create-tradejs 3.1.22 → 3.1.23

package/dist/skill-bundle/.codex/skills/strategy-release/references/historical-hypothesis-audit.md

@@ -0,0 +1,163 @@
+# Historical hypothesis audit
+
+Use this audit before inventing release-core hypotheses. Its purpose is to
+prevent a new Codex task from forgetting a stronger prior core result, rerunning
+an already rejected idea, or treating an unrelated refactor as research.
+When the objective fingerprint changed, follow this inventory with the complete
+candidate revalidation board from
+[research-objective.md](research-objective.md); do not revalidate only the
+previous winner.
+
+## Contents
+
+1. Build the inventory
+2. Match commits to evidence
+3. Bridge prior results to the current experiment
+4. Prioritize untested hypotheses
+5. Required artifact
+
+## 1. Build the inventory
+
+Inspect the complete strategy history, the current strategy files, and every
+shared helper imported by that strategy:
+
+```bash
+git log --reverse --date=iso-strict \
+  --format='%H%x09%aI%x09%s' -- packages/strategies/src/<Strategy>
+git show --stat --patch <commit-sha> -- \
+  packages/strategies/src/<Strategy> <referenced-shared-paths>
+yarn research:core index
+```
+
+Also inspect `notes/<Strategy>/`, relevant `notes/CrossStrategy/` records, the
+core-research family ledger/index, Redis config snapshots cited by those
+records, and retained release artifacts. Search by full/short Git SHA, config
+field, canonical config hash, behavior/diff hash, research id, and run id. A
+missing Git SHA in a note is not proof that no evidence exists.
+
+Classify every behavior-relevant commit or dirty patch as exactly one of:
+
+- `verified-result`: complete reconciled evidence exists for the same behavior;
+- `rejected-result`: complete evidence exists and rejected it;
+- `partial-result`: useful evidence exists but is not selection-grade;
+- `untested-behavior`: a causal behavior change has no result evidence;
+- `refactor-no-hypothesis`: no intended decision/economic behavior change;
+- `superseded-duplicate`: the same causal delta was tested under another
+  commit/config lineage;
+- `unreconstructable`: the behavior or required point-in-time inputs cannot be
+  reproduced honestly.
+
+Do not infer the class from the commit subject. Inspect the patch. Config,
+filter, engine, order-plan, exit-lifecycle, risk, and deterministic-gate changes
+are hypothesis candidates; tests, formatting, module moves, and performance-only
+refactors are not unless the patch changed decisions or execution semantics.
+
+## 2. Match commits to evidence
+
+Match evidence semantically, not only by filename:
+
+- exact or documented dirty Git lineage;
+- canonical resolved core config and behavior/diff hash;
+- ordered universe and checksum;
+- half-open window and cache coverage;
+- fees, slippage, entry delay, interval, connector, and context lineage;
+- completed manifest, export hash, and Redis N/W/L/PnL reconciliation;
+- result status and whether the evidence was control, candidate, diagnostic, or
+  release-selection grade.
+
+If a current behavior was introduced by several commits, collapse them into one
+causal hypothesis record. If one commit introduced several independent
+mechanisms, split them. Never spend several trial slots on semantic duplicates.
+
+When a change is already part of the current control but was never ablated, the
+current baseline proves only its combined economics. Reconstruct an explicit
+off/on ablation before claiming that the historical mechanism helped.
+
+## 3. Bridge prior results to the current experiment
+
+Before freezing new families, create a bridge table containing every previously
+tested or shortlisted core/gate/direction/rescue result, deduplicated by
+behavior/config fingerprint:
+
+```text
+source lineage | behavior/config delta | window | universe | costs | ALL/LONG/SHORT |
+N | PnL | PF | PnL/trade | DD | cadence | evidence grade | current comparability
+```
+
+Classify comparability as:
+
+- `exact`: logic, config, universe, window, costs, and context match;
+- `bridge-required`: the causal config is reconstructable but one or more
+  experiment dimensions differ;
+- `diagnostic-only`: evidence cannot support a direct economic comparison.
+
+Do not compare raw PnL from 1500d/300 symbols with 1800d/507 symbols as if it
+were the same experiment. Equally, do not ignore the former. Re-score every
+compatible retained candidate under the frozen objective. For every
+`bridge-required` result that could reach the new Pareto frontier, rerun its
+exact causal config on the new frozen window/universe/cost contract before
+testing novel ideas. Attribute the delta to source/config, membership, window,
+costs, or context with matched metrics where possible.
+
+The current release control is not accepted merely because it is named `:ai`.
+If it is materially worse than a prior comparable result, explain and reproduce
+that regression first. A final claim that “nothing works” is invalid while a
+stronger prior result remains unbridged.
+
+## 4. Prioritize untested hypotheses
+
+Use reconstructable `untested-behavior` records before generating novel
+families. Select at most three causally distinct records as round-1 anchors,
+ranked by:
+
+1. strength and completeness of any adjacent/prior evidence;
+2. direct relevance to the current ALL/LONG/SHORT failure mode;
+3. causal clarity and point-in-time validity;
+4. support and cadence potential;
+5. difference from hypotheses already tested in the ledger.
+
+An untested historical behavior consumes the normal trial budget; it is not a
+free extra look. Metric-only re-scoring and exact bridge reruns of previously
+tested behavior do not consume new causal slots, but remain in the global
+multiple-testing ledger. Remaining records stay in the immutable backlog. They
+may motivate the post-round-3 rescue board when they address a selected seed's
+measured failure. A changed threshold, feature, payoff, direction policy, or
+other decision behavior is a new candidate; prefer it as an anchor or, if
+discovered later, spend a rescue slot. Rebuilding the unchanged control or
+normalizing metadata does not count as a candidate.
+Do not return `STOP_RESEARCH` or claim the strategy is exhausted while a
+reconstructable untested historical hypothesis with stronger evidence than the
+tested set remains unbridged and fits an available anchor/rescue slot. If the
+bounded budget is exhausted first, report the remaining backlog explicitly as
+`INSUFFICIENT_EVIDENCE`, not as market unsuitability.
+
+## 5. Required artifact
+
+Persist and hash a machine-readable inventory before round 1:
+
+```json
+{
+  "schema": "tradejs-strategy-hypothesis-inventory/v1",
+  "strategy": "<Strategy>",
+  "headGitSha": "<sha>",
+  "entries": [
+    {
+      "sourceCommits": ["<sha>"],
+      "causalFamily": "<family-or-null>",
+      "configOrBehaviorDelta": {},
+      "behaviorSha256": "<sha-or-null>",
+      "status": "verified-result|rejected-result|partial-result|untested-behavior|refactor-no-hypothesis|superseded-duplicate|unreconstructable",
+      "evidenceResearchIds": [],
+      "priorResult": null,
+      "comparability": "exact|bridge-required|diagnostic-only|null",
+      "decision": "round1-anchor|rescue-backlog|closed|excluded",
+      "reason": "<bounded reason>"
+    }
+  ],
+  "bridgeRuns": [],
+  "unresolvedBacklog": []
+}
+```
+
+Reference its SHA from every child spec, causal handoff, final note, and release
+decision. An absent or stale inventory makes the release evidence incomplete.

package/dist/skill-bundle/.codex/skills/strategy-release/references/professional-research-loop.md

@@ -0,0 +1,198 @@
+# Professional algo-research loop
+
+## Contents
+
+1. Write the trading thesis
+2. Revalidate the candidate history
+3. Build an opportunity map
+4. Construct a hypothesis portfolio
+5. Diagnose before adapting
+6. Think in compositions
+7. Make a professional decision
+
+The workflow is a decision framework, not a compliance checklist. Behave like
+the owner of the strategy's future expectancy: understand how the edge is
+supposed to work, find where the realized process breaks, and spend the bounded
+trial budget on the highest-information causal interventions.
+
+## 1. Write the trading thesis
+
+Before selecting variants, explain in plain language:
+
+- who is expected to be forced or mispriced;
+- what observable setup identifies that condition at signal time;
+- why the proposed entry timing should improve price or confirmation;
+- how the stop expresses thesis invalidation;
+- how profit should be realized and why the target/trail is achievable;
+- which market regimes should help or hurt;
+- what cadence and event independence the mechanism can realistically support.
+
+Compare this thesis with the current code, figures, traces, and realized trades.
+List every semantic gap. A strategy with no defensible thesis is not rescued by
+threshold search; use one bounded falsification family to test whether any edge
+exists, then retire it if the mechanism remains unsupported.
+
+## 2. Revalidate the candidate history
+
+Freeze the versioned objective from [research-objective.md](research-objective.md),
+then re-score every reconstructable historical behavior candidate before
+inventing a new one. A new objective changes which candidate is preferred; it
+does not erase prior trials or make an exposed tail unexposed.
+
+Separate three activities:
+
+- metric-only re-scoring of compatible normalized trades;
+- exact bridge reruns when the current frozen window/universe/cost contract
+  differs;
+- genuinely new behavior, which belongs in the new causal-family budget.
+
+Persist the revalidation board and global trial ledger. Use the resulting
+Pareto frontier as evidence for the opportunity map, but still execute new
+causal improvement rounds in the new lineage. Neither a prior `STOP_RESEARCH`
+nor a newly rescored winner substitutes for the requested research attempt.
+
+## 3. Build an opportunity map
+
+Decompose the performance failure across these intervention points:
+
+1. **Opportunity formation** — too few/many raw setups, duplicated symbols per
+   event, stale zones/pivots, universe or timeframe mismatch.
+2. **Entry selection and timing** — premature entry, confirmation delay,
+   distance/chase, quality/context discrimination, side asymmetry.
+3. **Risk geometry** — stop not tied to causal invalidation, target unreachable,
+   fee/slippage economics, position sizing inconsistency.
+4. **Position lifecycle** — winners returned, losers held, wrong opposite exit,
+   missing thesis-invalidation exit, occupancy/cooldown spillover.
+5. **Regime and direction** — edge exists only in one side or causal regime;
+   current gate hides it or mixes incompatible policies.
+6. **Concentration and capacity** — one timestamp, symbol, sector, or correlated
+   event supplies most PnL; event fan-out overstates independent support.
+7. **Execution/parity** — delayed fills, protection, fees, runtime state, or
+   package boundaries differ from the researched behavior.
+
+For each point record evidence, estimated economic impact, independent support,
+signal-time observability, implementation complexity, and a falsifying result.
+Use `unknown`, not invented precision.
+
+The opportunity map must identify:
+
+- the largest loss budget that a causal intervention can plausibly remove;
+- the strongest existing edge worth protecting;
+- the main cadence bottleneck;
+- the main tail/drawdown source;
+- the cheapest experiment that distinguishes two competing explanations.
+
+When one isolated global config change improves one side and worsens the other,
+record a directional-parameter opportunity. Use the dedicated classifier and
+[directional-parameter-split.md](directional-parameter-split.md) rather than
+discarding the field, accepting the aggregate compromise, or duplicating every
+parameter speculatively.
+
+Persist and hash this map before round 1. Do not confuse a report containing
+many metrics with a diagnosis that ranks actionable causes.
+
+## 4. Construct a hypothesis portfolio
+
+Choose three causally distinct families from the opportunity map, not from a
+generic parameter menu. The portfolio must include these roles when possible:
+
+- **Exploit** — preserve or amplify the strongest evidenced edge;
+- **Repair** — attack the largest attributable loss or lifecycle failure;
+- **Explore/falsify** — test a distinct market mechanism or the counter-thesis
+  most likely to invalidate the current explanation.
+
+Generate at least two candidate mechanisms for every role before choosing the
+anchor. Rank them qualitatively by:
+
+`expected economic impact × information gain × support ÷ complexity/risk`.
+
+Do not pretend this score is statistically precise. Its purpose is to force an
+explicit choice between meaningful experiments.
+
+Good variants change one causal transition: setup formation, confirmation,
+invalidation, payoff, exit, or gate decision. A bundle is allowed only when its
+parts are inseparable for the thesis and a matching ablation is included.
+
+Bad variants include:
+
+- adjacent thresholds chosen only because a nearby value looked profitable;
+- filters defined from exit reason, realized PnL, future regime, or delayed fill;
+- broad indicator combinations without a market-mechanism claim;
+- deleting a losing side from raw evidence;
+- reducing cadence until a few correlated winners remain.
+
+## 5. Diagnose before adapting
+
+After each round, answer these questions before creating children:
+
+1. Did the intended trace transition actually change?
+2. Was the PnL delta caused by matched-trade improvement, removed losers, added
+   winners, or occupancy spillover?
+3. Is the improvement independent across events, symbols, folds, and regimes?
+4. Did costs, drawdown, or tail losses move consistently with the thesis?
+5. Which observation supports or falsifies the mechanism?
+6. What is now the highest-value uncertainty?
+
+Update a belief ledger for every family:
+
+```json
+{
+  "family": "<id>",
+  "priorClaim": "<causal claim>",
+  "observations": ["<metric/trace/identity facts>"],
+  "mechanismVerdict": "supported|falsified|inconclusive",
+  "protectedEdge": "<edge the child must retain>",
+  "remainingFailure": "<dominant failure>",
+  "nextExperiment": "<single discriminating intervention>",
+  "falsifier": "<result that retires the family>"
+}
+```
+
+- If supported, refine the remaining failure without discarding the protected
+  edge.
+- If inconclusive, prefer an ablation or higher-information contrast over a
+  smaller threshold nudge.
+- If falsified, change mechanism or retire the family; do not rescue the label.
+
+Outcome-derived cohorts may diagnose where money was lost. Translate them into
+a signal-time observable hypothesis and validate it on a later partition; never
+turn the outcome cohort itself into a filter.
+
+## 6. Think in compositions
+
+The product is core plus deterministic gate, not a raw-core beauty contest.
+
+- A high-cadence weak core may be valuable if a causal gate retains enough
+  independent positive events.
+- A profitable but sparse gate is a clue, not a release: identify why it wins,
+  widen support through causal neighboring setups, or preserve it as one sleeve
+  of the same strategy composition.
+- A profitable direction mixed with a losing direction requires the explicit
+  direction-policy checkpoint.
+- A recent regime pocket may justify micro-forward monitoring, but never a
+  silently fitted historical rule.
+
+Optimize the hierarchy from [research-objective.md](research-objective.md):
+out-of-sample expectancy per risk after costs, selection-adjusted Sharpe,
+drawdown/tail robustness, temporal stability, independent support, and
+executable cadence. Never maximize PnL, Sharpe, win rate, a streak, a recent
+window, or cadence alone.
+
+## 7. Make a professional decision
+
+At every checkpoint choose one concrete action:
+
+- run the next highest-information experiment;
+- repair invalid evidence and rerun;
+- carry a side-qualified candidate into gate research;
+- select one immutable prospective composition;
+- start/prepare the authorized risk-1 micro-forward;
+- retire a falsified family with a stated reason.
+
+Do not answer with a passive blocker when a safe in-scope experiment can reduce
+the uncertainty. Do not build unrelated infrastructure merely because it would
+make the study ideal. Add infrastructure only when it is the smallest path to
+distinguish candidates or establish runtime parity.
+
+The final narrative must say what was learned about the market mechanism, not
+only which thresholds won.