PRISM (Second-Brain)

index.md

PRISM — Wiki Index

Current state (compiled truth — 2026-07-12)

Finishing Build Plan written + phased execution decided, kicking off Phase 1. Live VPS/GitHub/local recon (not doc-derived) found the 2026-07-06 V2 handoff was only half-executed: role doors (AE/BDR/Marketer) + a real Dell/Nike landing-page builder DID ship 2026-07-06, but the architecture work (Hermes removal, executioner rebuild, chat-as-operator) never started — Hermes/Cassandra is still the live executioner, zero Claude Agent SDK code exists anywhere. Also found, unrelated: POST /api/chat on prism.chowmes.com is unauthenticated in prod today — a live security hole, being patched first regardless of the rest of this plan. Arijit locked 4 decisions: executioner = upgrade prism-runner.py (not an SDK rewrite — unproven, bigger lift, not needed until multi-tenancy/chat-as-operator demand it), full Hermes rip-and-replace (not staged, but cutover-sequenced to prove parity before removal), build the real Figma→Jahia landing-page integration (not just static-template reuse), and scope = full productization (Tracks A-H, sellable product not just internal tool). Execution model: 6 phases, each its own Fable-5 goal, hard review gate between phases — not one autonomous run (the 07-06 "run the loop yourself" framing already produced a silently-half-finished result once). Phases: (1) Foundation — auth patch + re-verify the 3 role doors against current data + design-system reconciliation; (2) Executioner + anti-handwaving Verification Pipeline (5-stage gate: mechanical→factcheck→3-agent adversarial panel→quality→legal) + Hermes removal — highest-risk phase; (3) Landing pages — Belk/Lululemon + a Figma/Jahia integration spike; (4) Modular rearchitecture; (5) Multi-tenancy + domain-pack, gated on finally running the concurrency test open since 07-02; (6) Prod-readiness + GTM + pricing. Full plan: PIP repo docs/plans/2026-07-12-prism-finishing-build-plan.md. Decision record: 2026-07-12-finishing-build-plan-phased-execution.

Prior state (2026-07-10)

Belk re-validation session, 2026-07-10 (post-midnight, picking up where 2026-07-09's crashed run left off). The SOP written the night before (docs/sop/AUDIT-REVALIDATION-SOP.md, PIP repo) got its first real test on a resumed company. Found the crashed session had left Belk in a dangerous half-state: research files were fixed but the deliverables (the actual audit-data.json, the live page, the PDF) were never regenerated from them — meaning fabricated hiring roles and a 3x-wrong competitor revenue number were still live on the served page. That page was pulled and replaced with a holding page the moment this was confirmed, before any further work. Root-caused and fixed 8 real defects across the night, several found only because Arijit kept pushing back on "done" claims rather than accepting them: a database-write agent that silently wrote from the wrong source file (caught via a nested-field check, not the top-level status columns that looked fine); a client-render field-name mismatch that flattened 6 partner relationships into identical fake-looking cards (pruned to the 1 actually-confirmed one, per Arijit's explicit "only validated stays" rule); a screenshot filename mismatch that made every finding image 404 silently; Arijit's own manually-captured, WAF-passed screenshot that existed on disk but was never wired into any finding; a signal mislabeled "Partnership" that was actually a competitive-threat entry; a self-admittedly weak signal that was still rendering as a HIGH-urgency red badge because the renderer hardcodes severity; a Solution Map with no real case-study proof; and — found only because the final gate was told to independently spot-check sections nobody had touched — a duplicated press quote mislabeled as a fabricated "leadership change." Final gate: v6, WARN, score 8.3, blocking_count 0. Belk is live again with the corrected data; the WARN label (not a clean PROCEED) is cosmetic-only, same class of carried warning Lululemon shipped PROCEED with. A real process lesson surfaced hard tonight: checking that a database write happened is not the same as checking the write happened correctly — every fix from here needs the field-level check, not the status-column check. Full detail: 2026-07-10 entry.

Prior state (2026-07-09)

PRISM V2 build is UNDERWAY, not just planned: the V2 manifesto (3-phase re-architecture: remove Hermes entirely, Postgres+pgvector + Claude Agent SDK, plug-and-play modules, domain-agnostic productization) kicked off 2026-07-04, and the first V2 vertical slice is LIVE — an AE-role cockpit on prism2.chowmes.com running off real Postgres audit data (dell/belk/lululemon), live-verified 2026-07-06. Next: BDR+Marketer cockpit fan-out, awaiting Arijit's go-ahead; chat-operator is backend-gated. Source: tracker log 2026-07-04→06.

Reliability crisis surfaced and partially fixed, 2026-07-09. A routine "fix Lululemon's SimilarWeb data" request turned into a 6+ hour forensic pass that exposed the real reason audits keep shipping broken: there is no single source of truth for the pipeline's code, its data, or its output shape. Three bugs found this session affect every company audit ever run, not just Lululemon — a data-quality flag that never reached the factcheck gate, a ROI calculator formula that inflated dollar-impact estimates roughly 5x, and a filename typo that has silently zeroed out the industry-benchmarks section on every single audit. The deeper cause is that the GitHub repo of skill code and the actual code running on the VPS have quietly drifted apart across roughly forty skills — one skill's instructions were caught still telling agents to use a dead, five-days-fixed API. Two skills were fixed and locked together (via a symlink, so they literally cannot drift apart again); the rest still need the same treatment. A first real step toward "the live page should just read the database" was also started: a working API endpoint that pulls straight from Postgres, verified returning correct live data — but wiring the page to actually call it hit a genuine login-session bug that was deliberately left unfixed rather than patched insecurely. A full written plan for finishing this cleanup (unify the code, move all data into the database, delete the old file-based copies, and make every skill check its own work before handing off) is saved at arijit-skills/docs/PIPELINE-UNIFICATION-PLAN.md on the VPS repo, ready for a fresh session to execute. Full detail: 2026-07-09 entry.

Read this file first. It maps the entire wiki.

Files

File	Purpose	Last Updated
wiki/overview.md	What PRISM is, architecture, personas	2026-06-29
wiki/requirements.md	Spec and acceptance criteria	—
wiki/decisions/	Architectural decisions (ADRs)	2026-06-30
wiki/dev-log.md	Build chronicle	2026-06-30
wiki/open-questions.md	Open questions	2026-06-30
wiki/sources/	Research source summaries	—
wiki/concepts/	Domain concept pages	2026-06-30
wiki/entities/	People, orgs, tools	2026-06-30
wiki/syntheses/	Cross-cutting analysis	—
wiki/IA-Redesign-Pending.md	PAUSED status: report IA redesign A/B prototype (resume: "resume IA-Redesign-Pending")	2026-06-30
wiki/2026-07-04-v2-manifesto.md	PRISM V2 — 3-phase re-architecture starting: (1) remove Hermes entirely, rebuild executioner/chat/data on Algolia tech, (2) plug-and-play modules, (3) domain-agnostic productization. Repo copy: `PIP/docs/PRISM-V2/00-manifesto.md` (canonical, keep in sync)	2026-07-04
wiki/research/2026-07-04-algolia-as-database-feasibility.md	Fresh research: Algolia as PRISM's data backend — good fit for content/search layer, weak fit for executioner state-tracking (no transactions/joins)	2026-07-04

Decision Records

ADR	Summary	Date	Status
2026-06-27-search-vendor-packet-detection	Search-vendor detection via live network-packet inspection (replaces substring source-scan)	2026-06-27	Accepted
2026-06-27-persona-journey-ui-architecture	UI = three role-shaped doors (AE→BDR→Leader) over one intel engine; calm Claude-desktop language	2026-06-27	⚠️ Superseded (chat agent, not a web app)
2026-06-27-prism-as-ai-sales-toolkit-stage-cockpit	PRISM = AI Sales Toolkit + FY27 5-stage motion (still valid); built "stage cockpit" UI superseded	2026-06-27	⚠️ Partially superseded
2026-06-28-prism-is-chowmes-prism-hermes-instance	PRISM = Chowmes-PRISM, a dedicated Hermes instance on the VPS; Temporal dropped; Gemini control / Claude exec	2026-06-28	Accepted
2026-06-28-grounded-report-qa-gate	Grounded report-QA via Hermes plugin (pre_llm_call inject + transform_llm_output Gemini gate, no fork); facts grounded, coaching allowed	2026-06-28	Accepted
2026-06-28-unified-chat-on-prism-hub	Unified grounded chat: Hermes one-brain on the published reports hub + Telegram; LIVE/proven (PetSmart 15.98%)	2026-06-28	Accepted
2026-06-30-about-splash-routing	About is the splash at `/`; reports list moves to `/reports/`; `/about/` redirects	2026-06-30	Accepted
2026-06-30-github-vps-auto-deploy	GitHub webhook to VPS git pull; prism-hub auto-deploys on push, no manual scp	2026-06-30	Accepted
2026-06-30-cassandra-naming-and-persona	One name "Cassandra" (Cass removed); Imagen-4 portrait; session auto_prune 7d	2026-06-30	Accepted
2026-06-30-detect-search-vendored	detect-search vendored into arijit-skills, symlinked, runs on the VPS	2026-06-30	Accepted
2026-06-30-no-em-dash-writing-standard	No em dashes in reader-facing copy; rewrite grammar instead	2026-06-30	Accepted
2026-06-30-wave2-similarweb-screenshots	Wave-2 traffic from manual SimilarWeb screenshots; Wave-2 now fully keyless	2026-06-30	Accepted
2026-07-12-finishing-build-plan-phased-execution	Finishing Build Plan: executioner=prism-runner.py (not SDK), full Hermes rip-and-replace, real Figma→Jahia integration, Tracks A-H scope, 6-phase gated execution	2026-07-12	Accepted

Entities

Entity	Type	Role
prism	product	Orchestrates the algolia-* skills; the whole engine
hermes	tool	Runtime that runs the skills and serves the agent
cassandra	agent	Grounded report-QA and sales-coach agent on Hermes
scout	tool	Acquires the target's first-party data
detect-search	tool	Keyless tech and search-vendor detection
gemini-grounded-search	tool	Open-web research
yfinance	tool	Public financials
chrome-browser	tool	Live search testing

Concepts

Concept	Summary
right-tool-for-each-job	Evidence-routing doctrine: each fact-job to its own tool, keyless, no fabrication

Sources

Title	Slug	Date
(none yet)

Recent Activity

Date	Event
2026-06-28	ADR: unified chat LIVE on prism-hub + Telegram (PetSmart 15.98% proven)
2026-06-29	Standardization Wave 1 complete; Cassandra persona built and dialed in
2026-06-29	Linked Scout as PRISM acquisition plane
2026-06-30	About page becomes the splash at `/`; reports list to `/reports/` (ADR)
2026-06-30	GitHub to VPS auto-deploy webhook live (ADR)
2026-06-30	Cassandra rename, Imagen-4 portrait, session auto_prune (ADR)
2026-06-30	detect-search vendored and running on the VPS (ADR)
2026-06-30	No-em-dash writing standard (ADR)
2026-06-30	Wave-2 SimilarWeb blocker resolved via screenshots (ADR)
2026-06-30	Entities and concept pages added (Scout, Hermes, Cassandra, detect-search, and more)
2026-07-04	PRISM V2 manifesto started: Arijit wants Hermes removed entirely (security blocker for Algolia VPN adoption) and PRISM rebuilt on Algolia tech (Agent Studio for chat, Algolia itself for data, executioner TBD — Temporal vs Agent Studio). Found + surfaced a critical prior-decision nuance: Temporal was dropped 2026-06-28 specifically because it was redundant alongside Hermes, not because it's unsuitable — that objection no longer holds once Hermes is gone. Fresh research done on Algolia-as-database (mixed: good for content/search, weak for transactional state-tracking). Also found and closed a real vault-sync gap: Agent Studio native-capability findings and the full Algolia API reference existed only in the Algolia-Central2 repo, never committed here, until now.
2026-07-10	Belk audit re-validation resumed post-crash: fabricated live data caught and pulled (holding page), 8 real defects found and fixed (DB-write source bug, partner-intel field mismatch, screenshot filename mismatch, unwired manual screenshot, mislabeled signal badge, false-HIGH severity on a weak signal, missing Solution Map proof, duplicated/mislabeled press quote found by mandated spot-check). Final gate v6: WARN, score 8.3, blocking_count 0.
2026-07-12	Finishing Build Plan written after live recon exposed the 07-06 V2 handoff was half-executed (doors+landing builder shipped, architecture rebuild never started, Hermes still live). Found unrelated live security hole (`/api/chat` unauthenticated). 4 decisions locked (executioner=prism-runner.py upgrade not SDK; full Hermes rip-and-replace; real Figma→Jahia integration; scope=Tracks A-H). Execution model set to 6 phases with mandatory review gates, not one autonomous run. See 2026-07-12-finishing-build-plan-phased-execution.