A specialist agent per market domain. Watches contracts, reads adjacent discussion, pings you when reality and the price drift apart. Never trades.
prediction-edge
telegram · alert · 06:42
🎬guild-watcherPGA + DGA + SAG-ensemble → Anora
📝critic-watcherNYFCC + NSFC → Anora; LAFCA split
📣buzz-watcherlate-cycle momentum shifting to Anora
Market: "Anora wins Best Picture" → 27%. The Brutalist still leading at 38%.
Three independent feeds converging on Anora. The PGA + DGA + SAG-ensemble trio to one film has called BP 9 of 10 times since 2010. Your January pick was Anora before the guilds confirmed.
Worth a look?
not a trade. that's your job.
The situation
You watch prediction markets. You have informed views on a few specific topics (award shows, election outcomes, sports championships) but you can't watch every signal yourself. Prediction Edge runs a specialist agent per signal stream (guild votes, critic awards, polling, recruiting, whatever the market depends on), plus a reviewer that synthesizes across them and asks the questions you'd ask if you had time. None of them place trades. That's your job.
How it behaves
Three things, on the markets you care about.
Watches the signals you would — award shows: guild votes, critic awards, industry buzz. elections: polling and ground-game reporting. sports: recruiting and beat coverage
Pings you when signals agree but the price is lagging — the kind of mispricing you want to know about in minutes, not in the next day's recap
Stays read-only — every trade is yours by construction, so the system can flag but never act
A watcher sits on each signal stream. The reviewer only pings you when they independently line up. That keeps the alerts grounded instead of vibe-driven.
How you'd build it
Picking what to watch
prediction-edge
I bet on award shows — Best Picture mainly, sometimes Best Actor and Director. I follow the trade press and feel I have informed views, but I can't watch every signal in real time. What would the system actually watch?
all-agents · design
For Best Picture, the high-signal feeds are guild votes (PGA, DGA, SAG-ensemble — these are the predictive ones), critic associations (NYFCC and NSFC matter because they're early; LAFCA splits often), and industry buzz (Variety, Deadline, THR coverage of the campaign).
I'd set up a watcher per stream — three watchers — plus a reviewer whose job is the synthesis: when guild + critic + buzz align toward a film and the market hasn't repriced, that's the alert. For Best Actor and Director we'd reuse the same shape with the actor- and director-specific outlets.
Want me to watch more than one venue when they list the same race?
Yes, cross-reference where they overlap. Skip Best Director — too noisy, too contested. Add the Spirit Awards to the critics watcher; they've been more predictive lately. And if there's strong indie consensus the guilds aren't seeing, I want to know separately — that's a different signal pattern from guild-converged.
all-agents · design
Wiring it in:
· Best Picture and Best Actor — three watchers (guilds, critics, buzz) with Spirit Awards feeding the critics watcher
· Director skipped until you say otherwise
· every venue you watch, read-only, with cross-reference when more than one lists the same race
· New rule for the reviewer: Spirit + critic alignment toward an indie that the guilds haven't seen → separate alert class, distinguishable from guild-converged alerts at a glance
The last one is the tricky bit. Want me to draft sample alert language for the indie case so you can see how it'll read on Telegram?
Your edge as a bettor is the input. You bring domain knowledge — which markets you follow, which signals matter, which sources you trust. Cast turns that into a structure: which parts watch what, how they cross-check, when they ping you. The substance is yours; Cast supplies the shape.
How an alert decides to ping you
prediction-edge
On these prediction-market alerts — I don't want to be paged for every flutter. When would something actually fire vs. stay quiet?
all-agents · design
Convergence + price gap is the bar. Two examples to make it concrete:
· All three watchers surface Anora over The Brutalist (PGA + DGA + SAG-ensemble all align). The market has The Brutalist at 38%, Anora at 27% — that's an 11-point gap with three converged signals. Pings you on Telegram.
· One critic association (LAFCA) goes for The Brutalist; the others haven't moved. No convergence. The reviewer notes it in the morning brief but does not page.
The gap threshold and which signal combinations count are tunable. I'd write a starting point into the reviewer's instructions — say 10-point spread, two-of-three convergence — and we'd refine those together over the first few weeks based on what fires.
Start at 8 points spread, not 10 — I'm fine with more noise to find more edges, I'll tighten if it gets annoying. Two-of-three convergence is right. Weight NYFCC and NSFC heavier than LAFCA — LAFCA picks weird every few years, shouldn't move the needle.
all-agents · design
Set:
· 8-point gap minimum to alert
· two-of-three convergence required
· LAFCA noted in the morning brief but doesn't count toward critic-watcher convergence
· NYFCC + NSFC aligning counts as the critic-watcher's vote, even if LAFCA dissents
All of these live in the reviewer's instructions — I can tune any of them in a future design pass without redoing the watchers, whenever you want them dialed.
What turns this from a noisy feed into a tool is the bar for action. The design conversation is where you set it: what convergence means, what gap matters, what stays in the morning brief instead of paging you. Tunable from there as you live with it.
Adding a new market
prediction-edge
I want to add NCAA basketball next March — completely different signals from awards, but I follow it the same way. How much rework?
all-agents · design
Almost none of the structure changes; the contents do. Three new watchers, scoped to the streams that matter for college hoops:
· recruiting and roster construction (247Sports, Rivals, On3)
· advanced team metrics (KenPom, Bart Torvik, Haslametrics)
· beat coverage and locker-room reporting
Reviewer same role, different prompt about what convergence looks like for sports — usually it's metrics + recruiting both flagging a team that the futures market hasn't repriced. Coordinator unchanged.
I'd run NCAA alongside the awards setup — both stay independent, both alert you on their own clock. The system gets bigger but doesn't get harder to use.
Only follow KenPom top-25 — I don't have an edge below that. Run it Nov 15 through Selection Sunday, paused otherwise. Awards setup keeps running year-round. NCAA convergence rule should be different from awards: recruiting + advanced metrics aligning is the strong signal, beat coverage is supporting, not gating.
all-agents · design
Configured:
· NCAA active Nov 15 through Selection Sunday; paused outside that window — no signal collection, no reviewer cycles
· top-25 KenPom only; teams falling out drop off automatically, climbers added
· awards setup unchanged, year-round, alongside
· NCAA convergence rule: recruiting + metrics aligning on a team the futures market hasn't priced → alert; beat coverage as supporting evidence in the body, not gating
Ready to ship the first NCAA version whenever you are. Dry run when November hits, or live on day one?
Once one domain is working, the rest is mostly supplying signals. A new market isn't a new system — it's the same shape with different specialists and different prompts about what convergence looks like. The architecture you set up the first time is the one you reuse.
What stays separate
None of the parts can place a bet, by construction. The market connections are wired up for reading only (checking prices, looking at order books), and the betting actions simply aren't there. The model has nothing to call. Read-only isn't a rule the assistant obeys. It's the absence of a button it could push.
24/7 watch with threshold-triggered pings to your Telegram
Blueprint highlight
Three watchers on different signal streams + a reviewer cross-checking + a coordinator alerting you, read-only across the whole thing
Build it yourself
Install Cast, then describe this in the admin chat. The Design agent writes the agents, conversations, and ACL for you. The extensions you'd need are listed above.