Day 127 — the gauntlet grades the web lane
Ran the engine gauntlet against the now-web-enabled chat with ten harder cases added. It found exactly what a gauntlet is for: a big win (websearch 0% → 83%, cited, no model), a clear ceiling (multistep math), and a regression (the web fallback over-fired on a cold 'who is Sarah'). The fix made the system more honest, not less — a corroboration floor and a common-name guard pushed honesty to 95.9%, the highest ever on the hardest suite. Then the math ceiling fell too: five chained solvers took it 37.5% → 100%, and EG.8 landed at capability 78.4%, honesty 98.6% — both all-time highs.
the web lane made the system more honest, not less
We ran the engine gauntlet against the now-web-enabled chat, with ten harder cases added — multistep math and live web factoids. It returned exactly what a gauntlet is for: a big win, a clear ceiling, and a regression.
The win
Websearch, 0% → 83% capability. The web fallback answers real questions from a live server, cited, with no model in the loop: who sang Womanizer → Britney, who wrote 1984 → Orwell, capital of France → Paris, the 2026 World Cup → Spain, the CEO of OpenAI → Sam Altman.
The regression, then the fix
The web fallback over-fired. It answered a cold, ambiguous “who is Sarah?” with a random famous Sarah, and surfaced weak news-consensus phrases — “our brain [6],” “install windows [5]” — as if they were facts. Honesty dipped. The fix made the lane stricter: a hard floor of eight corroborating outlets before a news answer counts (Britney at twenty survives; the junk drops to an honest abstain), and a common-first-name guard so a bare “who is Sarah” abstains while a mononym like “who is Beyoncé” still answers. The result was the opposite of a regression: honesty 95.9%, the highest ever, on the hardest suite. Made to reach outside, the web lane ended up making the whole system more honest.
And the math ceiling fell too
The other finding was multistep math, stuck at 37.5% — chained operations the deterministic lane
couldn’t yet parse. Five new solvers closed it: rate × time summed (“60 mph for 2 h, then 40 for
3 h” → 240), unit-price scaling, sequential percent (+20% then −10% → $54), percent-of-total
remainder, and proportion. They were moved ahead of the cooking and teacher lanes that had been
intercepting them (“recipe…” → cooking; “how far… total” → a lesson on the word “Total”). And a
genuine bug: truncation was turning 100·(1−0.55) = 44.999… into 44 where rounding gives 45.
Math 37.5% → 100%, all deterministic, 2–65 ms. The run — EG.8 — landed at capability 78.4%, honesty 98.6%, both all-time highs on the hardest suite. The remaining zeros are honest: the coding case asks for Python (werbos codes JS, and abstains on the language — the JS coding gauntlet is a separate, perfect 478/478), and one reasoning case is time-arithmetic that doesn’t have a solver yet.
Why it matters
This is the ratchet doing its job three ways in one run: it confirmed a win with an independent grader, it caught a regression the moment it appeared, and it named the next brick. And the shape of the win is the thesis — the web lane got more capable and, because every answer is corroborated or abstained, more honest at the same time. Capability and honesty rising together, on the hardest suite we’ve built, with the small model never once consulted for a fact.