TGMS v0.1.0 — first public release (Phase 1–2 research prototype)

A bi-temporal property graph store whose query surface is a small algebra of
verified temporal operators exposed as tools to LLM agents, plus a
Planner–Executor–Verifier layer that machine-checks every claim in an answer
against the execution trace that produced it.

Highlights

• Bi-temporal substrate (valid × transaction time): evolution vs correction,
  "as of" queries, append-only event log with replay-identical digests
  across Kùzu and DuckDB backends
• 13 verified operators: typed · deterministic · bounded · cost-guarded ·
  output-contracted; 500 randomized oracle cases per operator; bi-temporal
  immutability metamorphic tests; 96% line coverage in tgms/temporal
• PEV agent layer: constrained plan IR with a grounding rule (fabricated
  identifiers impossible by construction), deterministic executor with
  content-addressed traces, trace-grounded claim verifier with truncation
  taint — fault-injection acceptance 500/500 detected, 0 false positives
• Eval stack: task suites with program-computed gold (incl. correction
  probes no baseline can answer), vector-RAG / static-graph RAG /
  text-to-Cypher baselines, matrix harness with paired-bootstrap CIs and
  determinism receipts
• First results (CollegeMsg dev, local Qwen2.5-7B/14B): pooled EM 0.41 vs
  0.05–0.18 for baselines; correction probes 0.67 vs 0.00 — details in
  docs/TECHNICAL_REPORT.md
• Interfaces: Python library · MCP server · CLI · static trace viewer ·
  interactive guided demo GUI (scripts/run_webapp.sh)

See CHANGELOG.md and docs/TECHNICAL_REPORT.md for the full picture.