Algorithm
Architecture and diagrams | Testing | Agent workflows
Algorithm 1 uses numbered, IEEE-style pseudocode to describe the complete research workflow. It joins implemented Python paths and host-agent command instructions into one reference. It is not one executable function, a new controller, or a claim of IEEE publication.
A is the local artifact store. B bounds Python graph steps. K bounds host-agent attempts.
M contains missing required artifacts. F contains recorded graph failures. L is an optional summary model.
The host-run watch and completion commands are distinct from the Python graph.
The hosted implementation is outside this repository. Its branch shows the documented connection boundary only.
ALGORITHM 1: OPTION DESK RESEARCH, CONTROL AND REPORTING
Input: request Q; operation mode; skill edition; artifact store A;
graph step budget B; host attempt limit K; optional summary model L
Output: artifacts, quality observations, outcome and grounded report
001 Select the skill edition for the available local or hosted tools.
002 Read the relevant SKILL.md instructions and reporting requirements.
003 IF the tools run on the hosted service THEN
004 Connect through the separately configured remote MCP service.
005 Submit only operations exposed by that service with permitted input.
006 Receive its result; apply REPORT below; RETURN.
007 END IF
008 Check local tool availability and select the artifact directory.
009 IF Q needs an upload schema THEN return option_snapshot_schema.
010 IF an application uses the optional router prompt THEN
011 Assemble the router instruction and Q; obtain a proposed tool call.
012 The caller checks the proposal and invokes a supported local tool.
013 Do not use this model proposal to route the bounded Python graph.
014 END IF
015 Dispatch through CLI handlers, local MCP, or optional agent bindings.
016 Check each tool call for required arguments and supported names.
017 For a user snapshot, check rights, source, time and contract fields.
018 For a provider request, resolve the capability and check access and availability.
019 An explicit provider is strict by default; report failure if unavailable.
020 Route shell calculations through engine_bridge.
021 Preserve IV provenance; count missing IV and skipped calculations.
022 CASE mode OF
023 SINGLE_RESEARCH:
024 Run the requested handler using its declared inputs.
025 Build only structures supported by the available contracts, quotes and model inputs.
026 Retain payoff assumptions, friction, exclusions and degradation.
027 WHEN the requested handler is simulation:
028 Load permitted history and convert closes to daily log returns.
029 Require enough returns for the GARCH fit.
030 Fit GARCH-t with coordinate-wise Metropolis proposals.
031 Adapt step sizes during burn-in; freeze them for retained draws.
032 Compute split R-hat, per-parameter ESS and sampling acceptance.
033 Record convergence status and its diagnostic thresholds.
034 For each predictive path, draw parameters and independent shocks.
035 Propagate log price and conditional variance over the horizon.
036 Discard and count paths rejected by finite-value and upper-log-price guards; fail if none survive.
037 Compute the quantile fan and empirical underlying VaR/ES.
038 Evaluate intrinsic structure payoffs at horizon terminals; expiry alignment and surviving time value are not enforced.
039 Preserve diagnostics even when the posterior did not converge.
040 Withhold unconverged quantiles from the answer, as the reporting rules require.
041 END WHEN
042 BOUNDED_GRAPH: [Python graph]
043 R := [chain, greeks, exposure, comparison].
044 Read A for the selected underlying and optional expiry.
045 M := kinds in R that are absent; s := initial steps_taken or zero.
046 F := recorded failures.
047 REPEAT
048 IF F is nonempty THEN outcome := failed; BREAK.
049 IF M is empty THEN outcome := complete; BREAK.
050 IF s >= B THEN outcome := exhausted; BREAK.
051 k := first missing kind in pipeline order.
052 IF runner k is absent THEN record failure; s := s + 1; CONTINUE.
053 Invoke runner k; count the attempt.
054 IF runner raises THEN record failure; CONTINUE.
055 Re-read A for the expected kind, underlying and expiry.
056 IF expected artifact is absent THEN record failure; CONTINUE.
057 Remove k from M; record progress and any degraded result.
058 UNTIL an outcome is assigned.
059 Re-read report context; if a complete result has no records, fail it.
060 Artifact presence does not establish a successful trading thesis.
061 COMPLETE_DESK: [Host instructions]
062 Select the requested expiry or the nearest listed expiry beyond a week.
063 FOR attempt := 1 TO host limit K DO
064 Read the six criteria directly from artifacts:
065 IV coverage >= 90%; graded rows >= 50;
066 both walls and a smile; rankable structures >= 5;
067 converged simulation with horizon >= ceil(days_to_expiry);
068 no degradation reason except disclosed IV fallback.
069 IF every criterion holds THEN report criteria met; BREAK.
070 IF a structural blocker exists THEN report it and STOP.
071 Run only missing or failing stages.
072 For sparse input, consider a farther expiry or a wider Greek band.
073 For unconverged sampling, try 4000 then 6000 draws.
074 Stop after two failures at 6000 draws.
075 END FOR
076 Report each criterion, its observed value and any unmet condition.
077 Keep these quality criteria separate from BOUNDED_GRAPH's presence checks.
078 WATCH: [Host instructions]
079 Read baseline values BEFORE refreshing the selected expiry.
080 Refresh chain, Greeks and exposure.
081 Compare available old/new values: spot > 1%, ATM IV > 1 point,
082 RR > 0.5 point, regime flip, wall strike change, new degradation.
083 Report a newly unavailable value as unavailable, not a numeric move.
084 Report material changes with previous values, or no material change.
085 Stop this invocation; the host controls any later recurrence.
086 Do not open, mark or close a position in WATCH.
087 BACKTEST:
088 Load permitted underlying history and validate its length.
089 FOR each eligible entry index after the lookback window DO
090 Estimate trailing centred annualized volatility from past returns.
091 Price a synthetic chain for holding_days / 252 years.
092 Build the requested compatible structure.
093 IF volatility or plan is unavailable THEN record skip; CONTINUE.
094 Settle the payoff at the historical exit underlying price.
095 Record entry legs, model premiums, capital at risk and P/L.
096 Record return on risk only where its denominator is defined.
097 Advance entries by entry_every trading-day indices.
098 END FOR
099 Sum returns in risk units; compute statistics and drawdown.
100 Set block length to max(1, ceil(holding_days / entry_every)).
101 Run block sign-flip randomization and moving-block bootstrap; report their assumptions and degenerate samples.
102 Include the buy-and-hold benchmark and model-premium limitations.
103 FORWARD_OPEN:
104 Read an explicit plan or the newest matching saved plan.
105 Record a new open paper position: ID, time, entry legs/prices,
106 entry spot, source plan and optional thesis; save the ledger.
107 FORWARD_MARK:
108 FOR each selected open paper position DO
109 Find the newest on-disk chain matching underlying and expiry.
110 IF absent THEN report unmarkable; keep position open; CONTINUE.
111 Match option legs by symbol, then strike/type; mark stock at spot.
112 Use available mid, bid/ask midpoint, or labelled last trade.
113 IF any leg lacks a mark THEN append unmarkable result.
114 ELSE append mark_value minus entry_value and quality notes.
115 END FOR
116 Save the ledger; report successful marks and unmarkable positions.
117 FORWARD_CLOSE:
118 Require an ID for an open paper position; reject missing/closed IDs.
119 Use explicit settlement price or newest matching on-disk spot.
120 Compute intrinsic-payoff settlement and record closed status/time.
121 Save the ledger. Expiry and quote freshness are not enforced here.
122 This settlement path does not price a surviving two-expiry leg.
123 FORWARD_STATUS:
124 Read open/closed positions and summarize paper outcomes.
125 END CASE
126 For artifact-producing handlers, validate the corresponding JSON schema.
127 Preserve provenance, missing counts, quality flags and model assumptions.
128 Use the artifact writer: archive replaced content by default, then
129 write through a temporary file and atomically replace the target.
130 The preceding writes occur within their handlers, before graph re-checks.
131 REPORT:
132 Assemble saved artifact context for the question or graph outcome.
133 IF a summary model L was supplied to the graph THEN
134 system := REPORTING_RULES.
135 The prompt helper can append caller-supplied deployment rules.
136 human := artifact context followed by the question.
137 Invoke L once for the final summary; do not rerun calculations.
138 ELSE assemble the deterministic graph summary or caller's report.
139 Surface degradation before numbers; give source, time and units.
140 Preserve missing values and the word unlimited.
141 State dealer-sign assumptions and model-premium limitations.
142 Withhold unconverged simulation quantiles under the reporting rules.
143 Present backtest context and uncertainty; give no trade recommendation.
144 Model compliance with these reporting requirements still needs verification.
145 RETURN artifacts, outcome, observations and report.
| Algorithm block | Evidence |
|---|---|
| Skills and execution surfaces | Local skills, hosted skills, installation guide |
| Local tool arguments | MCP server |
| Uploads and provider selection | Chain command, provider registry, provider access |
| Bounded graph and outcomes | graph.py |
| Host completion and watch policies | desk-complete.md, desk-watch.md |
| Historical backtests | Runner, statistics, CLI orchestration |
| Paper open, mark, close and status | Ledger command, mark and settlement math |
| Prompt assembly and final summary | prompts.py, graph report node |
| Validation and replacement | Contracts, artifact writer |
The Python graph checks artifact presence. The completion command asks the host to check six stronger quality conditions. Neither condition proves an edge or a profitable strategy.
The watch thresholds are instructions in a command file. Recurrence and attempt limits belong to the host. Prompt requirements do not prove that a model follows every instruction. The optional router prompt proposes a tool call. It does not execute the call or route the Python graph.
Backtests use model premiums and omit spreads, slippage, assignment and early exercise. The runner does not supply the second chain needed by two-expiry builders. Forward close uses intrinsic settlement and does not enforce expiry or snapshot freshness. It does not perform general two-expiry valuation.
The reviewed source revision is 79e8d7f. Public reference archives retain their original contents and historical claims.
The IV solver rounds its accepted candidate before returning it. The rounded IV need not reprice within the internal tolerance. Partial option quotes can still produce an ok friction verdict.
Simulation rejects paths at its finite-value and upper-log-price guards. Its structure callback applies intrinsic payoff at the requested horizon without aligning plan expiries or retaining a far leg's time value. Sample means of unbounded payoffs need not estimate a finite expectation under Student-t log returns.
Block sign-flip randomization requires symmetry under the allowed flips. Block methods retain within-block dependence but can lose it at boundaries. A bootstrap with one available block has no resampling variation; its zero-width interval cannot support a conclusion. The current even-sample median field selects the upper middle observation. These are documented runtime limits, not additional guards in the algorithm.