Product · Sentinel Router v0.2
Sentinel Router
Cuts frontier LLM spend for unit-tested agentic coding without sacrificing quality — cheap first, cross-model rescue on fail, frontier only on the original task (EOF-cold).
Install Sentinel Router by cloning https://github.com/tricombinator-tec/Sentinel_Router_Alpha and following the setup steps in its AGENTS.md file.
The problem
Agentic coding pipelines typically send every subtask to a frontier model. That is simple to operate and expensive. A large share of those subtasks are well-specified and would pass on a cheaper local model.
Routing everything to a cheap model is not a fix. Failures are uneven: some easy tasks survive, others fail silently, and accuracy drops in ways that are hard to predict or bound. The useful question is not “cheap or frontier,” but which tasks need the expensive pass.
How it works
EOF-cold + cross-model rescue
- Cheap first pass. A local coding model, served with Ollama by default, writes the solution. Unit tests run. No frontier call is made yet.
- Same-model repair, then a different cheap model. On fail, the same model gets one repair pass (failed code, error, and tests). If that still fails, a second cheap model gets the same repair template.
- Frontier only on the original task (EOF-cold). If rescue still fails, a frontier model sees the original task only — not the failed attempts. Same-model continuation on leftover budget is a measured control, not the product path.
That flow is what v0.2 shipped. Escalation is driven by unit-test fail after a cheap attempt, not by a pre-generation difficulty score.
Validated results
Audited evaluation, not a marketing estimate
0.331
Cost ratio vs. always-frontier · Suite A n = 158 · 155/158 vs 150/158
Suite A (held-out commercial-style coding tasks, n = 158, paired against always-frontier): Sentinel Router passed 155/158 vs 150/158 for frontier-only. Cost was $0.094 / $0.283 = 0.331. Pre-registered quality and cost gates both passed.
On Suite A traces that stalled past cheap self-repair, cross-model rescue recovered 28/55 vs 2/55 for same-model continuation. Suite B (HumanEval hard-fail slice, n = 21): rescue recovered 8/21 vs 0/21 continuation.
v0.1 on the same n = 158 catalog (EOF-cold, no rescue) cleared quality and cost at 151/158 vs 150/158 and a cost ratio of 0.54. Rescue sits between “cheap gave up” and “call frontier”; it does not replace EOF-cold.
Switch
Operating modes
Sentinel Router has an explicit --mode switch, the
same way you would pick an effort level on other agent tools.
Only direct is behind the Suite A/B gates above.
- direct — Validated; the default. EOF-cold + cross-model rescue. This is the architecture behind the numbers above.
- planned — Implemented, tested, did not beat direct. On a held-out underspecified set (n = 21), planned passed 6/21 vs 9/21 for direct on the same prompts, at 3.44× the cost. It is not a validated product path. RESEARCH.md
- handback — Implemented. Escalation returns a JSON payload to the calling agent instead of the router calling a frontier API. Correctness-gated (payload complete, frontier never called). No Suite A/B quality or cost numbers. RESEARCH.md
sentinel-router modes
sentinel-router run --mode directDo not cite the Suite A/B figures for planned or handback. Planned has its own gate; handback is an integration mode, not a quality/cost result.
Where else it applies
Same company, separate products
Escalate-on-fail is not a coding-only idea. Related controls — act under a cheaper policy, then escalate when a cheap attempt has failed — have been validated in AML alert triage and in quantitative trading.
The public repo is the coding cascade only. AML (Chitin) and trading are separate products; they are not in Sentinel_Router_Alpha.
Evaluate
Request a technical review
For engineering leads, AI infra teams, and accelerator or partner reviewers who want the eval protocol, not a pitch deck. Or clone the repo and follow AGENTS.md.