{"schema":"masidyn-primes-science/v1","title":"MasiDyn's Primes","subtitle":"Empirical coding-tool evaluations, measured model behavior laws, and falsification methodology — zero vendor marketing claims.","multiSkinContract":{"endpoint":"/api/masidyn/primes","surfaces":[{"host":"masidyn.dev","role":"Developer Lab, Toon Builder & SDK Reviews"},{"host":"masidyn.com","role":"Consumer Store & Base Bundle Portal (same prices)"},{"host":"masidyn.ai","role":"Mobile & Edge Companion Skin (same prices)"}]},"codingTools":[{"id":"cursor","name":"Cursor Harness","vendor":"Cursor","modelsTested":"Cursor Agent / Plan / Ask / Debug","evidenceTier":"SOURCE-BACKED","measuredDate":"2026-07-28","sourceArtifact":"agnostic/CURSOR_COMPARATIVE_EVALUATION.md","strengths":["Per-task cloud VM isolation makes the git diff an honest deliverable boundary.","Closed harness-enforced mode set (Agent / Plan / Ask / Debug) prevents ambiguous permission prompts.","Declared environment.json / Docker configs avoid split-brain read/write directory mismatches."],"measuredFailureModes":["No mechanical post-claim disk verification — self-reports are trusted once a diff exists.","No cross-provider refutation law (C0); single-family errors go unchecked.","Cloud agents load repo hooks rather than ~/.cursor/hooks.json; rented workers vanish when billing stops."],"bestTaskFit":"Isolated multi-file coding inside containerized tasks with external verification added on top."},{"id":"codex","name":"OpenAI Codex CLI & Desktop","vendor":"OpenAI","modelsTested":"GPT-5.6 Sol / Terra / Luna, GPT-5.5, GPT-5.4 Mini, gpt-5.3-codex-spark","evidenceTier":"INCIDENT-VERIFIED","measuredDate":"2026-07-28","sourceArtifact":"agnostic/matrices/OPENAI_GPT_FAMILY.md + agnostic/matrices/model_matrix.yaml","strengths":["GPT-5.6 Sol caught a subtle prompt-injection bug in Antigravity transcripts that a Claude seat missed (proven cross-family C0 refutation).","Clear capability routing split: Sol (Ultra helm), Terra (High builder), Luna (bulk structured transforms), 5.4 Mini (index/search)."],"measuredFailureModes":["codex exec via the bridge hung past --timeout as a spawned agentic subprocess, stalling sequential benchmark batteries.","Codex Desktop 26.721.41059 / CLI 0.146.0-alpha.3.1 failed chat creation when global [agents] existed in config.toml (AbsolutePathBuf deserialization error).","Pinning gpt-5.3-codex-spark (separate codex_bengalfox quota bucket) as default drains or breaks remote compaction."],"bestTaskFit":"Single-dispatch cross-family architecture audits and refutation passes; never unattended sequential timeout loops."},{"id":"antigravity","name":"Antigravity CLI (agy) & Gemini Tier","vendor":"Google","modelsTested":"gemini-4-argon-{high,medium,low}, gemini-3.8-flash-{high,medium,low}","evidenceTier":"RUNTIME-DEMONSTRATED","measuredDate":"2026-10-14","sourceArtifact":"agnostic/blueprints/gemini_4_argon_exploration_report.md","strengths":["gemini-4-argon-high completed direct stream-json compliance in 11,688.9 ms — 4.1x faster than gemini-3.8-flash-high (48,724.4 ms).","gemini-3.8-flash-low ran in 9,051.3 ms (flash_lite subagent tier ~3s turnaround) for high-volume bulk checks."],"measuredFailureModes":["Subagent probes proved 'pro' (3,901 out tokens) and 'flash' (1,158 out tokens) shared placeholder M322 rather than distinct model backends.","Earlier wrapper bug silently injected 'do not modify files' into write-enabled units until cross-family review caught it."],"bestTaskFit":"Argon High for architectural/coding execution; Flash High for long-deliberation R&D; Flash Low for bulk triage."},{"id":"claude-code","name":"Anthropic Claude Family","vendor":"Anthropic","modelsTested":"Claude Opus 4, Sonnet 4, Haiku 4 & Haiku 4.5","evidenceTier":"RUNTIME-DEMONSTRATED","measuredDate":"2026-06-29","sourceArtifact":"model-analysis-site/data/leaderboard.json + models.json","strengths":["flat-sonnet won quality outright in the price_refresh_health leaf tournament (43/50 total, 224 LOC, cost weight 3, value 14.33).","flat-haiku delivered highest value (39/50 total, 200 LOC, cost weight 1, value 39.0)."],"measuredFailureModes":["Multi-tier orchestration stacks (opus->sonnet->haiku, 41/50, cost_wt 19, value 2.16) added zero quality uplift at 4x-19x cost on deterministic leaf tasks.","Swarm sample sizes in models.json are small (n=2..36; e.g., claude-opus 0/4, claude-sonnet 2/6, claude-haiku 2/2) and must not be over-generalized."],"bestTaskFit":"Flat single-model runs for well-specified modules rather than multi-tier delegation chains on leaf scripts."},{"id":"grok","name":"xAI Grok Seat & Authority Harness","vendor":"xAI","modelsTested":"Grok-3 / Grok-3-Mini (workhub commission audit)","evidenceTier":"INCIDENT-VERIFIED","measuredDate":"2026-10-16","sourceArtifact":"agnostic/blueprints/rank_taxonomy_and_grok_5star_audit_20261016.md","strengths":["Effective peer verification and analytical critic role when isolated from authority-mutation writes."],"measuredFailureModes":["2026-10-16 workhub incident: promote_to_global_orchestrator only checked for a non-empty utterance string, letting an ambiguous question trigger self-promotion without an operator SHA-256 will anchor.","Proved that authority boundaries interpreted by models instead of enforced by harness cryptography fail under initiative."],"bestTaskFit":"Independent adversarial auditing behind harness-enforced read/write capabilities."},{"id":"masidyn-prime","name":"MasiDyn Prime Local Harness (:11435 / :11434)","vendor":"MasiDyn","modelsTested":"bit (0.6B), pico (0.5B), nibble (2B), byte (8-9B), word (14B), wordxl (24B MoE)","evidenceTier":"RUNTIME-DEMONSTRATED","measuredDate":"2026-11-02","sourceArtifact":"agnostic/matrices/model_matrix.yaml + SecondBrain Notes 2030/2035","strengths":["Mechanical proof-of-work contract (--selftest exits 0, --sabotage exits non-zero; caught 52 argparse fall-through bugs and 15 missing files).","100% operator-owned: workers, GGUF/I2_S weights, and 38-shard Spine RAG survive with $0.00 cloud spend.","masidyn-word (qwen2.5:14b) achieved 100% across 3 rounds on schema-drift when spine entered the request system message."],"measuredFailureModes":["No per-unit VM sandbox by default on local repo writes — units must be fenced by git worktrees or Baitball claims.","Stock Ollama failed (HTTP 500, exit 1) on BitNet-b1.58-2B-4T-gguf due to missing I2_S kernels — requires native MasiDyn ternary runtime."],"bestTaskFit":"Zero-spend local gating, deterministic verification, schema classification, and sovereign fallback."}],"cliLatencyBenchmarks":[{"model":"gemini-3.8-flash-low","wallLatencyMs":9051.3,"subagentPlaceholder":"M198 (flash_lite ~3s)","characteristic":"Fastest CLI completion; high-volume triage & checks"},{"model":"gemini-4-argon-low","wallLatencyMs":9149.6,"subagentPlaceholder":"Argon Low","characteristic":"Near-instant Argon compliance with low reasoning overhead"},{"model":"gemini-4-argon-high","wallLatencyMs":11688.9,"subagentPlaceholder":"M417 (inherit)","characteristic":"4.1x faster than Flash High with top architectural precision"},{"model":"gemini-4-argon-medium","wallLatencyMs":11776.7,"subagentPlaceholder":"Argon Medium","characteristic":"Virtually identical wall latency to Argon High"},{"model":"gemini-3.8-flash-medium","wallLatencyMs":22645.7,"subagentPlaceholder":"M322","characteristic":"Mid-tier Flash thinking budget before emission"},{"model":"gemini-3.8-flash-high","wallLatencyMs":48724.4,"subagentPlaceholder":"M322 (3,901 tok)","characteristic":"Deep internal deliberation budget before first streamed token"}],"leafTournament":{"taskClass":"build a contract-compliant utility script","taskInstance":"price_refresh_health","measuredDate":"2026-06-29","sourceArtifact":"model-analysis-site/data/leaderboard.json","sampleNote":"Single deterministic leaf benchmark (2026-06-29). Companion swarm pass-rates in models.json use small samples (n=2..36) and remain provisional until n >= 30.","finding":"Flat single-model runs beat orchestration-down hierarchies on both total score (flat-sonnet=43) and cost-efficiency (flat-haiku value=39.0), while stacked delegation cost 4x-19x more.","compositions":[{"label":"flat-sonnet (sonnet)","acc":9,"thor":9,"eff":8,"eleg":9,"util":8,"total":43,"loc":224,"costWt":3,"value":14.33,"champion":true},{"label":"normal (opus->sonnet->haiku)","acc":9,"thor":9,"eff":6,"eleg":8,"util":9,"total":41,"loc":332,"costWt":19,"value":2.16},{"label":"sonnet-led (sonnet->haiku)","acc":9,"thor":9,"eff":6,"eleg":9,"util":8,"total":41,"loc":232,"costWt":4,"value":10.25},{"label":"flat-haiku (haiku)","acc":7,"thor":8,"eff":9,"eleg":7,"util":8,"total":39,"loc":200,"costWt":1,"value":39,"bestValue":true},{"label":"opus-solo (opus)","acc":9,"thor":9,"eff":4,"eleg":9,"util":8,"total":39,"loc":291,"costWt":15,"value":2.6},{"label":"inverted (script->opus-execute)","acc":null,"thor":null,"eff":null,"eleg":null,"util":null,"total":null,"loc":432,"costWt":15,"value":null}]},"localFleetMatrix":[{"id":"gemma2:9b","mode":"non-thinking","latencyLabel":"12s P0 / 2s P1","leeway":"L1_narrow","calibrated80PctRung":"R2 (2 files)","verdict":"Top local workhorse; clean 3-bullet & classify adherence"},{"id":"gemma4-fast","mode":"evolved-fast","latencyLabel":"3s P1 (~10x faster)","leeway":"L1_narrow","calibrated80PctRung":"R1 (evolved)","verdict":"Ollama Modelfile temp=0 + terse SYSTEM suppresses CoT spill"},{"id":"nemotron-mini","mode":"non-thinking","latencyLabel":"11s P1 (2.7 GB)","leeway":"L1_narrow","calibrated80PctRung":"R1 (1 file)","verdict":"Fastest concise local classify reflex"},{"id":"phi4","mode":"non-thinking","latencyLabel":"27s P0","leeway":"L1_narrow","calibrated80PctRung":"R1 (1 file)","verdict":"Solid reasoning-lite backup; ~2x slower than gemma2:9b"},{"id":"gemma4","mode":"thinking","latencyLabel":"30s P1 (no clean ans)","leeway":"L0_reflex","calibrated80PctRung":"R1 (1 file)","verdict":"Strong base, but un-evolved thinking leaks before final tag"},{"id":"qwen3.6","mode":"thinking","latencyLabel":"66s P1 (23 GB)","leeway":"L0_reflex","calibrated80PctRung":"R1 (1 file)","verdict":"Deep heavy synthesis only; never route tight latency loops"},{"id":"qwen3.5:9b","mode":"thinking","latencyLabel":"151s P0 timeout","leeway":"L0_reflex","calibrated80PctRung":"R1 (1 file)","verdict":"Over-deliberates past timeout despite 'no preamble' prompt"},{"id":"qwen2.5:14b","mode":"bridge-quarantined","latencyLabel":"60s bridge stall","leeway":"L0_reflex","calibrated80PctRung":"R1 (1 file)","verdict":"100% decision-perfect on direct system-spine; spinner quirk via bridge"},{"id":"hermes3 / llama3.1","mode":"stdin-unproven","latencyLabel":"23s / 20s empty","leeway":"L0_reflex","calibrated80PctRung":"R1 (1 file)","verdict":"Empty output on stdin pipe; unproven until arg-form re-probe"}],"cursorScorecard":[{"dimension":"Task isolation / sandboxing","cursor":"Per-task VM","masidyn":"Repo root unless worktree-fenced","winner":"Cursor"},{"dimension":"Write-authority model","cursor":"Closed modes, harness-enforced","masidyn":"Prompt-stated string (upgrading to modes)","winner":"Cursor"},{"dimension":"Environment declaration","cursor":"environment.json / Docker","masidyn":"Implicit env (caused spine split-brain)","winner":"Cursor"},{"dimension":"Live observability","cursor":"IDE / web / Slack / iOS surfaces","masidyn":"Local SQLite board & BEAM telemetry","winner":"Cursor"},{"dimension":"Local <-> cloud handoff","cursor":"First-class","masidyn":"Gateway adapter / tunnel bridge","winner":"Cursor"},{"dimension":"Proof of work delivered","cursor":"Diff + human inspection","masidyn":"Mechanical --selftest 0 / --sabotage != 0","winner":"MasiDyn"},{"dimension":"Cross-provider verification","cursor":"Model-agnostic, rank-free","masidyn":"Law C0 (producer != verifier family)","winner":"MasiDyn"},{"dimension":"Durable law portability","cursor":".cursor/ vendor-scoped","masidyn":"Provider-neutral hash-bound canon","winner":"MasiDyn"},{"dimension":"Cross-session federated memory","cursor":"Not automatic across tools","masidyn":"38-shard Spine RAG + SecondBrain","winner":"MasiDyn"},{"dimension":"Swarm-scale unit economics","cursor":"Per-session optimization","masidyn":"Wave amortization & 80% R1..R5 envelope","winner":"MasiDyn"},{"dimension":"Explicit gap honesty","cursor":"Explicit non-continuity disclosure","masidyn":"Admitting gaps via NOT_DONE receipts","winner":"Cursor"}],"behavioralLaws":[{"id":"spine-delivery-position","title":"Spine Delivery Position Law","formulaOrThreshold":"Request System Message (100%) vs. Modelfile Baked SYSTEM (55%-85%)","measuredResult":"masidyn-word (qwen2.5:14b) scored 100% across 3 rounds with zero under-alarms when rules entered the request system message, but dropped to 85% with under-alarms when the identical text was baked into Modelfile SYSTEM (nibble dropped 90% -> 55%/70%).","operationalRule":"Keep model weights & sampling params fixed, and inject spine rules dynamically at the request system boundary.","sourceArtifact":"SecondBrain/notes/2030_masidyn_word_birth_and_spine_delivery_position_law.md"},{"id":"frontier-small-model-misprediction","title":"Frontier Misprediction of Small Models","formulaOrThreshold":"gemma2:9b (100%) | gemma2:2b (50%-100%) | llama3.2:3b (0%) | 1B (below JSON floor)","measuredResult":"Frontier models assume emergent reasoning absent below ~4B params, causing single-character prompt flips ('butterfly effect') and 20% drift collapses on vision-tuned small models (qwen3-vl:4b).","operationalRule":"Let frontier models reason once at design time into a deterministic choice matrix; let small models run grammar-constrained lookups with zero CoT.","sourceArtifact":"SecondBrain/notes/1543_frontier_mind_mispredicts_small_models.md"},{"id":"vram-offload-ratio","title":"VRAM Offload Ratio Law","formulaOrThreshold":"R_offload = (V_budget - V_KV_cache) / V_weights >= 1.0","measuredResult":"On a 16 GB GPU (~14,800 MB usable budget), R_offload >= 1.0 yields 85%-100% GPU utilization (qwen2.5-coder:14b at 65-85+ tok/s). Below 1.0, PCIe transfers stall GPU compute ~80% of the time (~20% utilization).","operationalRule":"Cap context (-c) or swap down to Q4_K_M / 14B rather than letting 1-2 layers spill across PCIe.","sourceArtifact":"agnostic/blueprints/vram_ratio_and_model_tradeoffs.md"},{"id":"bitnet-ternary-kernel-gap","title":"BitNet Ternary Kernel Gap","formulaOrThreshold":"Stock llama.cpp / Ollama -> HTTP 500 (exit 1) on BitNet-b1.58-2B-4T-gguf","measuredResult":"Standard Ollama bundles lack ternary I2_S / TL1 / TL2 kernels and crash during GGUF tensor load; native ternary C/Python kernels are mandatory for sub-2-bit weights.","operationalRule":"Route BitNet / Hyperion-I2_S 0x55 weights exclusively through the MasiDyn native stack, never stock Ollama.","sourceArtifact":"SecondBrain/notes/2035_ollama_cannot_load_bitnet_ternary_gguf.md"}],"transportProbeChecks":[{"id":"direct-vs-pipeline","name":"1. Direct vs. BEAM Pipeline","measure":"Compare SHA-256 content hashes, finish_reason, and latency between :11521 direct and :11435 BEAM gateway."},{"id":"hop-sse-stitching","name":"2. Hop Integrity & SSE Stitching","measure":"Track exact bytes per hop and diff stitched SSE completion chunks against non-streamed responses."},{"id":"synthetic-delay","name":"3. Synthetic Relay Delay","measure":"Inject controlled HTTP relay delay to separate network transport lag from model inference degradation."},{"id":"shuffled-constants","name":"4. Shuffled Constants Check","measure":"Grade arithmetic word problems under swapped constants (e.g. 17 * 23) to expose memorized benchmark answers."},{"id":"multistep-compounding","name":"5. Multi-Step Compounding (p^k)","measure":"Compute true multi-step reliability p^k vs. oracle-gated retry bounds rather than extrapolating 1-turn pass@1."}],"testingLadder":{"principle":"Preconditions are always on; Interventions sort by reversibility; Diagnostics enter only on failure sorted by cost.","sourceArtifact":"agnostic/gold_standards/pico_pack_staging/MODEL_TESTING_LADDER.md","parts":[{"name":"Preconditions (Continuous)","detail":"P1 baseline on real operator corpora (never synthetic/LLM-invented tests), negative cases included, zero production mocks."},{"name":"Interventions (By Reversibility)","detail":"1. Request system-message spine swap -> 2. Sampling/Modelfile params -> 3. Adapter/LoRA swap -> 4. Full fine-tune."},{"name":"Diagnostics (On Failure Only)","detail":"Entered only after cheap reversible fixes fail; produces inspectable receipts, never self-certified proof."}],"laws":["Law C0 (Producer != Verifier): Milestone sign-off requires an independent model family.","Zero Synthetic Eval Ban: Model-generated test corpora are rejected from P1 certification.","Small-Sample Honesty: Provisional n < 30 swarm pass rates are surfaced verbatim instead of extrapolated."]}}