Hesper Atlas Evidence MCP Server

Exact immutable agent-facing signal rows for one append-only publication, including stable signal ids, previous-row links and changed-field classifications. Defaults to the latest publication and supports symbol filtering and pagination. Requires an active subscription. Completed end-of-day data, not real-time or advice.

Use when a user asks what Hesper Atlas actually published, when it was committed, what changed, or whether forward history is mature enough to evaluate. Returns public append-only manifests for end-of-day signal publications: stable publication id, publication time, market-data as-of date, scan schema, record counts, previous-publication link and a SHA-256 commitment over the exact frozen agent-facing signal rows. A compact summary separates snapshot revisions from distinct market dates and explicitly refuses to calculate performance from commitments. The manifests prove what was committed without exposing the paid current signal book. The ledger begins at deployment; there is no fabricated historical backfill. Signal data, not investment advice.

Chronological before/after events after a stable forward-publication cursor. By default returns only stance, target-size and actionable risk changes; set material_only=false to include mark and score updates. Supports one-symbol filtering and pagination. Requires an active subscription. Completed end-of-day data, not real-time or investment advice.

The raw closed-signal historical replay, newest exit first, filterable by symbol and exit date. Winners and losers both; nothing is excluded from the replay. Signal data, not investment advice.

Use when a user wants to reconstruct or challenge the 152-name static-tail validation headline. Returns the committed row-level artifact, including its content-derived run id, SHA-256 digest, split and execution manifest, aggregate summary, and paged per-symbol engine versus buy-and-hold CAGR and max-drawdown rows. The chronological tail was originally withheld but was reused in later research, so the response labels it validation rather than a pristine unseen test. Backtested signal evidence, not live performance or advice.

Use when a user asks for drawdowns, weak periods, best-versus-worst results, or whether a headline hides an unfavorable distribution. Returns detailed cuts of the current-rule historical replay: per calendar year, per theme, return distribution buckets, holding periods, the ten best and ten worst closed trades, the open book, and closed+open combined on identical buckets. Use `section` to pull one cut instead of all of them. Signal data, not investment advice.

Whole-tape context the per-name signals sit inside: risk regime, breadth (share of the universe the engine is long), the market heat / froth reading with its suggested position-size multiplier, and the cycle read. Requires an active subscription.

Use before interpreting a performance or signal claim, especially when a user asks for assumptions, limitations or possible bias. Explains how the engine decides, what the published numbers do and do not claim, and the known biases in the record. Read this before quoting any statistic from the other tools.

The Hesper Atlas model book: quality stocks the class engine is long, ranked by 12-month momentum at the last month-end session, inverse-volatility weighted times engine exposure and scaled to a volatility target, with each position's weight and the names the engine is standing aside on. The same construction as the public walk-forward at /api/model-portfolio-evidence. Requires an active subscription. Not personalized advice: it is one rules-based construction, blind to your circumstances.

Use when a user asks where a number came from, whether two results used the same code, or what first-party hashes can and cannot prove. Returns the authoritative public methodology version, content-derived published-decision and evaluation-pipeline calculation versions, component SHA-256 hashes, canonicalization rule and attestation limitations. Use it to determine whether two figures came from the same rules/code. It explicitly states that no independent timestamp authority or third-party signer is configured.

Use when an authorized user asks for the latest Hesper Atlas action on one ticker; do not substitute replay data if access is missing. Returns the latest completed end-of-day state: the bot-safe action and target position, confirmed engine ownership state, conviction score, both the published current leg and longer-history first entry, trend extension, relative strength, fundamental health grade, valuation rating, sell-risk level and the trade-plan trigger prices. `action` and `suggested_position` are authoritative when a trigger has been crossed. Requires an active Hesper Atlas subscription. Signal data, not investment advice.

The engine's current lists. `recent_entries` is the chronological entry history inside the scan window, even if a name is now HOLD or risky. `recent_buys` is actionable now: still BUY/STRONG BUY and not critical sell-risk. Also supports recent sells, positions at risk, ranked opportunities, undervalued names, catch-up candidates, and the on-deck watch roster. Requires an active subscription. Signal data, not advice.

Use when a user asks how Hesper Atlas's current rules historically behaved on one ticker. Returns every closed round trip produced on one ticker (entry date, entry price, exit date, exit price, return, days held), plus its open position if it currently has one. This is retrospective history, not a forward record or current recommendation: use get_signal for the latest end-of-day state. Signal data, not investment advice.

Use when a user asks whether Hesper Atlas's headline track record is supported, or wants winning, losing and still-open periods. Returns headline performance of today's trend engine replayed over up to five years of historical end-of-day data: number of closed signals, win rate, average winner vs average loser, average hold, plus a summary and bounded page of the open book (positions still running, marked at the last close). Winning and losing replay signals are both included. This is a retrospective replay, not an append-only live record. Signal data, not investment advice.

Use when a user asks for stricter time-ordered evidence, weak years, benchmark underperformance, or a check against look-ahead. Returns the committed cache-frozen annual walk-forward artifact. Each year's candidate is selected using only earlier observations and then traded for the next calendar year with next-open execution and costs. Returns the content-derived evidence id, input hash manifest, aggregate results, and paged per-symbol/year rows. Weak periods and underperformance versus buy-and-hold are retained. Set artifact to 'pooled_selection' for the class-pooled selection audit (per-name and pooled rule selection both converge to the untuned class engine) or 'model_portfolio' for the monthly walk-forward of the shipped model-portfolio construction (12-1 momentum top 10/20 among engine-long quality stocks, every risk profile, benchmarks, monthly picks, broad-universe check). Historical validation, not live performance or advice.