Skip to content

Document the champion's context length, and seeding a live miner's store - #26

Merged
Thykof merged 2 commits into
mainfrom
docs/context-days
Nov 9, 2026
Merged

Thykof merged 2 commits into
mainfrom
docs/context-days

Conversation

@Thykof

@Thykof Thykof commented Nov 6, 2026

Copy link
Copy Markdown
Contributor

Stacked on #25. Engine side: synthdataco/synth-lib#91.

Why

From campaign-7 a champion chooses its context length. Its modeling.py may set CONTEXT_DAYS, a
plain integer from 3 to 30; without it the champion gets 7 days, as every archived champion does.
The agent's backtests, the interface check, the verdict and the live miner read the same value. The
engine side lands in a companion synth-lib PR and reaches this repo with the next pin. This PR
documents it for operators and for anyone deploying a champion.

The deploy guide matters most. A deployed miner needs CONTEXT_DAYS + 1 days in its store, and
Hyperliquid serves only its last 5,000 minutes, about 3.5 days. Unseeded, a 30-day champion would
serve its 9 Hyperliquid-routed assets from about 3.5 real days, gaining one a day, for four weeks.

Stacked on build/antigravity-cli.

Merge order. This PR merges after the synth-lib stack, once a pin-bump commit to the merged
synth-lib lands on build/kimi-code-2. The pin it sits on today has no CONTEXT_DAYS code, so
these docs describe the engine that bump brings.

What

File Change
docs/deploy-a-champion.md A section on CONTEXT_DAYS; the price-store section rewritten (below)
docs/scoring.md The verdict's context is the champion's CONTEXT_DAYS, 7 when unset
docs/operate-a-campaign.md The verdict section's 7-day mentions become CONTEXT_DAYS (below); the loader's NaN head; the interface-check note
docs/reproduce.md The store must cover the window plus the champion's context
docs/install-and-first-run.md "the 7-day context" becomes "the context"; the smoke's summary line and interface-check note
CHANGELOG.md Under "Planned, not built", a Task group: the context length, landing with the next synth-lib pin

What operate-a-campaign.md now says about the verdict's prices:

  • The data table: the champion's input is CONTEXT_DAYS days per prompt, 7 when unset. The loader
    requires a partition for every day after the asset's first one; days before it are NaN.
  • Generation starts one horizon, up to a day, before the window, so the first contexts start on
    window start - CONTEXT_DAYS - 1 day. The store needs every day from there, for the largest
    CONTEXT_DAYS among the legs scored, and the audit runs with
    --require-from <window start - CONTEXT_DAYS - 1 day>.
  • The no-lookahead chain hands each prompt only the CONTEXT_DAYS days up to t.
  • The interface check's note starts with ok (24h + 1h), CONTEXT_DAYS N.

What deploy-a-champion.md now says about the store:

  • Each request reads CONTEXT_DAYS + 1 daily partitions: a D-day window touches D + 1, the oldest
    and today's both partial.
  • Binance-routed assets (BTC, ETH, SOL, XRP): the miner's own warm-up fetches every day. The
    old pre-warm loop with market_data --days 30 is gone.
  • Hyperliquid-routed assets (HYPE and the 8 com-equ-24h assets): with CONTEXT_DAYS above 3,
    seed the store with scripts/ingest_reservoir.py for the CONTEXT_DAYS + 1 days up to
    yesterday. 31 days covers any champion.
  • The seed and the miner both run from the champion's own folder: cd champions/<name>, the
    seed as uv run python ../../scripts/ingest_reservoir.py ..., then ./entrypoint.sh. Its
    market_data/ is then apart from the campaign store in the repo root, and uv still finds the
    repo's project.
  • Seed before the first start: the warm-up stores the venue's partly retained oldest day as
    settled, and the miner never fetches a settled day again.
  • Start each champion on an empty store, as a new champion folder is: a store reused from the
    campaign, another champion or an older engine keeps its old days.
  • The archive publishes with a lag and the venue covers the last three days, so seed on the day the
    miner starts.

Tests

Check Result
uv run black --check . 11 files unchanged
uv run pytest -q 42 passed

Docs only. The summary (CONTEXT_DAYS N): line and the interface-check note were checked against
the synth-lib PR's predict.py and nomination.py, and the window lead-in against its
run_verdict.generation_start. Re-read the engine claims against that PR before merging this one.

🤖 Generated with Claude Code

https://claude.ai/code/session_016kmUsyYSEMPGiwbSuvR9hh

@claude claude Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

Tip: disable this comment in your organization's Code Review settings.

@Thykof
Thykof added this pull request to stack #28 November 9, 2026 10:54
@Thykof
Thykof force-pushed the docs/context-days branch from 903b6e8 to 9fc749d Compare November 9, 2026 10:55
Base automatically changed from build/antigravity-cli to main November 9, 2026 10:56
Thykof and others added 2 commits November 9, 2026 11:56
…ive miner's store

From campaign-7 a champion's modeling.py may set CONTEXT_DAYS, an int from
3 to 30, 7 when absent. deploy-a-champion.md describes it and how to read
it, and its price-store section now says what a deployed miner needs:
CONTEXT_DAYS + 1 days, Binance-routed assets from the venue, and
Hyperliquid-routed ones beyond about 3.5 days seeded from the archive with
scripts/ingest_reservoir.py before the miner first starts, on an empty
store. scoring.md, the verdict section of operate-a-campaign.md,
reproduce.md and install-and-first-run.md replace the fixed 7 days with
CONTEXT_DAYS. The CHANGELOG lists the context length as landing with the
next synth-lib pin.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016kmUsyYSEMPGiwbSuvR9hh
…pion's own store

operate-a-campaign.md: the loader requires a partition for every day after
an asset's first one, and days before it are NaN. Generation starts one
horizon before the window, so the store needs every day from
CONTEXT_DAYS + 1 days before --window-start, and audit_store.py takes
--require-from <window start - CONTEXT_DAYS - 1 day>. predict.py prints
`summary (CONTEXT_DAYS N):`, and the interface check's note starts with
`ok (24h + 1h), CONTEXT_DAYS N`.

deploy-a-champion.md runs the seed and the miner from the champion's own
folder, so its market_data/ is apart from the campaign store.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016kmUsyYSEMPGiwbSuvR9hh
@Thykof
Thykof force-pushed the docs/context-days branch from 9fc749d to 5ecec2c Compare November 9, 2026 10:56
@Thykof
Thykof merged commit b1202ab into main Nov 9, 2026
4 checks passed
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.

1 participant