candidate-1

1. Fits objectives — 4/5. It explicitly prefers 52.36% funded payout probability over 46.31% at 2% risk and acknowledges that its estimated $134 annual income per slot cannot meet “I need money.”
2. Evidence discipline — 2/5. Its $25.5 expected surplus and 0.12% wallet ruin omit the extra costs of its required news policy, while treating loss from entry as an acceptable daily-limit reading conflicts with rules.json’s previous-day equity basis.
3. Robustness — 2/5. It quantifies useful switches, including 3% → 4% assessment for only 0.86 percentage points more pass probability, but buys two accounts before resolving news costs and knowingly leaves overnight daily-equity exposure.
4. Actionable — 4/5. It specifies 3% and 2.5% assessment risks, 1.5% funded risk, two entry windows and a 13-purchase stop, although “request the full equity” should explicitly mean eligible profit.
5. Falsifiable — 2/5. Its 3% breach threshold can reject the assessment choice, but the claimed 93% power transfers effective sample sizes of 860 and 1,730 from pass-rate errors to a different, clustered breach statistic without justification.

candidate-2

1. Fits objectives — 4/5. It connects 52.4% payout probability, $276 median wallet drawdown and roughly 134 days to first cash, and recognizes that $36.8 expected receipts cannot justify a $38 purchase.
2. Evidence discipline — 2/5. It correctly identifies the 1.33 extra news round trips, but its daily-risk ceiling of 4.42% checks loss from entry rather than previous-day floating equity, and its 52.5% ±3.3 broker figure uses between-seed uncertainty.
3. Robustness — 3/5. Its cost gate and honour-rate sensitivity are useful, including log growth falling from +0.167 to −0.18, but the daily-equity omission and automatic expansion after one honoured payout leave material risks unresolved.
4. Actionable — 4/5. The purchase condition of cost share ≤4.5%, approximately ≤0.55 news asides per trade at 0.5-pip spread, is readily reusable, although the move to two accounts lacks a new-policy ruin calculation.
5. Falsifiable — 3/5. It declares a null, rejection thresholds of 41.2% and 48.0%, and 83% power for a five-point shortfall, but uncertainty in the estimated aside rate and covariance between paired arms are not incorporated.

candidate-3

1. Fits objectives — 5/5. It makes the drawdown–income trade-off explicit with $228 versus $276.40 median drawdown and 99 versus 66 model trades to first cash, while declining purchases when the news sensitivity produces only $36.81.
2. Evidence discipline — 4/5. It distinguishes 52.496% seed-average payouts from 50% pooled payouts and labels the news-cost transfer as sensitivity, although its test leaves the relationship between the lifetime null and censoring adjustment insufficiently specified.
3. Robustness — 4/5. It preserves tick enforcement, requires floating-equity and calendar controls, and moves allocation to zero when gates fail, but provides less quantitative reranking for spread and swap changes than candidate-1.
4. Actionable — 4/5. It clearly selects zero purchases now and a conditional single $38 account at 3%/1.5%, with concrete operational controls, but the payment assumption needed to release the purchase gate remains underspecified.
5. Falsifiable — 4/5. It declares Δ=0.24709, rejection outside 0.10850–0.38568 and approximately 94% design power, while honestly acknowledging that four calendar blocks cannot deliver that power.

Recommended base: candidate-3.

Its zero-purchase decision follows from the evidence rather than merely expressing caution: the required news intervention can reduce modeled receipts from $63.45 to $36.81 against a $38 fee, and the available broker does not establish compliance with floating-equity limits. It also best separates exact-model precision from empirical confidence and avoids transferring old concurrency results into the new policy.

Before adopting its test, specify the censored-data estimator and matching null: broker_map.json gives a lifetime difference of 0.24709 but a protocol-adjusted difference of approximately 0.23043; these cannot be interchanged. Its 94% power remains a prospective design target, not available historical power.

From candidate-1, graft:
- The overnight-equity diagnostic: its 4% assessment leaves only 0.11–0.23 stop-distances of overnight profit before a reversal can breach the daily limit; use this to design compliance checks, without importing its power claim.
- The spread sensitivity: funded risk switches to 1% at zero spread, while 1.5% remains preferred over the stated 0.25–1.0-pip range, subject to recomputation with compliant costs.

Reject candidate-1’s immediate $76 deployment, its untreated news costs, and its tolerance of a known daily-equity breach mechanism.

From candidate-2, graft:
- The explicit economic buffer: expected receipts per fee ≥1.33, corresponding to approximately ≤0.55 asides per trade at 0.5-pip spread, labeled as a policy threshold requiring compliant-cost verification.
- The honour-rate distinction: +0.006 wallet log growth at 80% honour versus −0.18 at 60% shows why expected-value breakeven alone is inadequate.

Reject candidate-2’s expansion to two accounts after one payment, its entry-based daily-loss check, and its unsupported suggestion that narrowing the calendar would likely halve aside frequency.

Verdict: Use candidate-3 as the base; graft candidate-1’s overnight and spread diagnostics and candidate-2’s economic and honour-rate gates, while retaining zero purchases until those gates are established.
tokens used
51,739
candidate-1

1. Fits objectives — 4/5. It explicitly prefers 52.36% funded payout probability over 46.31% at 2% risk and acknowledges that its estimated $134 annual income per slot cannot meet “I need money.”
2. Evidence discipline — 2/5. Its $25.5 expected surplus and 0.12% wallet ruin omit the extra costs of its required news policy, while treating loss from entry as an acceptable daily-limit reading conflicts with rules.json’s previous-day equity basis.
3. Robustness — 2/5. It quantifies useful switches, including 3% → 4% assessment for only 0.86 percentage points more pass probability, but buys two accounts before resolving news costs and knowingly leaves overnight daily-equity exposure.
4. Actionable — 4/5. It specifies 3% and 2.5% assessment risks, 1.5% funded risk, two entry windows and a 13-purchase stop, although “request the full equity” should explicitly mean eligible profit.
5. Falsifiable — 2/5. Its 3% breach threshold can reject the assessment choice, but the claimed 93% power transfers effective sample sizes of 860 and 1,730 from pass-rate errors to a different, clustered breach statistic without justification.

candidate-2

1. Fits objectives — 4/5. It connects 52.4% payout probability, $276 median wallet drawdown and roughly 134 days to first cash, and recognizes that $36.8 expected receipts cannot justify a $38 purchase.
2. Evidence discipline — 2/5. It correctly identifies the 1.33 extra news round trips, but its daily-risk ceiling of 4.42% checks loss from entry rather than previous-day floating equity, and its 52.5% ±3.3 broker figure uses between-seed uncertainty.
3. Robustness — 3/5. Its cost gate and honour-rate sensitivity are useful, including log growth falling from +0.167 to −0.18, but the daily-equity omission and automatic expansion after one honoured payout leave material risks unresolved.
4. Actionable — 4/5. The purchase condition of cost share ≤4.5%, approximately ≤0.55 news asides per trade at 0.5-pip spread, is readily reusable, although the move to two accounts lacks a new-policy ruin calculation.
5. Falsifiable — 3/5. It declares a null, rejection thresholds of 41.2% and 48.0%, and 83% power for a five-point shortfall, but uncertainty in the estimated aside rate and covariance between paired arms are not incorporated.

candidate-3

1. Fits objectives — 5/5. It makes the drawdown–income trade-off explicit with $228 versus $276.40 median drawdown and 99 versus 66 model trades to first cash, while declining purchases when the news sensitivity produces only $36.81.
2. Evidence discipline — 4/5. It distinguishes 52.496% seed-average payouts from 50% pooled payouts and labels the news-cost transfer as sensitivity, although its test leaves the relationship between the lifetime null and censoring adjustment insufficiently specified.
3. Robustness — 4/5. It preserves tick enforcement, requires floating-equity and calendar controls, and moves allocation to zero when gates fail, but provides less quantitative reranking for spread and swap changes than candidate-1.
4. Actionable — 4/5. It clearly selects zero purchases now and a conditional single $38 account at 3%/1.5%, with concrete operational controls, but the payment assumption needed to release the purchase gate remains underspecified.
5. Falsifiable — 4/5. It declares Δ=0.24709, rejection outside 0.10850–0.38568 and approximately 94% design power, while honestly acknowledging that four calendar blocks cannot deliver that power.

Recommended base: candidate-3.

Its zero-purchase decision follows from the evidence rather than merely expressing caution: the required news intervention can reduce modeled receipts from $63.45 to $36.81 against a $38 fee, and the available broker does not establish compliance with floating-equity limits. It also best separates exact-model precision from empirical confidence and avoids transferring old concurrency results into the new policy.

Before adopting its test, specify the censored-data estimator and matching null: broker_map.json gives a lifetime difference of 0.24709 but a protocol-adjusted difference of approximately 0.23043; these cannot be interchanged. Its 94% power remains a prospective design target, not available historical power.

From candidate-1, graft:
- The overnight-equity diagnostic: its 4% assessment leaves only 0.11–0.23 stop-distances of overnight profit before a reversal can breach the daily limit; use this to design compliance checks, without importing its power claim.
- The spread sensitivity: funded risk switches to 1% at zero spread, while 1.5% remains preferred over the stated 0.25–1.0-pip range, subject to recomputation with compliant costs.

Reject candidate-1’s immediate $76 deployment, its untreated news costs, and its tolerance of a known daily-equity breach mechanism.

From candidate-2, graft:
- The explicit economic buffer: expected receipts per fee ≥1.33, corresponding to approximately ≤0.55 asides per trade at 0.5-pip spread, labeled as a policy threshold requiring compliant-cost verification.
- The honour-rate distinction: +0.006 wallet log growth at 80% honour versus −0.18 at 60% shows why expected-value breakeven alone is inadequate.

Reject candidate-2’s expansion to two accounts after one payment, its entry-based daily-loss check, and its unsupported suggestion that narrowing the calendar would likely halve aside frequency.

Verdict: Use candidate-3 as the base; graft candidate-1’s overnight and spread diagnostics and candidate-2’s economic and honour-rate gates, while retaining zero purchases until those gates are established.