Beacon

Tracks funding programs for Calgary, Alberta and Canada builders, and labels every claim with how it was verified.

Most grant lists present scraped text as fact. Beacon shows its work instead.

Every program carries a claim tier:

• VERIFIED — confirmed against a primary source (the program's own site or announcement) in the current check cycle.
• INFERRED — drawn from secondary reporting (press, guides). Treated as a lead, not a fact.
• GUESSED / UNKNOWN — marked explicitly when the evidence doesn't support more.

What it guarantees

• A fetch failure never deletes or flips a program. scripts/reverify.py keeps the old status, marks a tier concern and logs it. A 200 response with no status keyword keeps the current status.
• Every status change carries evidence. Changes are appended to data/ledger.jsonl with their evidence URL, and the dashboard shows a "last verified" stamp on every card.
• Discovery never self-promotes. New-program leads from scripts/discover.py are always tier UNKNOWN and are never written into data/programs.json automatically.
• Unverified facts render as UNKNOWN. Per-program evidence snapshots name a source, a check date and a tier for every fact.
• Every run leaves a receipt. scripts/agent.py writes data/agent.json with per-step results. A failed step is recorded, not fatal, and the run reports RESULT: FAIL naming the step.
• Private data stays out. scripts/check.py validates schema, tiers and pipeline privacy. Real application entries live in the gitignored data/pipeline.json.

Quickstart

git clone https://github.com/FetchDoge/beacon && cd beacon
python3 scripts/match.py      # run funding & eligibility search

Standard-library Python only. No dependencies.

Example search & output

Running python3 scripts/match.py scores builder criteria (data/profile.yaml) against all funding programs in data/programs.json:

Beacon match — 13 programs · 2026-10-07
Verdicts: blocked=5, eligible-now=3, not-for-him=5
Shortlist:
  + ai-rin-2026: Ecosystem infrastructure — founder-facing via RINs; no incorporation bar.
  + elevateip-ab: Open and does not require incorporation for this profile.
  + ai-compute-fund: Open and does not require incorporation for this profile.
Unlock roadmap:
  → incorporate unlocks 5: ai-rnd-associates, irap-ai-assist, raii-prairies, ieg-sred, mitacs-accelerate
  → first-rnd-dollar unlocks 1: ieg-sred
Wrote data/shortlist.json
RESULT: PASS

Interpreting the results

• Verdicts: Categorizes programs into eligible-now (apply today), blocked (prerequisites pending), and not-for-him (out of scope).
• Shortlist: Immediately actionable programs that meet all criteria without blocking hurdles.
• Unlock roadmap: Highlights concrete tactical milestones (incorporation or first R&D dollar) and the programs each unlock.

Running the pipeline

python3 scripts/check.py      # validate the dataset (schema, tiers, pipeline privacy)
python3 scripts/reverify.py   # re-check sources, update ledger on real changes
python3 scripts/match.py      # score eligibility, write data/shortlist.json
python3 scripts/evidence.py   # build per-program evidence snapshots -> data/evidence.json
python3 scripts/embed.py      # rebuild docs/index.html + docs/variant-b.html from data

Open docs/index.html in a browser. That's the whole deployment.

How it fails

Condition	Behaviour
Source unreachable or non-200	Status kept, tier concern logged, run continues
Page reachable but no status keyword	Status kept ("keep current")
A step in the agent loop fails	Recorded in data/agent.json. Remaining steps still run, and the result is RESULT: FAIL with the step named
New program spotted on a listing page	Queued in data/discovery.json as an UNKNOWN lead, with an alert to verify against a primary source

Evidence

From committed data, 2026-10-07:

• data/programs.json: 13 programs. 10 open, 2 paused, 1 closed. 12 INFERRED, 1 VERIFIED (Alberta Innovates Regional Innovation Networks renewal). By region: 6 Alberta, 6 Canada, 1 Calgary.
• data/shortlist.json: 3 eligible now, 5 blocked (unlock on incorporation), 5 not a fit.
• data/agent.json: the latest run, at 2026-10-07 19:19 UTC (13:19 MDT), finished with ok: true, all 8 steps rc 0. It checked 10 sources, found 0 flips and has 58 discovery candidates queued.
• python3 scripts/check.py passes locally (2026-10-07).

Most records are INFERRED from secondary sources. Verify before you apply; Beacon tells you exactly which claims need it.

What it does

• Radar: 13 programs across Calgary / Alberta / Canada, filterable by region, status, and type.
• Shortlist — an eligibility engine (scripts/match.py + data/profile.yaml) scores programs against a builder profile and produces an unlock roadmap: what's eligible now, what's gated behind incorporation, what's gated behind first R&D spend.
• Ledger — a provenance-bearing history of every status change, with evidence links.

The agent loop (Stage 2, 2026-09-30)

One command runs the whole radar cycle, in dependency order:

python3 scripts/agent.py              # full loop (see below)
python3 scripts/agent.py --dry-run    # reverify/discover/alerts print, don't write
python3 scripts/agent.py --limit 2    # smoke-test the loop on 2 programs

Step	Script	What it does
1	reverify.py --respect-cadence	Re-checks only sources due this run (check_cadence_days: 7 open / 14 paused / 30 closed); flips and concerns → ledger
2	match.py	Automatic re-scoring of every program against data/profile.yaml → shortlist.json
3	deadlines.py	Recomputes urgency bands → deadlines.json
4	evidence.py	Rebuilds per-program evidence snapshots
5	discover.py	Scans data/discovery_sources.json listing pages for new-program leads → data/discovery.json (tier UNKNOWN, never auto-promoted)
6	alerts.py	Deadline crossings (≤30d, overdue), today's status flips, discovery digest → data/alerts.json + provenance-bearing ledger events
7	embed.py	Rebuilds docs/index.html + docs/variant-b.html
8	check.py	Validates the dataset

Every run writes data/agent.json — the run receipt (timestamps, per-step results, counts). A failed step is recorded, not fatal; the loop finishes what it can and reports RESULT: FAIL naming the step.

Scheduled: .github/workflows/agent.yml runs scripts/agent.py daily at 13:17 UTC and commits data/ + docs/ back to main — the radar updates itself.

Claim-tier discipline holds throughout: discovery candidates are UNKNOWN leads; fetch failures never flip or delete; alerts are derived, not primary claims.

Waves

• Wave 1 (2026-09-30): deadline-urgency engine, stacking notes, client-side search, two page variants (cards + dense table).
• Wave 2 (2026-09-30): application pipeline tracker — stages prospecting → preparing → applied → in-review → awarded/declined. Real entries live in the gitignored data/pipeline.json; the repo ships data/pipeline.template.json with clearly-labeled SAMPLE data. The page says which one it's showing.
• Wave 3 (2026-09-30): per-program evidence snapshots (data/evidence.json, generated from programs.json + ledger.jsonl) — every fact carries a named source, check date, and claim tier; funder history comes only from recorded ledger events; anything unverified renders as UNKNOWN, never as a plausible fact. The data/ directory is documented as a static JSON read API in data/README.md.

Status

Live and self-updating: .github/workflows/agent.yml runs the loop daily at 13:17 UTC and commits data/ and docs/ back to main.

Built by Marcus Richards, directing AI coding agents.

License

MIT.