Built by people who think a comp ladder deserves the same engineering as a trading desk.

Bulk scan runs entirely in Safari over getUserMedia. Every ~180ms we sample the framing rect into a 120×<H> canvas, derive a luma map, and compute three signals:

• inner detail— stddev of luma inside the card-shaped guide rect (a real card has print + image variance; an empty desk doesn't)
• inner-vs-outer contrast— inner stddev minus outer stddev so a busy desk doesn't fake fill
• edge density— fraction of pixels along the rect's perimeter with a strong luma gradient (catches white-bordered cards on dark surfaces even when the card face is featureless)

All three must clear for ~1.2s of consecutive samples before we fire. Stack-bulk mode arms a 1.2s settling delay after each capture so the camera doesn't re-fire mid-flip on the residual scene. Concurrency on submit is capped at 3 in flight.

A single AI model reading its own work has a known failure mode: it agrees with itself. So we don't let the same model be both classifier and judge. Cardpulse uses Claude Haiku 4.5 — the faster, cheaper model — to read each forum post, news article, and performance event, and produce a sentiment score with a confidence weight.

Whenever Haiku's output is loud — high magnitude or low confidence — Claude Sonnet 4.6 (the slower, more careful model) takes a second pass. The auditor evaluates: does the article actually justify this magnitude? Is the kind right? Is this a confident-wrong answer? Three verdicts: pass, regenerate, reject. Rejected high-magnitude signals get logged for human review rather than silently dropped. Nothing big and bullish ever moves a price without paper trail.

The agent also runs a one-click scout on demand: search Reddit and ESPN for a player, score every fresh post, regenerate the forecast — about ten seconds end-to-end.

After the rules pipeline produces a candidate Buy/Hold/Sell, the reviewer agent (app/agents/conviction_reviewer.py) gets the same evidence and one permission: downgrade Buy/Sell to Hold when the evidence is thin. It can never upgrade. Cost-gated to fire only on directional candidates with sentiment consulted — the Hold path skips it entirely.

The system prompt lists durability bands (single missed practice → keep; ACL/retirement → keep; trade rumor with no named team → downgrade) and asymmetric defaults (Keep when in doubt, since the rules already gated). When it overrides, the reviewer's rationale replaces the formulaic one verbatim.

Every tick lands in a TimescaleDB hypertable — Postgres with a time-aware extension. Rows are partitioned automatically by 7-day chunks, indexed by (variant_id, ts DESC) so the catalog's "last 30 days" median runs as a tight index scan instead of a full table read.

A continuous aggregate rolls every variant up into a daily sold-and-ask bucket on a schedule, so the price chart on every detail panel renders from pre-computed data — no per-request aggregation latency. New ticks force a refresh of the affected chunk so the chart never lies about staleness.

A 1/1 sells once a lifetime. The standard 30-day median collapses, and most pricing tools either silently refuse or return a number with no methodology. Cardpulse runs an explicit borrowing graph with five named edges, each with its own statistical rationale.

Edges A–E pull comps from progressively further distance: cross-year siblings of the same insert (compact 4-year tricube kernel), same-year cross-player comps adjusted by an empirical-Bayes-shrunk player-tier multiplier, the player's broader 1/1 ceiling capped at 20% weight share, the pop class anchor multiplied by a fitted scarcity log-ratio, and a cross-product sanity floor capped at 15%.

Per-comp weight = tricube(distance) × tricube(recency) × (1 − suspicion)². Per-edge headline = weighted median of log-prices. Per-edge variance = 400-sample bootstrap. Cross-edge pool = inverse-variance with hard share caps and excess redistribution. The pooled posterior shrinks toward a product-line prior via Normal-Normal conjugacy, then the CI inflates by an explicit borrowing distance (1 − max admitted edge share, λ = 0.6) so a borrowed estimate is structurally wider than a direct-comp one.

Pathology checks refuse single-comp dominance (Logoman/Shield outlier z-test against leave-one-out cluster stats), high mean suspicion, and stale edges where the dominant 80% of weight is pre-540d. Cross-grade comps survive only when a learned grade-bridge multiplier admits at the product line — refused otherwise. Borrowed estimates cap at WARM confidence and demote to COLD/DARK as borrowing distance widens.

A separate calibration cohort tracks bucketed PIT histograms (A-dominant / B-dominant / C-dominant / mixed). Empirical 80% coverage on A+B-dominant ≥ 75% AND log-MAE < 0.35 is the gate that earns borrowed estimates promotion into the predictor's anchor blend.

Most pricing tools stop at a single number. Cardpulse keeps going. Every variant gets a 30-day and 90-day forecast as a true 80% prediction band — not a fudged ±8%. The buy / hold / sell conviction is a function of the band's overlap with current price, plus the agent's net sentiment.

The number itself is a sum of layered drivers. A regression baseline against the recent comp ladder. A peer-borrow from comparable variants when the card itself is thin. A sentiment drift from the agent's signal aggregate. A grade-tier multiplier (top tiers amplify sentiment harder than mids). A rookie premium when the card is an RC (rookies trade on narrative arc). A rarity premium for short-print, low-pop, and 1-of-1 variants. A calendar bump when a postseason game, set release, or theatrical premiere sits inside the horizon.

Each layer carries its own contribution percentage and confidence score. Click into any forecast and walk back through them.

Adapters re-emit the same eBay / PriceCharting listing every hourly refresh, often at the same price. Without a guard, price_ticks bloats by ~24× per live listing per day — same data, no new information. The persister checks for an identical-price tick in a 90-minute window before insert and skips when matched. Real price changes still create a new row.

The same hygiene runs at every agent boundary: news_articles dedup on URL or content_hash; sentiment_signals via partial-UNIQUE on dedup_key; forecast_eval has a NOT EXISTS guard with a 23h window; the predictor cron only re-runs on variants whose ticks moved in the last 90 minutes — so same-price reposts no longer trigger re-prediction either.

The hourly cron triggers the refresh. Every adapter pass yields ticks that get persisted, and every persisted tick fires a broadcast onto a Redis pub/sub channel. The web app keeps a WebSocket open to that channel. The tick lands in your browser within hundreds of milliseconds of the database write — no polling, no manual refresh.

The same signal triggers the alert engine: any active rule whose variant just got a new tick gets re-evaluated. If the rule trips, an event lands in your inbox with a 24-hour debounce so a flapping price doesn't spam you.