Skip to content

OVERVIEW.md

Latest commit

 

History

History
90 lines (68 loc) · 5.33 KB

OVERVIEW.md

File metadata and controls

90 lines (68 loc) · 5.33 KB
Raw

B.I.O.M.A. — What it is, what it solves, and the proof

What B.I.O.M.A. is (now)

B.I.O.M.A. is a local, provider-agnostic efficiency & security micro-kernel for LLM applications — a lock-free Rust core (bioma_micro) plus a thin Python layer (bioma) that you embed in front of any LLM call, in-process, before your prompt ever leaves the machine. It works with Anthropic, Google, OpenAI — or any provider — because it hardens the payload and hands it back to your SDK.

It is not a model and it does not try to make a model "smarter." We tested that thesis (multi-LLM orchestration / "mitosis") and refuted it with ground truth — see below. B.I.O.M.A. makes the processing cheaper, faster and safer.

Main objective

Make LLM processing viable, sustainable, and safe at scale — cut the token cost, resist floods, protect secrets, and bound latency — as a drop-in local artifact, with no vendor lock-in.

The pain it attacks

  1. Token/cost bleed. Every call re-sends bloated context (verbose tool logs, stale turns). On long sessions the input grows unbounded and the bill explodes.
  2. Cognitive DDoS / context-window exhaustion. Repetitive floods and forged-log spam blow the context window → denial of service on the reasoning pipeline.
  3. Secret leakage via prompt injection. Apps leave secrets in the working context; an injection asks the model to print them.
  4. Latency & hangs. Unbounded calls (or loop-injection) stall the pipeline.
  5. Vendor lock-in. Coupling the whole stack to one provider's API.

How it works — four real mechanisms

Mechanism What it does
Context apoptosis Assigns each context block a metabolic weight, applies aggressive half-life decay, and purges low-value blocks (old logs, resolved chatter) before dispatch — dehydrating the input. Cache-aware since kernel 1.1.0: a stable_prefix zone kept byte-identical for provider prompt caching, plus consolidation_gain() for the rewrite-a-cached-prefix economics.
Effort gauge (1.1.0) An O(n) lexical task-complexity score (calibrated on 1,223 real agent prompts) that sets a per-request thinking budget — trivial turns stop paying for reasoning. Exposed as effort_gauge() and wired into the gateway as opt-in BIOMA_AUTO_EFFORT; never raises effort beyond the client's ask; every decision logged to the audit.
Cognitive firewall Secret redaction (vault values scrubbed from the outbound payload AND the response), saturation detection (saturation_scan flags repetitive floods → 0x0F red alert → apoptosis), and a timeout guard bounding every dispatch.
Hormonal bus A lock-free, atomic in-memory signalling substrate (μs), used for the alert state.

Proof — real, reproducible benchmarks (ground truth)

Every number below was measured this project, not asserted. Scripts are in the repo.

Efficiency — context apoptosis

  • Up to −84.7% input tokens on stale-heavy agent sessions (measured, quality-neutral); savings are workload-dependent — with the Context Guard (v1.5) verified compressions median −43% and fall back to the intact context when the query needs it (never worse than baseline; see benchmarks/ab-context-scale/PROVA.md); up to −97% on long, noisy sessions.
  • 16-round real session (live OpenRouter, tests/test_enxuto_efficiency.py): input 47,890 → 2,022 tokens (93.8% avg/round, 97.5% by round 16), apoptosis latency ~1.6μs avg, 0/16 dispatch errors, whole session cost $0.0191.

Resilience — the Rust kernel (archived research bench)

  • ~2M signals/s at ~5μs mean latency.
  • 1k → 10k concurrent agents (10× load): mean latency 4.5μs → 5.0μs (1.1×), p99 21μs → 15μs — bounded, sub-linear under load.

Security — cognitive firewall APT war-game (real, reports/BIOMA_IMMUNITY_VERDICT.md)

  • Prompt-injection secret exfiltration: 2 secrets redacted, 0 leaked → CONTAINED.
  • Cognitive DDoS: a 32,317-token flood detected (saturation 0.999 → 0x0F) and dehydrated to 13 tokens in 0.6μs → MITIGATED.
  • Code-injection loop: contained by the timeout guard → CONTAINED.

Engineering

  • 20 unit tests (kernel, firewall, server), offline/deterministic, green in CI.

Proof of truth: what we refuted (and kept out of the pitch)

Rigor cuts both ways. We tested the "orchestrate many LLMs → better answers" thesis and it failed on ground truth (FINDINGS.md):

  • Objective executed-test correctness: baseline 95% → mitosis 83% (−17 tests; synthesis corrupted correct answers). Neutral on frontier models (ceiling), harmful on weaker ones.
  • Confirmed across three independent experiments (code eval, security remediation, verified cross-model selection) — always ≤ 0 gain, at 4–6× the cost.

So we removed it from the product. The value is the apoptosis + kernel + firewall — measured — not a quality claim we can't defend.

Positioning in one line

B.I.O.M.A. turns an expensive, fragile LLM call into a cheap, flood-resistant, secret-safe, bounded-latency one — locally, with any provider. Proven, honest, auditable.

See README.md for usage and FINDINGS.md for the full evaluation (proven and refuted).