is better than RAG.

Configuration

$env:PYTHONPATH = "src"
python -m benchmarks.run --events 1000 --seed 7 --budget-tokens 512 --json
Setting Value
Events 1,000
Queries 12: 9 answerable and 3 retrieval-abstention probes
Source bytes 221,586
Suite SHA-256 109c196d585532bf6ee7f9f1f96d8b33a1d8cd57f0f1c17ec042b37fc6b59f7a
Runtime-source SHA-256 8f5a4cc0f922eedba3a6a7d1f5be757925684fb085ea7e46fbc9a64b7c0ea4f7
Candidate limit 40
Baseline chunks 320 characters, 48-character overlap
Token estimate UTF-8 bytes / 3
Reader model None; this measures evidence support, not generated-answer accuracy
Runtime Python 3.12.0, SQLite 3.42.0, local Windows machine

The worktree was dirty, so the report records both Git HEAD and a deterministic path+content digest of all runtime source. Both databases are cleanly closed, symmetrically VACUUM- compacted, closed again, and then measured; compaction time is excluded from ingestion.

Both systems received the same canonical events, timestamps, namespace restrictions, query stream, and serialized-context budget. Times are single-machine measurements with only 12 query probes and should be treated as directional.

Systems compared

The baseline is fixed-size chunks ranked by SQLite FTS5 and packed to the same budget. It is a deliberately mechanical lexical baseline—not dense retrieval, hybrid RAG, GraphRAG, or a tuned production stack.

The candidate is the deterministic EviCortex kernel: immutable evidence, explicit temporal claims, lexical/entity/graph/time retrieval, stale-claim filtering, and bounded capsule compilation. Neither system used an LLM. The generator supplies EviCortex with oracle structured claims while placing the same canonical facts verbatim in baseline text. This isolates representation/retrieval and excludes real extraction errors and cost.

Single-seed result

Metric Fixed-chunk FTS5 EviCortex Interpretation
Evidence Recall@Budget 88.89% 100% Candidate found every required source ID
Supported-answer rate 88.89% 100% +11.11 percentage points
Current-fact accuracy 50% 100% Update handling improved
Historical as-of accuracy 100% 100% Equal
Stale-fact leakage 100% 0% Superseded evidence excluded from current recall
Retrieval-level no-evidence abstention 33.33% 33.33% Both fail the two partial-OOD probes
Namespace isolation / budget compliance 100% / 100% 100% / 100% Equal
Estimated context tokens, total 3,701 2,216 Candidate used 40.1% fewer
Retrieval latency p50 0.532 ms 3.725 ms 7.01x; only 12 probes
Retrieval latency p95 2.865 ms 4.614 ms 1.61x; only 12 probes
Retrieval latency p99 4.656 ms 4.834 ms 1.04x; descriptive near-maximum only
Ingestion time 9,617.0 ms 11,643.5 ms 1.21x baseline
Compacted database size 1,265,664 B 1,900,544 B 1.502x baseline
Storage/source-byte ratio 5.712x 8.577x Candidate stores more structure
Query errors 0 0 Equal

The candidate met the quality/context and temporal gates. It did not establish robust abstention, and it narrowly missed the optional 1.5x compacted-storage target by 0.16%.

Three-seed scale check

$env:PYTHONPATH = "src"
python -m benchmarks.sweep --seeds 7,11,19 --events 1000 --budget-tokens 512 `
  --bootstrap-samples 2000 --bootstrap-seed 20260721
Paired aggregate Mean Descriptive 95% bootstrap interval
Supported-answer delta +11.11 percentage points +11.11 to +11.11 pp
Evidence-recall delta +11.11 percentage points +11.11 to +11.11 pp
Retrieval-abstention delta 0 0 to 0
Context-token ratio 0.594x 0.586x to 0.599x
Retrieval p50 ratio 6.99x 6.89x to 7.06x
Retrieval p95 ratio 1.538x 1.458x to 1.610x
Retrieval p99 ratio 0.960x 0.880x to 1.038x
Ingestion-time ratio 1.208x 1.199x to 1.213x
Compacted database-size ratio 1.502x 1.502x to 1.502x

All three streams share the same task templates and vary only generated values and distractors. The intervals are unstable small-N descriptions, not evidence of statistical significance. They do show the cost shape: the candidate uses about 40.6% fewer context tokens and fixes the baseline’s update failure, while paying materially higher median lookup, ingestion, and storage costs.

Gate status

Gate Status Evidence or limitation
Namespace filtering, digest verification, budget compliance, deterministic ordering, preserved history, and per-item traces Pass Last verified suite: 144 tests passed
≥10% relative support improvement or ≥20% fewer tokens at equal support Pass Support rose to 100%; context fell 40.1%
No more than two-point direct-recall regression Pass in this suite Direct evidence remained supported
Lower stale leakage; no worse historical accuracy Pass 100%→0% stale leakage; historical stayed 100%
Candidate p95 no greater than 2x baseline Pass, smoke-level only 1.61x in the retained run
Additional database target no greater than 1.5x baseline Narrow miss 1.5016x after symmetric compaction
Multi-seed paired bootstrap report Run, small-N only Three template-sharing streams
Robust unsupported-query abstention Fail Both systems abstain only on the double-OOD case: 1/3
Tuned dense-hybrid comparison Not met Phase B currently has only a hashing diagnostic

Interpretation and next evidence

Prototype A supports one narrow conclusion: on this deterministic workload, explicit temporal structure and budget-aware compilation fixed a mechanical chunk baseline’s update failure and reduced prompt material. It does not show that the approach stores more information than RAG, scales better, or beats a modern retriever.

The next decisive comparison needs a pinned dense encoder, equivalent metadata/version history, a shared reranker and reader, exact tokenizer accounting, partial-OOD support verification, and official external workloads. See the Phase B diagnostic result for the implemented intermediate step.

Biology and naming caveats

“Genome,” “Engram,” and “Transcript” are software metaphors. Human autobiographical memories are not stored as ordinary records in DNA or RNA. EviCortex is the selected project identity, but it has not undergone formal trademark clearance.