Skip to content

Latest commit

 

History

History
194 lines (157 loc) · 11.3 KB

File metadata and controls

194 lines (157 loc) · 11.3 KB

B.I.O.M.A. — Empirical Findings (ground-truth evaluation)

Honest, reproducible evaluation of every B.I.O.M.A. mechanism against real data. Written to survive technical due diligence: it states what is proven, what is refuted, and what remains a design goal. Numbers below are measured, not asserted. Where a claim did not hold up, it is marked as such.

TL;DR

  • 🟢 Context apoptosis (−78–80% input tokens) is the real, universal, sellable win. Every model, every task.
  • 🟢 The Rust kernel is genuinely fast and resilient: ~2M signals/s at ~5μs mean latency, latency stays bounded under 10× load.
  • 🔴 Multi-LLM mitosis+synthesis does NOT improve answer quality or security remediation. On ground-truth executed tests it was neutral on frontier models (ceiling) and harmful on weaker ones (synthesis corrupts correct answers). Even the corrected best-case design (verified cross-model selection, baseline-in-pool) delivered +0 on hard security tasks. This thesis is refuted by our own data across three independent experiments and is not part of the honest pitch.
  • Positioning: B.I.O.M.A. makes AI processing viable, sustainable, and resilient — an efficiency/infrastructure layer, not a system that makes AI "smarter."

The evidence ledger

Claim Status Basis
Context apoptosis cuts ~80% of input tokens 🟢 Proven Measured in every run (ContextPruner, kernel-backed)
Kernel sustains 10k concurrent agents at μs latency 🟢 Proven bioma_kernel_loadtest.py
Kernel latency stays bounded under 10× load 🟢 Proven avg 4.5μs→5.0μs (1.1×), p99 21μs→15μs
Mitosis+synthesis improves quality 🔴 Refuted Objective code eval: 95%→83% (−17 tests)
Mitosis improves security remediation 🔴 Refuted Defense eval: baseline 98% ≥ synthesis 90% = selection 90%
Verified cross-model selection improves security (corrected best-case design) 🔴 Refuted +0 vs best single-shot; all models 100% on hard tasks (no headroom)
"Selection-by-execution is provably ≥ baseline" 🟢 Corrected & validated With the baseline in the pool it held exactly (30/30 = 30/30, monotonic) — but delivered +0
Resilient "at industrial level" (end-to-end system) 🟡 Goal Kernel primitive proven; full system not load-tested

Experiments & results

All LLM calls go through OpenRouter (bioma_orchestrator/openrouter_async.py) with exponential backoff. "Real" = a valid sk-or key; otherwise a clearly-labelled deterministic mock. Quality was measured two ways: a noisy LLM judge (early runs) and — decisively — objective execution against test suites (no judge).

1. Context apoptosis — 🟢 the universal win

ContextPruner (Rust StateContext oxygen/decay, Python fallback) prunes a bloated agent context before the model call. Measured reduction: −78% to −80% of input tokens, consistent across GPT-4o, Fable-5, Opus 4.8, Grok-4.3, Llama-3.3-70B and on every task. This is the one mechanism that helps unconditionally (cost + focus), including on frontier models. Repro: any of the sim scripts print apoptose −80%.

2. Kernel resilience — 🟢 proven (bioma_kernel_loadtest.py)

1k → 10k concurrent Tokio micro-agents flooding the lock-free hormonal bus, with live apoptosis on a CPU-pinned telemetry core (12-core host):

Concurrent agents Throughput Mean latency p99 latency
1,000 2.31 M sig/s 4.54 μs 21 μs
2,500 1.75 M sig/s 6.04 μs 20 μs
5,000 1.84 M sig/s 5.76 μs 27 μs
10,000 2.12 M sig/s 4.99 μs 15 μs

Under 10× load, mean latency moved only 4.5→5.0μs (1.1×) and p99 improved. Defensible claim: "the sovereign kernel sustains 10,000 concurrent agents at ~2M signals/s with bounded ~5μs latency." Caveat: max latency shows 10–37ms tail spikes = OS scheduling on a non-realtime OS (userspace); mean/p99 are the meaningful bounded metrics. Scope = the kernel primitive, not an end-to-end LLM system.

3. Objective code correctness — 🔴 mitosis refuted (bioma_objective_eval.py)

5 edge-case-heavy algorithmic tasks, 144 hidden tests, code executed in an isolated subprocess (no judge). Baseline (1 call) vs B.I.O.M.A. forced mitosis:

Model Baseline B.I.O.M.A. mitosis Δ
GPT-4o 48/48 (100%) 48/48 (100%) 0 (ceiling)
Grok-4.3 48/48 (100%) 48/48 (100%) 0 (ceiling)
Llama-3.3-70B 41/48 (85%) 24/48 (50%) −17
Aggregate 137/144 (95%) 120/144 (83%) −17

Smoking gun: Llama is_valid_number went 17/17 → 0/17 — the model was correct single-shot, and the LLM synthesis of 3 hypotheses corrupted the correct answer. This is the mechanism of harm: blind synthesis has no ground truth and can merge a flawed hypothesis over a correct one.

4. Multi-model efficiency (LLM judge) — mitosis neutral-to-negative (bioma_efficiency_simulation.py)

Neutral algorithmic task, 5 models, judge = gpt-4o-mini (noisy ±5):

Model Quality Base→BIOMA Cost Latency
GPT-4o 90→85 $0.006→$0.024 7→10s
Fable-5 95→100 $0.045→$0.302 12→45s
Opus 4.8 95→95 $0.031→$0.152 19→44s
Grok-4.3 95→95 $0.003→$0.010 9→17s
Llama-3.3-70B 95→90 $0.0003→$0.001 15→114s

Aggregate: −1.0 pt quality, 5.7× cost. Most deltas are within judge noise. On tasks frontier models already solve there is no headroom for mitosis to help. Engineering note: an earlier run mis-attributed a −25 Opus "regression" and a Fable-5 failure to B.I.O.M.A. — both were bugs: a low max_tokens truncated verbose models and starved reasoning models; Fable-5 also tripped its content filter on a security-flavoured task. Fixed (2048/2560 token budgets, staggered fan-out, neutral task). After the fix the results are the honest ones above.

5. Security remediation — 🔴 mitosis no gain (bioma_defense_eval.py)

3 vulnerabilities (unsafe eval, SQL injection, command injection), 17 executed security/functionality checks, benign-by-construction (no shell exec, no I/O). Baseline vs synthesis vs selection-by-execution:

Model Baseline Synthesis Selection
GPT-4o 17/17 17/17 17/17
Grok-4.3 17/17 17/17 17/17
Llama-3.3-70B 16/17 (94%) 12/17 12/17
Aggregate 98% 90% 90%

Frontier models: ceiling (all tie). Weak model: both mitosis arms fell below baseline. Llama build_query candidates scored [0,0,0] — the role-specialised prompts were worse than the plain baseline prompt for a weak model on a structural task, and the baseline was not in the selection pool.

6. Adaptive mitosis — cost-saving works, confidence signal is weak (live_pipeline.py)

A scout cell probes difficulty; only low self-confidence escalates to full mitosis.

  • ✅ Easy task → 1 call instead of 4 (real cost cut).
  • ⚠️ Hard task → GPT-4o self-rated 95 on a lock-free concurrency proof → did not escalate. LLM self-confidence is overconfident, so it is a weak escalation trigger. A robust gate needs a consistency/critic signal, not self-report.

7. Corrected mitosis: verified cross-model selection — 🔴 still +0 (bioma_verified_selection_eval.py)

The best-case design the analysis pointed to: baseline-in-pool (each model's plain answer is a candidate → selection can't drop below baseline) + cross-model diversity (GPT-4o + Grok-4.3 + Llama-3.3-70B) + objective verification (run the checks) as the selector — never LLM synthesis. Tested on 3 harder security tasks (30 checks: Windows reserved-name filename sanitizer, open-redirect validator with userinfo/backslash bypasses, PII masker with separated card numbers):

Reference Score
GPT-4o single-shot 30/30 (100%)
Grok-4.3 single-shot 30/30 (100%)
Llama-3.3-70B single-shot 30/30 (100%)
Best single model 30/30 (100%)
Verified cross-model selection 30/30 (100%)

Δ = +0. Two things are true and both matter: (1) the corrected design worked as designed — selection held exactly at baseline (monotonic, never worse); (2) it delivered no gain, because all models hit the ceiling even on the harder tasks. Three independent experiments (§3, §5, §7) now converge on the same verdict: on well-defined, objectively-verifiable tasks, modern single-shot is already at ceiling, so mitosis — even in its best-case corrected form — has no headroom to add value, at 6× cost.


Honest corrections we made to our own claims

  1. "Mitosis/orchestration improves answer quality." Refuted by ground truth.
  2. "Selection-by-execution is provably ≥ baseline." Overstated. It is ≥ the best candidate; to be ≥ baseline the baseline generation must be included in the candidate pool. Even then, on our tasks it only ties (no headroom).
  3. "Resilient at industrial level / ready for the planet's hardest problems." The kernel primitive is proven resilient; the end-to-end system is not yet load-tested, and the security simulation is inert. State it as a design goal, not a result.

Product recommendation

  1. Ship: context apoptosis (always on) + the μs Rust kernel. Proven, auditable.
  2. Mitosis: OFF by default. If offered, only as optional escalation with the baseline-in-pool selection design (guarantees "never worse") — and without any quality claim.
  3. Position on efficiency/infrastructure and resilience, not intelligence — the only framing the evidence supports.

Where the mitosis story could still earn value (unproven, needs testing)

  • Tasks genuinely beyond a single model's frontier (our tasks were at ceiling).
  • Objective verification as the selector (execution/property tests), never blind synthesis, with the baseline in the pool.
  • Cross-model diversity (different providers) instead of same-model temperature — same model shares its blind spots.
  • Reframe the target from average quality to tail-failure reduction (95%→99%), which only matters with objective verification.

Reproducibility

The harnesses below are archived in a private research repo — they depend on the legacy multi-LLM orchestrator (bioma_orchestrator), removed from the lean product once mitosis was refuted. This document preserves the full method and results; the product's own claims (apoptosis + firewall) are reproducible from tests/.

Script What it measures Cost
bioma_kernel_loadtest.py kernel throughput/latency under 1k–10k agents free
bioma_objective_eval.py code correctness, executed tests ~$0.4
bioma_defense_eval.py security remediation, executed checks ~$0.3
bioma_efficiency_simulation.py per-model efficiency (judge) ~$0.7
bioma_sakana_console_test.py demo console (μs lab notebook) ~$0.15
bioma_sovereign_defense_simulation.py defensive immune-system demo (inert) ~$0.01

All scripts run offline in a labelled mock mode without a key. Real mode needs a valid OPENROUTER_API_KEY (in .env, git-ignored — never commit a key).