jaxperro
c948f0f413
Chain-sweep backfill (784k conds ≈ ~1k calls, not 8 days of per-cond queries) → shadow audit → consumer flips shallowest-to-deepest → trust.py simplification last. Ground rules codify the 2026-07-08 alignment-audit discipline: overlay never rewrite, one flip per day, never coupled with follow-set changes, calibration continuity notes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
130 lines
6.7 KiB
Markdown
130 lines
6.7 KiB
Markdown
# Etherscan migration — true resolution time through the whole stack
|
||
|
||
**Goal:** every consumer of `res_t` reads the exact on-chain
|
||
`ConditionResolution` timestamp instead of endDate metadata. Measured rot
|
||
that motivates this (2026-07-08): the Jul-7 Brewers market carried
|
||
`end_date_iso` **2026-05-05** and cache res_t **29h before** true resolution
|
||
(04:51 next morning); a Jul-5 tennis match carried res_t **Jul-14** (future);
|
||
an entire class of in-play sells could never pass a timestamp test.
|
||
`payouts.resolution_time(cond)` (wired, verified) is the source; this plan
|
||
is the *consumption* migration. **No code until the plan is agreed.**
|
||
|
||
## Ground rules (scar tissue from the 2026-07-08 alignment audit)
|
||
|
||
1. **Overlay, never rewrite.** Chain timestamps live in `resolution_times`
|
||
(own table). `bets.res_t` is never mutated — provenance stays intact,
|
||
every flip is a one-line revert per consumer.
|
||
2. **One consumer per phase**, each with a before/after diff reviewed
|
||
*before* publishing. Never two flips in one day.
|
||
3. **Never couple a flip with a follow-set change** — attribution dies.
|
||
4. **Calibration continuity:** any phase that changes model output adds a
|
||
note (and a `rules_version` marker) to `live/history/calibration.csv`
|
||
that day, so the live-vs-model series stays interpretable.
|
||
5. Docs land with the phase, not after: FINDINGS gets the measured deltas,
|
||
README gotchas get amendments, HANDOFF tracks the phase counter.
|
||
|
||
## Scale facts (measured 2026-07-08)
|
||
|
||
- 974,155 distinct conds in the bets cache; **783,911 resolved**;
|
||
141,396 in exits; 66,095 in `resolutions` (payout cache); 1 in
|
||
`resolution_times` (the Brewers proof).
|
||
- Per-cond Etherscan queries are a non-starter: 784k ÷ 5 req/s ≈ 43h of
|
||
calls against a 100k/day cap ≈ **8 days**.
|
||
- **The chain sweep is the enabler:** ONE topic0-only `getLogs` walk over
|
||
the CTF contract returns every ConditionResolution ever, 1,000 events per
|
||
call ≈ **~1,000 calls ≈ minutes**, resumable by block cursor. Nightly
|
||
increment afterward = 1–3 calls in `daily.sh`.
|
||
- `res_t` consumer weight (grep): trust.py 26 · portfolio.py 23 ·
|
||
cache.py 11 · conviction_scan.py 5 · validate_timing.py 4 · skill.py 4 ·
|
||
sync_floors.py 1. Flip order below runs shallowest → deepest.
|
||
|
||
## Phase 0 — full-chain backfill (enabler, no behavior change)
|
||
|
||
Build `live/resolution_sweep.py`: windowed topic0-only walker over the CTF
|
||
contract (chainid=137), resumable `last_block` cursor, throttled ≤5 req/s,
|
||
writes `(cond, res_ts)` into `resolution_times`. Wire the incremental call
|
||
into `daily.sh` (cache is single-writer — it runs inside the pipeline slot).
|
||
|
||
Exit criteria / audits before any consumer flips:
|
||
- **Coverage:** ≥99% of *resolved* conds that appear in any scoring path
|
||
(trusted rows for scanned wallets, exits, portfolio replay) have a chain
|
||
timestamp; sample the misses by hand.
|
||
- **negRisk verification:** confirm ConditionResolution fires for negRisk
|
||
conds (oracle = the adapter) on a handful of known negRisk markets.
|
||
- **Uniqueness:** assert one event per cond in the sweep (if re-reports
|
||
exist, keep the LAST and document).
|
||
- **The rot histogram:** distribution of `res_ts − metadata res_t` across
|
||
all matched conds → FINDINGS. This quantifies 18 months of silent error
|
||
and predicts which downstream metrics will move.
|
||
|
||
## Phase 1 — read-only shadow audit (no behavior change)
|
||
|
||
Recompute both ways, publish nothing:
|
||
1. **validate_timing shadow:** med_lead_h and the held/timing gates with
|
||
chain res_ts — does the 35-sharp membership change? Which wallets flip?
|
||
2. **portfolio shadow:** replay with (a) settle/capital-recycle at true
|
||
res_ts (today's game-day-midnight settles free capital EARLY — expect an
|
||
equity haircut from honest capital lock), (b) timestamp sold-test
|
||
primary with the price test retained as fallback. Record equity/record
|
||
deltas per wallet.
|
||
3. **trust shadow:** agreement rate between consensus res_t and chain
|
||
res_ts — the empirical case for how much of trust.py survives Phase 5.
|
||
|
||
Deliverable: one FINDINGS section with the three deltas. Flip order below
|
||
is confirmed or reshuffled *on this data*.
|
||
|
||
## Phase 2 — flip validate_timing (lowest stakes)
|
||
|
||
Lead metrics + timing gates read chain res_ts (fallback: old logic where a
|
||
cond has no event, e.g. genuinely unresolved). If sharp-list membership
|
||
changes, that's honest — document the entrants/leavers like a follow-set
|
||
change review. Sharps table republished; dashboards unchanged structurally.
|
||
|
||
## Phase 3 — flip portfolio.py (moves published numbers)
|
||
|
||
Settle times, sold-test primary, threshold windows (`res_t < START`) all
|
||
read chain res_ts. Then: full row-audit vs the live bot book (the 7/9-style
|
||
check — agreement must not regress), republish, calibration note (rule 4).
|
||
Expect and pre-announce the equity change; the honest number wins.
|
||
|
||
## Phase 4 — flip conviction_scan + skill + floors (selection layer)
|
||
|
||
Train/test splits and conviction windows on chain res_ts. Floors re-pin via
|
||
`sync_floors` and may shift a few % → coordinate a same-day
|
||
`deploy_bot.sh` so bot and backtest keep gating identically (rule 3 still
|
||
holds: no wallet adds/drops that day).
|
||
|
||
## Phase 5 — simplify trust.py (LAST, biggest payoff)
|
||
|
||
With chain res_ts + chain payouts covering all scoring conds, the consensus
|
||
res_t machinery, the v2 self-certification split, and the 14-day pulled_at
|
||
fallback reduce to: *"resolved on-chain? what payout? when?"* Keep consensus
|
||
only as a fallback for conds absent from chain (Phase 0 says ~none).
|
||
Regression bar: the sharp list before/after must be identical or every
|
||
difference individually explained. This deletes the most complex code in
|
||
the repo — do it after the chain data has soaked a week+.
|
||
|
||
## Parallel track (independent, any time)
|
||
|
||
Port insider.py's funding-cluster tracer from Alchemy `getLogs` (now capped
|
||
at 10 blocks on free tier) to Etherscan V2 logs. No interaction with the
|
||
res_t work.
|
||
|
||
## Risks
|
||
|
||
| risk | mitigation |
|
||
|------|------------|
|
||
| Etherscan limits/downtime | cache-forever table; resumable cursor; env `ETHERSCAN_KEY` override; sweep re-runnable idempotently |
|
||
| negRisk event semantics differ | Phase 0 verification gate before any flip |
|
||
| UMA disputes / duplicate events | uniqueness assert in Phase 0; disputes precede first report, immutable after |
|
||
| DuckDB single-writer | sweep runs in the daily pipeline slot only |
|
||
| numbers move and surprise | shadow audit first; pre-announced deltas; rules_version in calibration.csv |
|
||
| Etherscan V2 API drift | pin chainid=137 param form; smoke test in sweep preamble |
|
||
|
||
## Sequencing
|
||
|
||
Phase 0+1 fit one session (~2–3h, mostly sweep runtime + audit reads).
|
||
Each flip (2, 3, 4) is its own session with regen + publish + docs.
|
||
Phase 5 waits a week+ of soak. Total: ~4 working sessions spread over
|
||
~2 weeks, zero downtime, every step revertible.
|