What changed in the engine, build by build

The AI Automation Researcher in v1.8.12026-09-08

A capability note rather than a version bump: the engine is unchanged at v1.8.1 · build g7f3a2c. Everything below runs in the free notebook demo on simulated data — no account, nothing installed.

• An AI Automation Researcher control on the standing bar above the sheet (N, or from the command palette): agents that compose, run and validate candidate notebooks on their own schedule
• A six-question wizard — risk profile · direction · timeframe or let it choose · edge families · budget in configurations · the house limits it cannot touch — each answer printing what it costs before you take it, and a review step that states what it will do, why, what it will cost and what it will never do. It can be left half-answered and resumed
• Eight steps a candidate, each logged with what it did, why, and the number: build · run · price the cost model at 1 bp and 45 bps · sweep the neighbourhood · race 400 random entries · cut six folds · resample 1,000 paths · report, including reporting that it found nothing
• An overview: agents with budget spent of budget set, a ranked results list, a shortlist of what still clears every check, and a one-click list of what it got wrong — candidates that looked good early and failed a later fold
• A library of fourteen estimators and tests and every dataset the notebook reads, each stating what it assumes and where it fails; every candidate names the dataset it was measured on

• Every configuration a researcher evaluates goes on the multiple-testing counter in cell [7], exactly like one you set by hand, and cell [7] now breaks the count out three ways — by hand · from the sweep · from the researcher. Measured on one saved assignment: 49 → 121 configurations, the agent reporting 72 spent, and the deflated Sharpe of your own unchanged result falling 76.9 % → 65.8 %
• The results list is ranked on the out-of-sample Sharpe the folds printed minus the noise bar in force when the candidate turned up — not raw Sharpe, which is what a long search always maximises. Raw Sharpe is still available as a sort, and the list says in words why it is the wrong one
• The shortlist is re-read every time it is drawn: three of the six checks move with the configuration counter, so a candidate that cleared them at 54 configurations comes off the list when it stops clearing them at 400, and the list names it

• A researcher with no budget ceiling cannot be configured. There is no unlimited option in the wizard and there will not be one
• Undo takes a candidate off the results list and refunds nothing: the budget stays spent and the counter does not move back, because the configurations were evaluated
• It will not move a parameter you set, will not size above quarter-Kelly, will not model leverage, will not search a very short-term or day-trading variant, and will not put a number on what anything returns

• A researcher left running for a long session makes every result in the workspace harder to believe — including its own. That is the arithmetic, not a defect, but it is a reason to set a small budget and read the counter rather than leaving one armed and forgetting it
• Assignments are remembered in this browser; their spend is not, because the configuration counter starts again on a reload. A restored assignment comes back stopped and says why
• The six folds and the resample cost no configurations but spend the sample: data a researcher has tested on is no longer untouched data for the next question you ask

The quant reviewer in v1.8.12026-09-08

A capability note rather than a version bump: the engine is unchanged at v1.8.1 · build g7f3a2c. This is what the reviewer does, and what it costs. Everything below runs in the free guest notebook — no account, nothing installed.

• The ✦ key in every cell rail: the same reviewer, entered already holding that cell, with three to five actions written for it
• Six analyst plans it runs itself — is this a real edge · is it better than luck · does it survive its costs · which parameter actually matters · where does it break · can I make it more robust
• Four standing mandates, said once: tell me when something I am running stops being an edge · keep watching my search count · re-validate anything I adopt · only speak when something actually breaks
• A second register in the research log (cell [12]) for what the reviewer did unasked, with what it checked, what it found and what it did — exportable as CSV
• Zero-input onboarding: one question, five answers, and the notebook configures itself — then states what it chose, why, and what the choice costs

• Every configuration the reviewer prices now counts against your multiple-testing bar, tagged with who ran it. On the guest notebook's default run, asking “is this a real edge?” takes cell [7] from 49 configurations · deflated Sharpe 76.9 % to 50 configurations · deflated Sharpe 76.7 %. Autonomy is not free, and the counter says so
• Undoing a standing-review action restores the parameter but not the count — the log now states this outright: a configuration it already priced cannot be un-counted
• Repeated identical “nothing crossed” entries fold into one row with a consecutive-check counter rather than filling the record with duplicates

• It will not move a parameter you set. Where it wants one moved it asks, and the log reads “Reported it and stopped. I do not move a parameter you set.”
• It will not size above quarter-Kelly, will not model leverage, and will not put a number on what a strategy returns — monthly, annually or as a range
• It will not rank two configurations whose gap is inside the standard error, and it will not plan on a result the engine has disowned or while a run is in flight

• Every verdict is conditional on one sample and one seed. Its best verdict says so in those words
• A standing mandate left armed through a long session contributes trials you did not think of as searches — watch cell [7], and pause it when you are only exploring
• Written up in full: WP 2026-09-08 — what a quant reviewer can be handed, and the four things it gets wrong

v1.8.1 latest internal build2026-08-30

• Monte Carlo plot cap: fans above 1,000 paths render a density band plus the P5/P50/P95 lines instead of every path — six times faster on 10k-path runs
• Portfolio-level slippage: same-bar legs are aggregated per venue and impact is computed on the total (config costs.portfolio_level)
• Trades tab: a per-trade realized-slippage column

• The walk-forward efficiency ratio showed n/a when the last fold ended exactly on the data-range end
• Heatmap month labels shifted by one in ranges starting mid-year
• macOS: the engine did not restart after sleep

• Default block size for the block bootstrap: 10 → 20 trades
• The report header now prints both the engine hash and the config hash