Skip to content

Add Predict reference tick admission#1086

Merged
0xaslan merged 10 commits into
mainfrom
at/specific-strike-unified-grid
Jun 25, 2026
Merged

Add Predict reference tick admission#1086
0xaslan merged 10 commits into
mainfrom
at/specific-strike-unified-grid

Conversation

@0xaslan

@0xaslan 0xaslan commented Jun 25, 2026

Copy link
Copy Markdown
Collaborator

Summary

  • Add Predict reference-tick admission on the unified tick grid, including exact-history Pyth setup and exactly-once event emission.
  • Widen packed order tick fields to 30 bits and split cadence tick granularity from admission tick granularity.
  • Update Predict unit coverage and simulation mirrors so existing simulations continue to run on the current admission grid.

Key decisions

  • The reference tick is derived from the previous-window exact Propbook Pyth observation and may bypass the coarser admission grid while still using the fine tick grid.
  • Simulations do not add a reference-tick scenario mode; they keep existing behavior by setting admission tick size equal to tick size.
  • Localnet setup passes the new cadence admission tick argument, and Python/TS mirrors use the 30-bit tick/order-id layout.
  • Unrelated local predict-audit skill changes were left out of the commits.

Test plan

  • sui move build --path packages/predict --warnings-are-errors
  • sui move test --path packages/predict --gas-limit 100000000000
  • bash -n run.sh
  • npx tsc --noEmit
  • python3 -m py_compile python_replay.py python_indexes/liquidation_book.py charts/chart_market_overview.py data/generate_scenario.py tools/analyze_liquidation_priority_encodings.py
  • python3 data/generate_scenario.py --mode normal --out /tmp/predict_normal_scenario.csv
  • python3 python_replay.py --scenario /tmp/predict_normal_scenario.csv --out /tmp/predict_normal_replay.json --max-rows 1000
  • python3 data/generate_scenario.py --mode long --out /tmp/predict_long_scenario.csv
  • python3 python_replay.py --scenario /tmp/predict_long_scenario.csv --out /tmp/predict_long_replay.json --derived-out /tmp/predict_long_derived.json --long-run
  • bash run.sh --sim_max_rows=1 --skip-analysis

@0xaslan 0xaslan changed the title Add Predict specific tick admission Add Predict reference tick admission Jun 25, 2026
0xaslan and others added 3 commits June 25, 2026 11:23
… verify, per-lane retirement

orchestrator.workflow.js:
- Sequence trim — per-lane retirement: each lens carries its own consecutive-dry
  counter and drops out once it converges, so later rounds re-run only the lenses
  still surfacing issues instead of the whole fleet (dryRounds applies per-lane;
  maxRounds:1 = single pass). Index by the dispatched lane key, not the
  agent-returned lane, so identity is stable.
- Cross-model verify: route the refute + repro verifiers to the codex
  (codex:codex-rescue) agent and keep the settled-decision check on Claude, so
  every High/Critical finding clears two different models (finder-bias check).
  Medium gets one codex verifier; verify prompts bounded + JSON-only. Degrades to
  the Claude verdict(s) if codex is unavailable.
- Depth preset (depth: low|standard|max), orthogonal to breadth — 'low' still runs
  all lenses, just maxRounds:1. Explicit caps override the preset; config log echoes it.

SKILL.md:
- Document the above; the launch questionnaire now asks two independent axes
  (breadth + depth) with a depth-tier table; bounded-by-construction agent-count
  framing throughout.

Validated: the codex agentType+schema workflow hop + a refutation counter-spike
(codex correctly refuted a known-false claim and caught a specific/reference
symbol-name error the Claude lenses shared), plus a mocked end-to-end dry-run
covering the loop, retirement, depth precedence, and verify routing. Read-only on
packages/*; the rule-sweep / ownership-walk siblings are not yet ported.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@0xaslan 0xaslan merged commit 542280e into main Jun 25, 2026
7 checks passed
@0xaslan 0xaslan deleted the at/specific-strike-unified-grid branch June 25, 2026 17:23
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants