/map/ — the learning map (asked for by the owner)

Scanner audit, part 1 — each pattern tested against other hours (Backlog #1)

• Need: the scanner's raw win rates ("RSI overbought wins 70%") mean nothing without a baseline — in a falling market every bearish finding "wins" — and some patterns looked worse than chance. Traders see these findings on coin pages and in the bot with no warning.
• Built: scripts/scanaudit.py (CI step after scanlog): for every finding in scanlog.json, the 24h move in the signal's direction from the next full-hour open; baseline = all findings of a pattern shifted together by every k hours (circular, exact test over all shifts — keeps the cross-coin clustering); two-sided p around the shifts' median; verdict better / worse ("under review") / no_edge / too_few (n<30) at p<0.05. Writes data/scanaudit.json. Coin pages show the tested line instead of the raw win % when a verdict exists; the bot marks "under review" under each finding in Opportunities and lists every pattern's test in the Track section (fa/ckb/en). Nothing hidden.
• Verified: py_compile; unit tests on synthetic random walks — null findings significant 0/10 seeds, planted +4% move caught (bull "better" p=0.0012, its bear mirror "worse" p=0.0012), one move hand-checked (+21.82%); coin pages and bot sections rendered on a scratch copy with fake audit data; screenshots 1280/390, zero JS errors. Real figures come from the next CI run — not yet seen.
• Next: Backlog #1 — read the real audit, then the /opportunity/ table.

Public /autopilot/ log page (Backlog #1)

• Need: the loop changes the site every hour; readers (and the owner) had no way to see what changed and why on the site itself — accountability for the site's own changes, matching the "measured honesty" strategy.
• Built: scripts/build_autopilot.py parses this file's Log and renders /autopilot/ (fa/ckb/en frame, entries in English, bold/code only, everything escaped) with links to AUTOPILOT.md's git history and a search of autopilot: commits; footer link "Autopilot log" next to Methodology (chrome.py); builder added to update.yml and pages.yml; sitemap entry. No bot entry: it is a meta page with no natural home in the bot.
• Verified: py_compile; parser unit test on synthetic markdown (continuation lines, missing title, section end, empty input, HTML escaping); built on a scratch copy; screenshots fa/en at 1280/390, zero JS errors.
• Next: Backlog #1 — scanner audit.

One honest number everywhere (Backlog #1)

• Need: coin pages, the home/dashboard "post track record" box and the Entekhab Score still used the raw average 24h move after posts (no baseline) — in a drifting market that flatters the channel.
• Built: score.py channel_component() — the Score's channel part is 50 (neutral) unless the tested record (data/trackrecord.json, post hours vs matched random hours) has p < 0.05; then 50 + 8 × (post avg − random avg). build_pages.py tested_line() replaces the raw sentence on every coin page and the server-rendered dashboard box; the home page's client-side box (index.html, fa/ckb/en) now loads trackrecord.json and shows post vs random-hour BTC move, p and the verdict, linking to /track-record/. /methodology/ wording updated.
• Verified: unit tests on synthetic inputs (empty → 50; real p=0.28 → 50; p=0.01, +1.3 gap → 60.4; clamp at 0; p=0.05 boundary → 50); score + page builders run on a scratch copy (BTC/ETH channel = 50.0); screenshots home fa/en and BTC coin page at 1280/390, zero JS errors. Bot "track" section already quotes the tested figure (previous run).
• Next: Backlog #1 — public /autopilot/ log page.

/track-record/ — the channel's real record

• Need: every reader's first question — "does following this channel help?" — had only a misleading answer: the raw average 24h move after posts (+0.53%, 58% up over 524 posts) with no comparison. In the same window bitcoin's average day was −0.04% (48% up), so the number needed a real test.
• Built: scripts/trackrecord.py (CI, Binance 1h): entry at the open of the first full hour after each post, 24h hold; the same measures for every hour of the window; posts in one hour count once; matched-cluster randomisation (each post-day → a random day with the same number of random hours), 10,000 draws, seed 20031222; also the 4h move before the post (reaction) and the 24h range (eventfulness), per category, a histogram and a day-cluster bootstrap CI. scripts/build_trackrecord.py renders /track-record/ (fa/ckb/en) with a verdict chosen only by p < 0.05.
• Mirrored: menu (reality group), site assistant, first card on /reality/, Telegram bot section trackrec in the reality group; the bot's existing "track" section now quotes the tested number instead of the raw average.
• Verified: unit tests — null data never significant (0/10 seeds), planted +1.5% effect caught (p=0.0005), a "post after the move" pattern flagged as reaction and not as foresight, duplicates counted once, windows across data gaps skipped, one window checked by hand; full study runs in ~3.5 s; page screenshots desktop/mobile, fa/en, no JS errors. Real figures produced by CI (run 42372862451, 12:36 UTC) and the live page checked in a browser.
• Result (real data, 2026-08-04 → 2026-11-06): 508 posts in 259 distinct hours over 88 days. Bitcoin 24h after posts +0.43% vs +0.20% at matched random hours (95% of random results between −0.19% and +0.64%), p = 0.29 — not significant. Share of gaining windows 56.4% vs 52.9% (p = 0.47). No sign of posting after moves (4h before: −0.02% vs +0.04%, p = 0.32); volatility after posts like any hour (3.07% vs 2.84%, p = 0.22). No category significant. Note: a quick estimate earlier the same day (posts +0.53% vs a daily-close baseline of −0.04%) overstated the gap — it used a shorter window and daily closes. This result supersedes it.
• Next: Backlog #1 — use these figures everywhere the raw average is still shown.

/map/ — the learning map (asked for by the owner)

Scanner audit, part 1 — each pattern tested against other hours (Backlog #1)

• Need: the scanner's raw win rates ("RSI overbought wins 70%") mean nothing without a baseline — in a falling market every bearish finding "wins" — and some patterns looked worse than chance. Traders see these findings on coin pages and in the bot with no warning.
• Built: scripts/scanaudit.py (CI step after scanlog): for every finding in scanlog.json, the 24h move in the signal's direction from the next full-hour open; baseline = all findings of a pattern shifted together by every k hours (circular, exact test over all shifts — keeps the cross-coin clustering); two-sided p around the shifts' median; verdict better / worse ("under review") / no_edge / too_few (n<30) at p<0.05. Writes data/scanaudit.json. Coin pages show the tested line instead of the raw win % when a verdict exists; the bot marks "under review" under each finding in Opportunities and lists every pattern's test in the Track section (fa/ckb/en). Nothing hidden.
• Verified: py_compile; unit tests on synthetic random walks — null findings significant 0/10 seeds, planted +4% move caught (bull "better" p=0.0012, its bear mirror "worse" p=0.0012), one move hand-checked (+21.82%); coin pages and bot sections rendered on a scratch copy with fake audit data; screenshots 1280/390, zero JS errors. Real figures come from the next CI run — not yet seen.
• Next: Backlog #1 — read the real audit, then the /opportunity/ table.

Public /autopilot/ log page (Backlog #1)

• Need: the loop changes the site every hour; readers (and the owner) had no way to see what changed and why on the site itself — accountability for the site's own changes, matching the "measured honesty" strategy.
• Built: scripts/build_autopilot.py parses this file's Log and renders /autopilot/ (fa/ckb/en frame, entries in English, bold/code only, everything escaped) with links to AUTOPILOT.md's git history and a search of autopilot: commits; footer link "Autopilot log" next to Methodology (chrome.py); builder added to update.yml and pages.yml; sitemap entry. No bot entry: it is a meta page with no natural home in the bot.
• Verified: py_compile; parser unit test on synthetic markdown (continuation lines, missing title, section end, empty input, HTML escaping); built on a scratch copy; screenshots fa/en at 1280/390, zero JS errors.
• Next: Backlog #1 — scanner audit.

One honest number everywhere (Backlog #1)

• Need: coin pages, the home/dashboard "post track record" box and the Entekhab Score still used the raw average 24h move after posts (no baseline) — in a drifting market that flatters the channel.
• Built: score.py channel_component() — the Score's channel part is 50 (neutral) unless the tested record (data/trackrecord.json, post hours vs matched random hours) has p < 0.05; then 50 + 8 × (post avg − random avg). build_pages.py tested_line() replaces the raw sentence on every coin page and the server-rendered dashboard box; the home page's client-side box (index.html, fa/ckb/en) now loads trackrecord.json and shows post vs random-hour BTC move, p and the verdict, linking to /track-record/. /methodology/ wording updated.
• Verified: unit tests on synthetic inputs (empty → 50; real p=0.28 → 50; p=0.01, +1.3 gap → 60.4; clamp at 0; p=0.05 boundary → 50); score + page builders run on a scratch copy (BTC/ETH channel = 50.0); screenshots home fa/en and BTC coin page at 1280/390, zero JS errors. Bot "track" section already quotes the tested figure (previous run).
• Next: Backlog #1 — public /autopilot/ log page.

/track-record/ — the channel's real record

• Need: every reader's first question — "does following this channel help?" — had only a misleading answer: the raw average 24h move after posts (+0.53%, 58% up over 524 posts) with no comparison. In the same window bitcoin's average day was −0.04% (48% up), so the number needed a real test.
• Built: scripts/trackrecord.py (CI, Binance 1h): entry at the open of the first full hour after each post, 24h hold; the same measures for every hour of the window; posts in one hour count once; matched-cluster randomisation (each post-day → a random day with the same number of random hours), 10,000 draws, seed 20031222; also the 4h move before the post (reaction) and the 24h range (eventfulness), per category, a histogram and a day-cluster bootstrap CI. scripts/build_trackrecord.py renders /track-record/ (fa/ckb/en) with a verdict chosen only by p < 0.05.
• Mirrored: menu (reality group), site assistant, first card on /reality/, Telegram bot section trackrec in the reality group; the bot's existing "track" section now quotes the tested number instead of the raw average.
• Verified: unit tests — null data never significant (0/10 seeds), planted +1.5% effect caught (p=0.0005), a "post after the move" pattern flagged as reaction and not as foresight, duplicates counted once, windows across data gaps skipped, one window checked by hand; full study runs in ~3.5 s; page screenshots desktop/mobile, fa/en, no JS errors. Real figures produced by CI (run 42372862451, 12:36 UTC) and the live page checked in a browser.
• Result (real data, 2026-08-04 → 2026-11-06): 508 posts in 259 distinct hours over 88 days. Bitcoin 24h after posts +0.43% vs +0.20% at matched random hours (95% of random results between −0.19% and +0.64%), p = 0.29 — not significant. Share of gaining windows 56.4% vs 52.9% (p = 0.47). No sign of posting after moves (4h before: −0.02% vs +0.04%, p = 0.32); volatility after posts like any hour (3.07% vs 2.84%, p = 0.22). No category significant. Note: a quick estimate earlier the same day (posts +0.53% vs a daily-close baseline of −0.04%) overstated the gap — it used a shorter window and daily closes. This result supersedes it.
• Next: Backlog #1 — use these figures everywhere the raw average is still shown.

/map/ — the learning map (asked for by the owner)

Scanner audit, part 1 — each pattern tested against other hours (Backlog #1)

• Need: the scanner's raw win rates ("RSI overbought wins 70%") mean nothing without a baseline — in a falling market every bearish finding "wins" — and some patterns looked worse than chance. Traders see these findings on coin pages and in the bot with no warning.
• Built: scripts/scanaudit.py (CI step after scanlog): for every finding in scanlog.json, the 24h move in the signal's direction from the next full-hour open; baseline = all findings of a pattern shifted together by every k hours (circular, exact test over all shifts — keeps the cross-coin clustering); two-sided p around the shifts' median; verdict better / worse ("under review") / no_edge / too_few (n<30) at p<0.05. Writes data/scanaudit.json. Coin pages show the tested line instead of the raw win % when a verdict exists; the bot marks "under review" under each finding in Opportunities and lists every pattern's test in the Track section (fa/ckb/en). Nothing hidden.
• Verified: py_compile; unit tests on synthetic random walks — null findings significant 0/10 seeds, planted +4% move caught (bull "better" p=0.0012, its bear mirror "worse" p=0.0012), one move hand-checked (+21.82%); coin pages and bot sections rendered on a scratch copy with fake audit data; screenshots 1280/390, zero JS errors. Real figures come from the next CI run — not yet seen.
• Next: Backlog #1 — read the real audit, then the /opportunity/ table.

Public /autopilot/ log page (Backlog #1)

• Need: the loop changes the site every hour; readers (and the owner) had no way to see what changed and why on the site itself — accountability for the site's own changes, matching the "measured honesty" strategy.
• Built: scripts/build_autopilot.py parses this file's Log and renders /autopilot/ (fa/ckb/en frame, entries in English, bold/code only, everything escaped) with links to AUTOPILOT.md's git history and a search of autopilot: commits; footer link "Autopilot log" next to Methodology (chrome.py); builder added to update.yml and pages.yml; sitemap entry. No bot entry: it is a meta page with no natural home in the bot.
• Verified: py_compile; parser unit test on synthetic markdown (continuation lines, missing title, section end, empty input, HTML escaping); built on a scratch copy; screenshots fa/en at 1280/390, zero JS errors.
• Next: Backlog #1 — scanner audit.

One honest number everywhere (Backlog #1)

• Need: coin pages, the home/dashboard "post track record" box and the Entekhab Score still used the raw average 24h move after posts (no baseline) — in a drifting market that flatters the channel.
• Built: score.py channel_component() — the Score's channel part is 50 (neutral) unless the tested record (data/trackrecord.json, post hours vs matched random hours) has p < 0.05; then 50 + 8 × (post avg − random avg). build_pages.py tested_line() replaces the raw sentence on every coin page and the server-rendered dashboard box; the home page's client-side box (index.html, fa/ckb/en) now loads trackrecord.json and shows post vs random-hour BTC move, p and the verdict, linking to /track-record/. /methodology/ wording updated.
• Verified: unit tests on synthetic inputs (empty → 50; real p=0.28 → 50; p=0.01, +1.3 gap → 60.4; clamp at 0; p=0.05 boundary → 50); score + page builders run on a scratch copy (BTC/ETH channel = 50.0); screenshots home fa/en and BTC coin page at 1280/390, zero JS errors. Bot "track" section already quotes the tested figure (previous run).
• Next: Backlog #1 — public /autopilot/ log page.

/track-record/ — the channel's real record

• Need: every reader's first question — "does following this channel help?" — had only a misleading answer: the raw average 24h move after posts (+0.53%, 58% up over 524 posts) with no comparison. In the same window bitcoin's average day was −0.04% (48% up), so the number needed a real test.
• Built: scripts/trackrecord.py (CI, Binance 1h): entry at the open of the first full hour after each post, 24h hold; the same measures for every hour of the window; posts in one hour count once; matched-cluster randomisation (each post-day → a random day with the same number of random hours), 10,000 draws, seed 20031222; also the 4h move before the post (reaction) and the 24h range (eventfulness), per category, a histogram and a day-cluster bootstrap CI. scripts/build_trackrecord.py renders /track-record/ (fa/ckb/en) with a verdict chosen only by p < 0.05.
• Mirrored: menu (reality group), site assistant, first card on /reality/, Telegram bot section trackrec in the reality group; the bot's existing "track" section now quotes the tested number instead of the raw average.
• Verified: unit tests — null data never significant (0/10 seeds), planted +1.5% effect caught (p=0.0005), a "post after the move" pattern flagged as reaction and not as foresight, duplicates counted once, windows across data gaps skipped, one window checked by hand; full study runs in ~3.5 s; page screenshots desktop/mobile, fa/en, no JS errors. Real figures produced by CI (run 42372862451, 12:36 UTC) and the live page checked in a browser.
• Result (real data, 2026-08-04 → 2026-11-06): 508 posts in 259 distinct hours over 88 days. Bitcoin 24h after posts +0.43% vs +0.20% at matched random hours (95% of random results between −0.19% and +0.64%), p = 0.29 — not significant. Share of gaining windows 56.4% vs 52.9% (p = 0.47). No sign of posting after moves (4h before: −0.02% vs +0.04%, p = 0.32); volatility after posts like any hour (3.07% vs 2.84%, p = 0.22). No category significant. Note: a quick estimate earlier the same day (posts +0.53% vs a daily-close baseline of −0.04%) overstated the gap — it used a shorter window and daily closes. This result supersedes it.
• Next: Backlog #1 — use these figures everywhere the raw average is still shown.