Ten questions a buyer actually asks.
Ten plain answers.

They tell you what your agent did. We decide whether it should have. Different insertion point: observability sits after the run, fewwords sits in the dispatcher. A post-hoc dashboard can’t stop a DROP DATABASE that already executed. Run both — the grid above shows how.

A judge is another model grading this model. Same input, different day, different verdict. Our contracts are YAML, compiled to deterministic checks: a Python expression, a JSON Schema, or an LTL formula compiled to a Büchi automaton. Same trace, same verdict, forever. Nothing for a compliance officer to audit except the config file. LLM-as-judges measured on our tau-bench comparison hit 54–62% agreement with a strong auditor; fewwords hits 80–82%.

Those are per-turn input/output validators. They tell you “this string is safe.” We tell you “this tool call is allowed right now, given the sequence that came before.” A lazy agent skipping validate_payment produces a perfectly safe-looking string; a trajectory contract catches the missing step. Use both at different layers.

If that were true, every observability vendor would have shipped it in a weekend. They haven’t. The hard parts: normalising six trace formats (OpenAI, OTel, LangGraph, Anthropic sessions, VCR, native) into one typed graph; compiling plain-English contracts into Python + JSON Schema + Büchi automata; running the whole stack in 0.01ms in-process so you can actually put it in the dispatcher. The YAML is the UX. The engine is not the YAML.

Nothing. Contracts are YAML, not weights. The Gate fires on tool calls, not on model probabilities, so moving Sonnet 3.5 → Sonnet 4.6 → Opus 4.7 doesn’t invalidate a single line of your .trajeval.yml. Your LLM-as-judge eval set has to be re-verified on every model bump; ours doesn’t.

Self-host today, two paths. Basic gate — the runtime guard, the SDK, adapters, assertions, and the CLI — is MIT and installable with no invite: pip install git+https://github.com/abhishek5878/fewwords.git. Runs in your process; zero data leaves your VPC. Full self-host stack (adds the drift layer, the contract-suggestion engine, the 14-incident corpus, the full benchmark harness) is invite-only during early access — email for a repo invite, then pip install "git+ssh://git@github.com/abhishek5878/fewword_ai.git" or uv sync from a clone. PII redaction hooks are in the adapter layer on both paths (apply before persistence). SOC 2 Type II on the hosted endpoint is in progress for Q3 2026 — if you need it sooner, both self-host paths have no egress and no audit surface, because there’s no data crossing a trust boundary. SAML/OIDC on the hosted dashboard is the next enterprise line item; email for the roadmap.

0.01 ms median, 0.03 ms p99 on the warm in-process path. Hard upper bound under 1 ms on a cold subprocess. The Gate is pure Python — no model calls, no network I/O, no GPU. If you can afford the tool call, you can afford the Gate.

trajeval init scans your repo for framework signals plus any traces you have on disk and synthesises a starter .trajeval.yml tight to observed behaviour. You edit from there. Each contract line compiles to a check; you can see exactly what it blocks before you ship it. The point is a conservative default that you loosen, not a permissive default you tighten after an incident.

MIT. Source is in early-access while we onboard the first paying customers — email for a repo invite and you get write access the same day. The Gate, the adapters, the benchmark harness, and the CLI all ship under the same license. We charge for hosted dashboards, the Ledger history, and concierge onboarding — not for the enforcement primitive.

The Gate and the CLI are free forever. The hosted analyzer at /analyze is free for single traces. Concierge onboarding — we take one trace, ship back a Dossier with the YAML that would have caught your last incident — is free for the first five teams that send one. Paid tiers land once the first five have validated the product; we’re not selling a price page we can’t back up yet.

You shouldn’t, yet. What you can trust: fourteen public production incidents reconstructed from postmortems, every one caught by a 1–14 line YAML, receipts published on-domain at /benchmarks. 1,980 real tau-bench trajectories scored, 80–82% agreement with a strong auditor vs 54–62% for Claude Sonnet 4.6 as judge, at 55,000× lower latency and $0 API cost. The receipts exist; the scale doesn’t, yet. That’s the honest pitch.

Three things. One: the engine compiles plain-English to three different check types (Python expression, JSON Schema, LTL/Büchi) behind one YAML, so you don’t need a formal-methods specialist to author a rule. Two: the Gate hits 0.01ms in-process, which is the only latency profile that makes pre-execution enforcement viable outside a research lab. Three: the Corpus — fourteen reconstructed incidents with trace + YAML + latency — is on-domain evidence, not synthetic ARC-style benchmarks. The academic work (ToolGate, Solver-Aided, TraceSafe, Verifiably Safe, all 2026) validates the axis. We’re the one you can pip-install.