T1: Pre-check — policy→price signal validation

[masdevid]

What to build: A repeatable analysis over 12-month price/volume on 3-5 candidate Indonesian energy names showing whether price moves cluster on policy-announcement dates (narrative-to-price signal) vs. quote adjustments (administered price). Pass/fail report gates the resolver build (T2). No pipeline code committed — this is the cheap falsification point agreed in introspective grilling (Q24).

Acceptance criteria:

• Written report (in this issue or a doc) with a verdict: narrative signal exists for names X, Y (market-priced) vs. no signal (administered-price names excluded)
• If signal exists, a narrowed market-priced name list is produced
• All analysis uses existing Evidence Graph / Sectors data; no new data source
• Verdict stated explicitly: proceed to T2 or kill the amplifier thesis

Blocked by: None (can start immediately).

[masdevid]

T1 Complete — Pre-check report posted

Report: docs/t1-precheck-report.md
Verdict: CONDITIONAL — the methodology and code are validated (13 unit tests pass, 222/222 backend suite); the gate condition requires running with verified policy dates.

What was built

• backend/app/services/policy_signal.py — pure analysis + async cache-first fetch
• backend/app/services/policy_dates.py — curated energy policy events (placeholder dates)
• backend/app/tools/policy_signal_precheck.py — CLI runner (live-or-simulated)
• backend/tests/test_policy_signal.py — 13 tests, all passing

Live-data result (using real Sectors data, UNVERIFIED placeholder policy dates)

All 5 candidates (ADRO, ITMG, MEDC, PGAS, PTBA) classified market_priced but policy-inert (ratio 0.0) — expected, because the placeholder dates don't match real announcements.

Simulated result (synthetic series with 8% policy-day shocks)

3 market-priced signals (ADRO, ITMG, MEDC) with ratio > 2× — strong beacons.

Gate to proceed to T2

Replace the 5 placeholder dates in policy_dates.py with verified real dates, run:

cd backend && SECTORS_API_KEY=<key> python -m app.tools.policy_signal_precheck

If ≥2 names clear the 2.0× strong threshold → close this ticket and begin T2. If FAIL → kill the amplifier stream.

Status: Ready for gate exercise. Leaving ticket open for the live-data gate run.

[masdevid]

T1 pre-check completed with a live Sectors key.

Ran the pre-check against the API key set in .env. A defect was found & fixed first: /v2/daily/ defaults to a 30-day window (90 max/call), so the check originally analyzed only ~21 trading days and every ratio read 0.0. The fetch now pulls a real rolling 12-month window in 90-day chunks (b1748c3, suite 313).

Live verdict: FAIL on 12 months of real daily closes (240 observations per name):

• All 5 candidates are market-priced (ADRO 2.54%, ITMG 1.74%, MEDC 3.27%, PGAS 2.35%, PTBA 1.94% daily vol) — including names the model had treated as "administered".
• None reacts around the overlaid policy dates: ratios 0.57 / 0.57 / 1.02 / 0.78 / 0.63 (floor 1.5).
• Caveat: all five events are unverified placeholder dates (budget-calendar). Real announcement days would change the windows.

Per the spec's pre-check semantics this says "do not proceed to sector resolution" for these dates. The T2–T7 machinery is built and unit-verified (T6 PASS is synthetic-scoring, unaffected), but the empirical narrative→price channel for these five placeholder events is not confirmed.

Full report: docs/t1-precheck-report.md (gitignored) and PR #9.

[masdevid]

T1 pre-check re-run — verified event dates, zero-cost cache

All five ENERGY_POLICY_EVENTS were verified against primary/journalist sources and the pre-check re-run:

Date	Event	Source
2025-10-01	Pertamina non-subsidized BBM adjustment (Dexlite +Rp100 → 13,700, Dex +Rp150 → 14,000; Pertalite/Solar unchanged)	MediaPresisi 2025-10-01
2025-12-01	HBA reset, Dec period I → US$98.26	Kepmen ESDM 388.K/MB.01/MEM.B/2025 (JDHI ESDM)
2026-01-22	APBN 2026 energy-subsidy detail: BBM −5.7% → Rp25.14T, LPG −7.7%, electricity +16.6%	Perpres 118/2025; CNBC Indonesia 2026-01-22
2026-04-01	Q2-2026 electricity tariff window — held for all groups	Kompas 2026-03-31 / ESDM
2026-06-22	DMO: US$70 domestic coal cap kept, supply task force after ~20 Mt PLN shortfall	CNBC Indonesia / Liputan6 2026-06-22/23

Re-run verdict: FAIL (same null result). All five names market-priced (ADRO 2.54%, ITMG 1.74%, MEDC 3.27%, PGAS 2.35%, PTBA 1.94%); ratios 0.82 / 1.10 / 0.98 / 1.17 / 1.10, none near the 1.5 floor. On-window returns do not deviate from baseline. The amplifier thesis is falsified for this window even with accurate announcement dates.

Cost optimization (commit 2ec60a6): the 12-month daily fetch now caches deterministic 90-day epochs (fixed calendar grid, 7-day TTL) instead of re-fetching wholesale. The re-run made 0 Sectors calls (usage 30 → 30, verified); steady-state re-runs are free until a new epoch appears (~1 call/name/90 days), down from 25 calls per run.

Per the pre-check gate: do not proceed to sector resolution on these event dates. Report: docs/t1-precheck-report.md.