README.md

metadata

license: mit
tags:
  - cortex
  - sft
  - synthetic
  - chat-format
pretty_name: CORTEX SFT 300
size_categories:
  - n<1K

cortex-sft-300

Phase 1 output of the Ornith-9B-CORTEX build (see docs/CUSTOM-MODEL-BUILD-PLAN.md in Johnny5b/portable-ai-runner).

304 validated chat-format examples ({"messages": [user, assistant]}), grown deterministically (seed 20260920) from the 14 seed examples in data/cortex_sft.jsonl of portable-ai-runner plus the repo's own source corpora: symbols/symbol_evolution_index.json, docs/CORTEX-SYSTEM-PROMPT.md, docs/EVO-SKILLS-BRIDGE.md, docs/CORTEX-LOOP-PHASE.md, docs/CORTEX-POINTER-ARCHITECTURE.md, docs/SYMBOL-TEMPLATE.md, docs/BENCHMARKS.md, docs/CORTEX-AWAKE-PHASE.md, docs/CORTEX-REPAIR-PHASE.md.

Validation (every example, data/validation_report.json)

• exactly one user + one assistant message
• assistant answer starts with the verbatim trace header [TRACE: §R:<target> | §T:<tool> | §S:<AWAKE|REPAIR|REST>]
• RST scaffold present (Claim: … Synthesis:)
• length bounds, no placeholders, sha256 dedupe (0 dropped)

Category mix

category	n
seed (original 14)	14
symbol_lookup	100
staged_refusal	36
doc_grounded (quotes the docs)	30
dual_verification (health/finance/nutrition/legal UNVERIFIED rule)	20
unknown_refusal	15
repair_workflow	12
module_mirror	12
lifecycle	10
trace_protocol	9
routing_benchmarks (measured numbers + job IDs)	9
registration	8
invariant_refusal	8
tiers_drive	8
invariants	7
excluded_repo	6

Honesty ledger

• Scripted generation, no LLM in the loop — templates + parsed registry/docs, so every symbol, commit hash, and tok/s figure is quoted from the source files, not invented.
• Benchmark numbers embedded in routing_benchmarks examples are the measured values from docs/BENCHMARKS.md with their HF job IDs.
• Additive: the 14 seed examples are included unmodified.