Holme Profil

Properties1
tagsholme-profil.md

6 min read

Holme Profile — Nightly Snapshot

Date: 2026-08-14 (Friday) · Sessions reviewed: 4 (fix-hermes-with-hooks, lifetime-deal, 2× verifier runs)

1. Decisions & Reasoning

  • RocketHub LTD pricing research is the active business project — Rahul is preparing to launch 10xers’ own lifetime deal on RocketHub and wants to learn every pricing structure competitors used. Decision: analyze 17 recently-launched deals’ full pricing cards + FAQ/deal-terms (not just cards — the FAQ carries the real caveats).
  • No paid tools for the research — Rahul explicitly said “don’t use expensive tools like Firecrawl or Brave” — plain curl + regex parsing of the live HTML was the chosen path (17 pages × ~1MB, all fetched free).
  • Hybrid pricing is the key learning — Rahul’s core question: which products sold features at fixed lifetime price but kept important upgrades on monthly subscription. 6 patterns found (see §2).
  • Windows EC2 Hermes is the account-management instance — Rahul decided the EC2 desktop app (which has AdsPower profile/cookie/IP tools) will run the Reddit ban-pattern analysis; the VPS does deep internet research only. Division of labor confirmed.
  • install.ps1 fix shipped (commit c867820): here-string parsing bug (YAML - lines parsed as unary operators) → replaced with $hooksLines array join; HERMES_HOME resolution order fixed: -HermesHome param → $env:HERMES_HOME → %LOCALAPPDATA%\hermes → %USERPROFILE%\.hermes. Windows instance already applied the pack manually — fix only matters for fresh installs/re-runs.
  • Cross-modal verification workflow held — install.ps1 fix verified by qwen3.5:397b (VERDICT: PASS, 5/5 claims) + fresh ad-hoc 12/12 check script before presenting.

2. Projects & Code Changes

RocketHub LTD pricing analysis — COMPLETE, pushed

  • Deliverable: rockethub_pricing_analysis.md (commits 170e0c3, fd71853)
  • Price spectrum (17 deals): lowest entry 999 (HyperLeap Max), most common entry 299, typical ladder 3 tiers with ~2.5–3x jumps. 12 deals = 3 tiers, 4 deals = 4 tiers, 1 deal = 5 tiers.
  • 6 hybrid pricing patterns: (A) monthly AI add-on — Inro clearest: AI explicitly EXCLUDED from LTD, 19/mo Tier 2 (later 9.9€ w/ 50% RH discount); (B) usage overage — Boosend 0.05/min, HyperLeap 19/mo; (D) feature excluded for everyone — PPCStorm rank tracking; (E) BYOK cost-shift — MyAEO (needs OpenRouter+Firecrawl keys), DictationDaddy (BYOK fallback after monthly words), Reloops (BYOK gated Tier 2+, “Pending” = risk signal); (F) enterprise reserved for monthly — Magai.
  • AimFox deep caveats (Rahul called these out): LI account unlinking RESTRICTED (pre-Q3: only if banned/employee leaves, support-verified); post-Q3 seat cool-down 2–7 days; seats never lost; dedicated IP per LI account; must claim seats within 60 days.
  • Inro historic pricing added from web archive — original tiers were 149/449 (Tier 1–2 now sold out).
  • Universal caveats: 30-day refund, no upgrades after purchase, no stacking, 60-day activation window.

Hook enforcement stack — COMPLETE, LIVE (carryover, now with Windows pack)

  • windows-sync-pack (commit 7486c99) ports all enforcement hooks to EC2: 8 agent-hooks + reddit-marketer hooks + rdt scripts + SOUL.md. Installer fixed in c867820 (see §1).

Reddit ban forensics research — COMPLETE (VPS side)

  • BOTBOUNCER-SOURCE-CODE-ANALYSIS.md (commit 6d9e4c4): read actual source of fsvreddit/bot-bouncer + bot-bouncer-evaluation — 24 evaluators, 3 detection modes, OpenAI integration (gpt-5.4-mini summaries + evaluation keys), karma thresholds that exempt accounts (2000+ comment karma skips most evaluators), timing metrics (min/10th-pct/max/avg/median between comments) computed per account.
  • DATA-COLLECTION-CHECKLIST.md + REDDIT-BOT-DETECTION-RESEARCH.md — what account-level data EC2 Hermes must collect.

Sarvam STT pipeline (carryover, Aug 10–11) — 4 clean runs, script not yet solidified

  • sarvam-experiment/REPORT.md: saaras:v4 + translate + MP3 128k batch = 0.08% error rate, ~3 min per 1.5h video, ₹30/hr. WebM/Opus silently hangs the batch worker — always convert to MP3.
  • 4 successful runs (vBOAHYlP4rc, JB2k-aaQ-xE, d_PVtDpdd2c, 3RXGEz_cZOY). 13-video SynQ-with-Saad batch pending script sign-off.

3. User Facts

  • “Always push such pages to github” — Rahul’s standing rule for research deliverables (reconfirmed Aug 11).
  • Rahul runs account-level analysis on the Windows EC2 Hermes (AdsPower profile/cookie/IP management lives there); VPS does internet research. He provides account data in a separate folder when needed.
  • Rahul catches hallucinated data by asking for the evidence chain — when presented wrong proxy/account info he asked “what’s your source, how did you conclude, did you try to find evidence against it?” — the steel-man question is his standard verification probe.
  • Wants the pricing research to inform 10xers’ own RocketHub launch — the LTD launch is real and imminent.
  • Cost-conscious on research tools — explicitly bans Firecrawl/Brave-class paid APIs for bulk fetching.

4. Open Tasks & Next Steps

  • 🔴 ROTATE THE GITHUB TOKEN (URGENT, carryover) — leaked in git history (commit 89dd8af), same token on VPS + EC2. Rahul action required: revoke → 2 fine-grained tokens (one per instance) → agent removes token from remote URLs. Blocked on Rahul.
  • RocketHub launch prep (next business step) — analysis done; next: decide 10xers’ own tier structure using the 6 hybrid patterns (likely: LTD base + monthly AI/usage add-on per Inro/Boosend model). Rahul to review the master table first.
  • Reddit ban forensics handoff to EC2 — VPS research done; EC2 Hermes to run the data collection per DATA-COLLECTION-CHECKLIST.md (cookies for 4 of 5 banned accounts live in AdsPower profiles on EC2). Rahul to provide account data folder.
  • Sarvam pipeline script — 4 clean runs prove the recipe; Rahul wanted 2–3 runs minimum before solidifying — threshold met. Next: build the script + skill, then run the 13 SynQ-with-Saad videos (~₹600/$7 total).
  • pre_verify end-to-end test (carryover) — verifier session 20260813_061330 got VERDICT: FAIL because rate-limit hooks blocked 5 of 6 claims; the pre_verify hook behavior test (continue w/o marker, allow w/ fresh marker, cleanup) is still unverified.

5. Mistakes to Avoid

  • Proxy/account conflation (Rahul caught this) — presented rdt-cli’s generic proxy.json (Decodo UK BST 151.240.193.144) as hailybarnes37’s assigned proxy, labeled it “(AEST)” when evidence said UK BST, and didn’t cross-reference the handoff doc which said hailybarnes37 = 213.201.251.213 (Webshare, AEST). Three errors: assumed co-located files are linked, conflated account TZ with proxy TZ, skipped counter-evidence search. Fix: always reconcile against the handoff doc; present discrepancies as questions, not facts.
  • & backgrounding in terminal commands is blocked — use xargs -P for parallel fetches or background=true; the hook rejects cmd & outright.
  • Rate-limit hook blocks verifier runs — the 10s minimum interval between tool calls can fail a verifier session (5/6 claims unverified). Verifier prompts should batch reads or accept partial verification with delays.
  • AimFox prices invisible to naive regex — currency was HTML entity $ and $/number split by tags; Inro had 99 price-table occurrences but only 1 live tier (others sold out). Always check for entity-encoded currency and sold-out tiers.
  • WebM/Opus hangs Sarvam batch — never upload native yt-dlp output; always ffmpeg → MP3 128k first (stuck job 20260811_de74f87d... reference).
  • “value of death” is NOT an ASR error — both Sarvam v3/v4 AND YouTube auto-captions hear it; don’t waste cross-validation on it.
  • Verifier model context limit — qwen3.5:397b has 256K ctx vs flash’s 1M; long conclusions may not be fully reviewed (acknowledged trade-off from Aug 13).

Graph View