monte-neo repair: repair defects without overfitting

Every other command says what is wrong. repair fixes it, and shows that the fix is not a new curve fit. It also says so when there is nothing left to fix.

monte-neo repair --demo                                    # three strategies, three honest outcomes
monte-neo repair --ohlcv data.csv --strategy s.py --out repair/ [--commission-bps 10] [--rebalance-every 21]

Optimisers maximise the return, so they overfit. repair minimises defects, from a short list of declared repairs, and lets a part of the history it has never seen decide.

The protocol

1. Diagnose the original on all the data, as verify does.
2. Stop on defects no transform can honestly repair: broken data, survivorship, files or network read by the strategy, a non-deterministic signal, a model whose profit lives in fitting on the whole file, an edge that exists only before a language model's training cutoff. The outcome is NOT_REPAIRABLE, with what you need to do.
3. Correct look-ahead and repainting first. Look-ahead gets causal rewrites of the source (the same rules as --suggest-fix); repainting gets a one-bar delay. These have no parameters, so there is nothing to choose. If the leak is still there afterwards, the outcome is NOT_REPAIRABLE.
4. Choose on the first 70 % of the history only, among a few declared repairs, by the number of defects left (a fail counts 3, a warning 1), never by return. Every candidate is added to the number of trials, so the Deflated Sharpe gets stricter with each one.
5. Open the lockbox once: the chosen version on the last 30 %. The outcome is REPAIRED only when the whole-history certificate passes and the lockbox still earns after costs.

Outcome	Exit code	Meaning
REPAIRED	0	Defects removed; the lockbox (data the choice never saw) still earns after costs and is not rejected. It can still be NEEDS_MORE_EVIDENCE: read its failed checks
REPAIRED_NO_EDGE	1	Defects removed, and no edge is left: the profit was in the defect
NOT_REPAIRABLE	2	What blocks it and what you need to do
NOTHING_TO_REPAIR	4	No defect the catalogue addresses; the strategy is not touched

The catalogue

Id	Defect	Repair (fixed values)
R1	Look-ahead (shift(-1), centred windows, bfill, whole-series statistics)	Causal rewrite of the source
R2	Repainting	Act one bar later
R5	Costs and turnover eat the profit; the edge lives in the next fill	Smoothing (span 5, 10, 20), a minimum holding period (5, 10, 20 bars), a no-trade band (0.25)
R7	The result depends on the rebalance day (--rebalance-every)	Four staggered tranches
R8	The profit is market exposure	Cross-sectional neutralisation (universes)
R9	Trades are large against the volume	Weight cap 0.1, no positions in the least liquid 20 %
R11	Unstable risk, one period makes the profit, a policy breach	Volatility target 10 % a year, cap 0.5

Pairs of repairs from different families are tried too, up to --budget (24 by default).

What you get

--out holds:

• strategy.repaired.py: the corrected original, renamed, the repair functions it uses, and a new signal(df). It is self-contained and runs without Monte-Neo.
• strategy.corrected.py: the original after the look-ahead rewrite, before any economic repair (the repair loads it the way verify loads a strategy file).
• repair.diff, certificate.json (the repaired strategy on all the data, with the trials counted).
• repair-report.json and repair-report.html: before and after, the lockbox, every candidate and its defect count.

The original file is never changed. Recheck the certificate with monte-neo verify --recheck repair/certificate.json --ohlcv data.csv --strategy repair/strategy.repaired.py.

What the demo shows

• leak: reads tomorrow's close. R1 rewrites it and R2 delays it. REPAIRED_NO_EDGE: on a random walk, the profit was the leak.
• flicker: a real 40-day trend signal that flips almost daily, with 10 bps costs. It loses 85 % before the repair. Smoothing (span 20) is chosen on the first 70 %, and the lockbox earns after costs: REPAIRED.
• unseeded: random positions without a seed. NOT_REPAIRABLE: seed every random generator.

How it is tested

Zoo A16 runs repair on every leaking strategy of the Trap Suite on a random walk. A REPAIRED there would mean the repair invented an edge. In the full run, 68 leaking strategies were repaired with no REPAIRED outcome, and 9 were skipped because they need other data columns. The original files are checked to be unchanged.

Do not run repair again and again on the same data to look for a better outcome. That is the search the protocol exists to prevent.

Not investment advice: a research example, not a call to trade or a promise of profit. Simulated results do not predict future returns.