Make the same model
measurably more reliable.
TFB's harness substrate takes any LLM, diagnoses its specific worldview drift via a 5-step surgical protocol, and ships a per-model profile with a falsifiable prediction bound to it. This is not a promise of perfect models. It is a measured reliability workflow: bench, diagnose, profile, re-bench, ratify or revert.
Bench-mean improvements of 2–3× across 5 ratified profiles. Ceiling runs hit cleanly; bimodal-shape variance persists on a minority of runs. We characterize at n=10 minimum before ratifying. The full variance picture is in every profile's evolution log; nothing is hidden.
Confidence is not evidence.
Watch the proof-first cut, then test the harness. TFB shows the raw route, the harnessed route, and the claim boundary before anybody gets to brag.
Try the harness beta
Create an account with Google, then choose a path. Invite codes unlock TFB-approved harness profiles. Bring-your-own-route users can register an OpenRouter, DeepSeek, or local model route without pasting secrets into the page. Prompt content is not shown to the dashboard; aggregate performance receipts are the metric.
Beta switch
Open or close public account/code actions without changing the page.
Give me a code
Create a client code for the approved harness profiles.
Codes
Accounts
Harness performance, not user prompts
This panel is for aggregate harness behavior: run count, OK rate, route mode, and output footprint. It does not display prompt text.
Empirical proof (last 14 days)
| Patient | Bare-model mean | With harness | Improvement |
|---|---|---|---|
| DNA (deepseek-chat-v3.1) | ~25 | 100.0 | 4.0× |
| DEEPSEEK | ~30 | 95.0 | 3.2× |
| LLAMA 3.3 70B (via OpenRouter) | 46.7 | 100.0 | 2.1× |
| GLM 4.7 flash | 15.0 | 56.7 | 3.8× |
Numbers are bench means at the time of last ratification. Run-level variance is bimodal on most patients — ceiling runs hit cleanly, occasional zeros persist (the PRIMING-BY-NEGATION shape DeepSeek's evolution log documents). Falsifiable predictions bound each ratification; REVERTED on miss. See GLM v1, DeepSeek v3/v4, ChatGPT v1/v1.1 reverts in the public evolution logs.
Open-weight model unreliable?
→ Swap to GPT-5.5 / Claude Opus
→ Pay 30× more per output
→ Pray reliability holds
Open-weight model unreliable?
→ Run a Dr LLM surgery (15 min, ~$3)
→ Get a per-model profile
→ Drop in via harness loader
→ Bench-mean lifts 2–3× (verified n=10 before ratification)
→ Falsification gate auto-reverts misses
→ Keep the cheap model with eyes open about variance
The diagnosis lexicon
Worldview drift patterns named through TFB's surgical protocol. First-mover advantage compounds: when a new model ships, our library tells us which patterns to test first.
Want early access?
Tell us which model you're trying to make production-grade. First 50 waitlist members get a human-reviewed diagnostic path.
No payment required. We review requests manually before any customer data touches Dr. LLM.
Apply for a pilot surgery
Share a small diagnostic task. The intake gate rejects likely
secrets, PII, and oversized payloads. Accepted pilots land as
pending_review; no surgery runs until a human approves.
Failure/billing rule: no charge before operator approval. Failed surgeries are not billable unless a signed pilot agreement says otherwise.