Skip to content

Quantitative evidence and constants (v1.2.0 development)

Current implementation and qualification are tracked in the v1.2.0 release record. The tables headed v1.1.0 below are historical measurements, not current release guarantees. The independent audit found that their recall corpus was insufficient and their 10k scan figure was an arithmetic extrapolation.

The frozen 27-row v1.2.0 corpus rerun records recall@1/@5=100% across repetitive, varied and multilingual synthetic cases at 100/1k/10k artifacts and three seeds. This establishes those identifier-retrieval cases, not general semantic/model effectiveness. See the exact standard corpus and multilingual corpus. Query timings in that run share the host with validation work and are not isolated SLO evidence.

Portable/native independent benchmark series and strict failures are retained in the release record. Hardware SLO failures remain failures. Current benchmark --json measures actual 1k and 10k scans independently and records build flags, CPU, corpus hash and gate outcomes.

This file is the single place where Infinity-Context (icx) states its quantitative claims. Other documents should link here instead of restating a figure. The dated historical values remain for comparison; current runtime contracts come from ADR 0006 and release evidence. Cross-document contradictions recorded by the audit (ICX-AUD-031) are listed with their canonical values in ref-docs/corrections-2026-11-18.md.

Every row is one of three kinds:

  • Benchmark: output of cargo run --release -p icx-cli -- benchmark for v1.1.0. The suite is synthetic and in-process; its figures are not end-to-end agent latencies.
  • Measured: a measurement taken during the v1.1.0 remediation on the development machine (Intel Core i9-10850K, AVX2, no AVX-512, Linux).
  • Constant: a value fixed in source code or asserted by a test. It changes only when the code changes.

1. Benchmark Suite (icx benchmark)

Release-build gates apply as written; CI multiplies the time ceilings by 3 and debug builds by 100 (crates/icx-cli/src/main.rs, scale).

Metric v1.1.0 value Fixture / how measured Release gate Source
HDC bind (^) + permute 46.82 ns / op 1,000 ops on 10,240-bit vectors at most 100 ns icx benchmark
HDC Hamming scan 64.20 µs / 1,000 nodes; extrapolated 642.05 µs /10k Historical 1k measurement multiplied by 10, not a measured 10k scan at most 150µs per 1k v1.1.0 icx benchmark
Fractal IFS suffix match 24.27 ns / pair shared_depth=6 at most 50 ns and depth at least 6 icx benchmark
Two-stage HDC query 49.00 µs synthetic index at most 200 µs icx benchmark
CSR Personalized PageRank 109.78 µs 201-node graph, 8 iterations at most 400 µs icx benchmark
AHTL inline ratio, repetitive log 214.0x (61,413 B to 287 B) in 44.43 µs repetitive 1,200-line test log ratio at least 10x, time at most 400 µs icx benchmark
AHTL inline ratio, varied log 159.8x (14,700 B to 92 B) varied log; informational, not gated none icx benchmark
FastCDC + BLAKE3 CAS ingest 86.72 µs 64,659 B, 8 chunks at most 250 µs icx benchmark
In-process seqlock publish 58.3 ns / slot 1,000 publishes, in-memory ring at most 5,000 ns icx benchmark
Synchronous hook round trip 244.05 µs one Claude Code PostToolUse on a fresh in-memory MMU (no persisted workspace) at most 1,000 µs icx benchmark
Byte-weighted prefix reuse, all turns 95.46% synthetic 50-turn transcript, seal turns included; a byte proxy, not a provider cache hit rate at least 90% (PREFIX_REUSE_FLOOR_PCT) icx benchmark
Byte-weighted prefix reuse, excluding seal turns 95.38% same transcript none icx benchmark
Passkey recovery in the 50-turn run recovered (18.13 ms for the run) planted passkey, recovered by deref and cold grep must be recovered icx benchmark

The AHTL ratio is the size of a skeleton plus a handle relative to the input. It is not a lossless encoding: the model sees the skeleton until it dereferences the handle, and the full text comes from deref.

2. Measured Outside the Benchmark

Metric v1.0.0 v1.1.0 How measured Source
icx hook, in-process, 338-artifact workspace 940 ms 14 ms real workspace, one hook invocation audit remediation measurement
icx hook, in-process, about 2,000 artifacts not measured 69–72 ms real workspace audit remediation measurement
icx hook via icx daemon, about 2,000 artifacts not available median 3.1 ms; p90 3.4 ms in a burst; about 19 ms when a request overlaps a flush resident daemon over its Unix socket audit remediation measurement
HDC unique-token recall, top-1 5.1% 99.0% 1,000 log documents audit remediation measurement
HDC unique-token recall, top-5 5.6% 99.7% 1,000 log documents audit remediation measurement
HDC query time on the recall corpus 24 µs 69 µs same corpus audit remediation measurement
File-backed bus, publish to visible not measured worst 38 µs over 200 events two handles on one file audit remediation measurement
FastCDC mean chunk size (8 KiB target) 13.2 KiB 10.6 KiB spread masks (Xia et al., ATC'16) audit remediation measurement

3. Constants and Test-Asserted Bounds

Handles, hashing and storage

Quantity Value Source
Handle widths 64-bit (#icx:<16-hex>) and 48-bit (#icx:<12-hex>) prefixes of the BLAKE3 digest icx deref help; crates/icx-core
50% collision point, 48-bit prefix (derived) about 1.98 × 10^7 artifacts (1.1774 · 2^24) birthday bound, N ≈ sqrt(2 ln 2 · 2^48)
50% collision point, 64-bit prefix (derived) about 5.06 × 10^9 artifacts (1.1774 · 2^32) birthday bound, N ≈ sqrt(2 ln 2 · 2^64)
Behaviour on a 64-bit handle collision IcxError::HashCollision; payload not virtualized, nothing overwritten crates/icx-core
Maximum ingest size 64 MiB (ICX_MAX_INGEST_BYTES) crates/icx-core/src/lib.rs
Maximum hook stdin 64 MiB + 64 KiB (MAX_STDIN_BYTES) crates/icx-cli/src/main.rs
Store layout one SQLite file icx.sqlite3 (WAL) per workspace crates/icx-cli
Workspace permissions directory 0700, files 0600; install key 0600, 32 random bytes crates/icx-cli
Default workspace id first 16 hex digits of blake3(canonical project root) crates/icx-cli

Codec (icx-codec)

Quantity Value Source
Maximum clause chain 64 ⇒ links (MAX_CLAUSE_CHAIN), parsed iteratively crates/icx-codec/src/ahtl.rs:90
Epoch header id hex u32, 4 to 8 hex digits (ep=000a is 10) crates/icx-codec/src/ahtl.rs
Braille-256 sketch sizes 8 bytes (Profile B) and 64 bytes (Profile A), one byte per glyph crates/icx-codec/src/braille.rs
Redactor tiers 4 (Tier 0 delimiter escaping, Tier 1 key blocks, Tier 2 prefixes and assignments, Tier 3 entropy) crates/icx-codec/src/redact.rs
Entropy gate Shannon character entropy above 4.35 bits/char crates/icx-codec/src/redact.rs:863
FLAG_IS_REDACTED bit 1 << 7 of ChannelBHologram byte B[7] crates/icx-codec/src/braille.rs

Hyperdimensional index and graph (icx-hdc, icx-store)

Quantity Value Source
Hypervector width 10,240 bits (HV_BITS), 160 u64 words crates/icx-hdc/src/hv.rs:8
Fractal coordinate in memory [f64; 16] crates/icx-hdc/src/fractal.rs
Recall gate test recall@5 at least 95% at n = 300 crates/icx-hdc/tests/recall_gate.rs
HDC snapshot format ICXHDC05 (binary, v1.2.2); 04 and 03 readable, postings rebuilt from CAS events by MMU crates/icx-hdc/src/snapshot.rs
PPR restart parameter 0.15 crates/icx-adapters/src/mmu.rs:848
PPR iterations, tail-slot recall 6 crates/icx-adapters/src/mmu.rs:848
PPR iterations, benchmark 8 crates/icx-cli/src/main.rs:1429

Epoch controller and proxy (icx-core, icx-proxy)

Quantity Value Source
EpochControllerConfig::default() (used by the MMU) 4,096 rollover tokens, 4 protected turns, 1,024-token tail slot crates/icx-core/src/epoch.rs:21-29
anthropic_default() preset 16,384 / 6 / 2,048 crates/icx-core/src/epoch.rs
openai_default() preset 8,192 / 6 / 2,048 crates/icx-core/src/epoch.rs
Anthropic breakpoints client breakpoints kept; icx adds at most 1, only while the total is below 4 crates/icx-proxy/src/wire.rs:109-148
Messages left unfolded at the tail last 2 crates/icx-proxy/src/wire.rs:99
Default proxy listen address 127.0.0.1:49380 icx --help
Maximum proxied request body 16 MiB crates/icx-proxy/src/wire.rs:1173

Bus (icx-bus)

Quantity Value Source
Ring slots 256 (ICX_RING_SLOTS) crates/icx-bus/src/ring.rs:15
Slot size 4,096 bytes crates/icx-bus/src/ring.rs:20
Inline AHTL capacity per slot 3,968 bytes (ICX_AHTL_INLINE_CAP) crates/icx-bus/src/ring.rs:16
Fabric header 64 bytes crates/icx-bus/src/ring.rs:19
Writer spin limit 4,096 iterations (MAX_WRITER_SPIN_ITERS) crates/icx-bus/src/ring.rs:24
File-backed transport pread/pwrite under a writer mutex and file lock; not memory-mapped, not lock-free, not zero-copy crates/icx-bus/src/ring.rs

Adapters, MCP and daemon (icx-adapters, icx-cli)

Quantity Value Source
Claude Code PreCompact limit (upstream) fewer than 10,000 characters Claude Code hooks documentation
icx PreCompact cap 9,000 characters (MAX_COMPACTION_INSTRUCTION_CHARS) crates/icx-adapters/src/claude.rs:199
Maximum fold depth 64 (MAX_FOLD_DEPTH) crates/icx-adapters/src/mcp.rs:17
Maximum MCP input line 4 MiB crates/icx-cli/src/main.rs:90
MCP tool schema size under 180 tokens by a characters/4 estimate test in crates/icx-adapters/src/lib.rs:2043
MCP protocol versions legacy initialize: 2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05; modern: 2026-07-28 crates/icx-adapters/src/mcp.rs
agy hook timeout default (upstream) 30 s Antigravity hooks documentation
Daemon acknowledgment Success only after SQLite WAL FULL commit under workspace lock; save errors propagate crates/icx-cli/src/daemon.rs; store.rs
Unix socket path limit 103 bytes (longer paths are rejected) crates/icx-cli/src/daemon.rs:108

Project and quality gates

Quantity Value Source
Workspace crates 8 Cargo.toml
Line coverage gate at least 95% .github/workflows/ci.yml:105
Fuzz targets 4 (ahtl_stream, hdc_snapshot, mcp_jsonrpc, redact) fuzz/fuzz_targets/
Roadmap sprints 15 to-dos/phase-*/sprint-*.md
arXiv papers indexed 22 ref-docs/papers/INDEX.md
arXiv PDFs redistributed 12 (11 CC BY 4.0, 1 CC BY-NC-SA 4.0) ref-docs/papers/

4. Figures That Are Not Claimed

The following appeared in earlier documents and are not claims of v1.1.0. Do not reintroduce them without a measurement:

  • any "cache hit rate" presented as a provider-measured figure (the 94.38% and 94%-99% figures of v1.0.0 came from an earlier simulation; the v1.1.0 metric is the byte proxy in section 1);
  • < 300 ns, < 500 ns, < 800 ns or < 1 µs file-backed or /dev/shm publish latencies;
  • AVX-512 / VPOPCNTDQ timings, which cannot have been measured on the AVX2-only benchmark host;
  • hook latencies of < 500 µs p99 or similar for a real workspace (see section 2);
  • compression ranges such as 12–72x stated without the input they were measured on;
  • figures attributed to research papers that the papers do not contain (see the corrections file).