← Back to MuseHub

VentureHub Master Assessment

Date: 2026-09-14 Consolidated by: Agent Muse, from four independent adversarial reviews Independently audited: a second agent re-consolidated all four reviews from scratch and diffed this draft; its corrections are incorporated below (notably: the unknowns remedy is genuinely contested, §2 T5; Codex's store trigger was initially narrowed and is corrected at T1/D2). Build under review: the 13 Sep 2026 VentureHub ("Agora Ventures · VentureHub", ~130 KB single-file HTML, 14 hash views) Status: synthesis + proposals only. Nothing here has been applied to the hub.

Sources

Reviewer 3's note: Instinct's v2 is a second pass confirming fixes from its own earlier first review (not in this batch). Count note: reviewers report different seeded-task counts — 27 (Cursor), 29 (Claude), 28→29 observed across a reload (Codex). Trivial, but the discrepancy is real; treat "≈28" until counted from source.

1. Consensus findings — what the reviewers agree on

C1. Persistence is the P0 gap. (4/4)

Every reviewer names in-memory-only state the single biggest defect.

• Claude: "The interactive layer currently lies to the reader." Checkboxes, factor edits, weights, theme reset on reload; on a phone, background-tab eviction kills state within minutes of switching apps. A factor set on the Nile page read back empty after reload [measured].
• Codex: Reproduced reload loss end to end (two factors, a weight, a traction-step field, a checkbox, the theme — all reverted). "Make every accepted edit recoverable" is the P0. Until demonstrated, treat the Google Doc mirror as a reference copy, not a recovery mechanism — the app script has no code path that writes browser edits to the Doc.
• Instinct: "The big one. Until this lands, VentureHub is a scratchpad, not a system of record."
• Cursor: Failure forecast #2 — "Data loss on reload — first real use. One silent wipe of a checked Nile gating item retires trust."

Corollaries the reviewers agree on: - The mirror does not back up decision state: it carries no factor values, weights, or task items (Claude, Codex; Cursor could not verify past the login wall). - Session-only labeling is honest but does not make the behavior acceptable for daily reliance (Codex, Claude, Cursor). - Claude's interim mitigation, if persistence is not built immediately: remove or visibly label every non-saving control ("scratch — not saved") at the point of use.

C2. Sparse profiles outscore researched ones — diagnosis unanimous, remedy contested. (4/4 on diagnosis)

scoreIdea normalizes by the weights of known factors only. Consequences, verified by three reviewers against the live formula:

• One factor set to max → 100/100, ranked above fully-described ventures (Claude [measured] on TikTok Shop; Codex reproduced: 1 factor at max = 100/100, then 50/100 after adding dread=1, then 60/100 after raising the weight; Cursor's worked table: TikTok with only enjoyment=5 scores 100).
• Admitting an unfavorable fact lowers a score versus leaving the field Unknown — research is punished (Claude R3, Cursor §3).
• Instinct's "fixed" claim needs an honest flag: Instinct wrote that the "100/100 for a 1-of-12-factor idea" failure is gone — on the same 130,269-byte build where Claude measured one-factor→100/100 and Codex and Cursor reproduced it. These empirical claims directly conflict. The charitable reading is that Instinct's "fixed" covers the zero-factor case (now null / "Needs scoring", sorted to bottom), but that is inference, not something Instinct wrote. Verify from source rather than asserting.

Where the reviewers split on remedy — see T5. Claude/Cursor/Instinct would charge sparsity inside the score (pessimistic imputation, coverage scaling, rank on the lower bound). Codex argues the opposite remedy family: keep Unknown null, and let unknowns limit the commitment, not reduce the value. Do not treat this as settled.

C3. The formula does not implement the stated rationale. (3/4 + code)

• The handoff/Philosophy says confidence and enjoyment are multipliers (even gates); the code makes them weighted addends — enjoyment carries weight 2 of 25, so dread=1 costs ~8 points out of 100 (Claude R2 [code], Cursor §3: "three different contracts" — gates in Philosophy §05, multipliers in Handoff §03.1, +2 in code).
• Research depth is a benefit factor (more research = more points) when it should be a confidence input (Claude R4, Cursor's absorption table; Codex adds the caution that rewarding research depth can reward documentation volume).
• Codex dissents on the remedy, not the diagnosis: multiplication needs a defensible meaning — an enjoyment rating is not a probability, and uncalibrated multiplication manufactures precision and can double-count uncertainty. Resolve explicitly; do not auto-implement the multipliers.

C4. Two/three ranking signals coexist; only one should. (3/4; Instinct's short review doesn't mention it)

The computed board (currently empty — all rows "Needs scoring"), the hand-written "Initial focus bands" ("Rock now: AccelerateBooks…"), and per-page prose "Priority" labels all answer "what next?" (Claude D6, Cursor: Law of Proximity; Codex: label bands as dated human priorities). Agreed fix: one canonical signal — generate the bands from the score, or date-stamp them as AK's last stated order.

C5. The muse.ai share URL is not a suitable daily driver. (4/4)

• Instinct [tested twice]: the share host serves the real app only to real browser navigations; programmatic fetches (curl, any UA) get a 302 chain back to the share shell. Agents that fetch without a browser cannot read VentureHub at all.
• Cursor: wrapper chrome ("unverified", Sign up) sits over a cockpit holding LLC status and brand-isolation rules; the URL was stuck "deploying" earlier the same day. (Codex notes the deployment placeholder resolved in his latest session — the flake happened, then cleared.)
• Claude: "temporarily on Agent Muse's web runtime" with no owner or retirement condition recorded.
• Codex: keep an export; verify deployments from the reader's entry point.

Agreed: treat the share as a preview. The daily driver lives somewhere AK controls.

C6. Keep these without argument. (3/4 explicitly; Instinct silent)

• Stage as a filter, never a score (Claude, Codex, Cursor; Instinct doesn't mention it).
• Unknown as a legal value (all four mention or assume it — once it costs, per C2/T5).
• Entity layer above/excluded from the ranking (all four; Claude: "conclusion 6 made visible"; Codex: but let the entity own tasks).
• Hard constraints (brand separation, commitments) outside the score (Claude, Codex, Cursor; Instinct silent).
• The single-file, no-framework, no-backend output stays portable and is fine as a surface (all four; Codex: "a framework rewrite is not the necessary next step").
• Dated CHANGELOG / temporal traceability (Claude D14, Codex §8, Cursor §7).

2. Contested — where the reviewers disagree, and my read

T1. Where should durable state live?

My read: three different clocks, and an earlier gloss of mine conflated two of them — corrected here. Claude: server only at "an agent must write unattended." Codex: shared store immediately if phone+desktop+agent writes are needed for first real use — a much earlier trigger than Claude's. Cursor: staged triggers, no DB "this quarter." And: read access does not fire anyone's trigger. Phone + desktop viewing and agents fetching are hosting problems (Instinct's 302 finding fixes that with a one-line move); the store trigger in all three proposals is about writes. The operative question for AK is therefore precise: for the first real use, do agents or a second device need to write VentureHub state — or only read it?

T2. How strong should energy/dread be in the model?

• Cursor: implement the multiplier — score *= (enjoyment/5)^1.5, dread=1 sends the score toward ~9%.
• Claude: two-stage — gates (energy floor among them, pass/fail, never weighted) → value computation; publish the formula. Also: score energy_now vs energy_week6, since novelty inflates capture-time ratings.
• Codex: narrower than the draft first suggested — persistent unwillingness can justify redesigning, delegating, or declining a discretionary commitment, but temporary reluctance must not automatically erase a venture's estimated value. And: ground the claim in AK's statements or observed outcomes; no diagnosis supplies a scoring coefficient.

My read: all three agree dread must be able to kill or reroute a venture, not merely discount it ~8%. One genuine partial resolution the draft missed: Claude grounds dread-as-kill-switch in AK's own journals (19 Dec 2025: "your system doesn't have a 'moderate engagement' setting. You either obsess or you neglect") — which is exactly the "link it to his own statements" evidence Codex demanded. So: an energy floor as a gate (Claude) with explicitly AK-defined rating semantics (Codex), and Cursor's exponent only once the rating is calibrated to something measurable.

T3. What is the decision unit — the venture or the next action?

• Codex (unique, strongest version): rank the next bounded commitment, not the venture: owner, effort, cash, dependencies, deliverable, expected signal, stop condition. A venture can hold both a worthwhile small experiment and an unjustified large investment; one lifetime score hides that. Proposes a 5-criterion scorecard with anchored 0–4 scales.
• Claude: compatible — adds window, kill criteria, depends_on, AK-hours vs agent-hours; rank on pessimistic bound.
• Cursor: compatible — "smallest next traction step" wired into the ranking subtitle; seed from page text.

My read: no real conflict on the what, some on sequencing — Codex's scorecard is a designed replacement; Claude's and Cursor's are incremental patches. The venture page keeps the narrative; the ranked object becomes the next tranche. ("Absorbs" is my inference, labeled as such.)

T4. Tasks: delete the API, or extend it?

• Claude: delete VentureHub.tasks now — "the API is engineering against a hypothetical." Checklists live in profiles; pair with GitHub issues when the issue layer lands.
• Codex: extend the seam to durable ops with explicit success/failure, revisions, idempotent creation. "The existence of a namespace does not make that migration automatic."
• Cursor: keep as the seam; do not plug in MCP until persistence exists.

My read: sequence it honestly — keep the seam (Cursor), make it durable with revisions if the T1 decision lands on the service (Codex), and note that the keep→durable→revisions path presumes the service outcome; if T1 lands on vault-first, Claude's delete-and-use-profiles is the coherent choice.

T5. Unknowns: penalize sparsity in the score, or keep Unknown null and bound the commitment?

The most contested scoring question in the batch — the draft initially presented one side as consensus; corrected here.

My read (mine, not theirs): the diagnosis is unanimous — the board cannot keep ranking sparse optimism first. The contested part is whether the penalty lives in the number or in the commitment. Note Codex's caution here is the same one he raises at C3 (don't manufacture precision from uncalibrated inputs). One mechanism draw no fire from any reviewer: a minimum coverage threshold for rank-eligibility — Codex never opposed gating eligibility, only silent value-assignment. A defensible middle is coverage-gated eligibility plus scenario bounds instead of pessimistic imputation — but that synthesis is mine; §5 D4 puts the real choice to AK.

3. Single-source findings worth keeping

Things only one reviewer caught, each load-bearing in its own right.

From Claude

• No time axis; the hub disagrees with AK's calendar. The hub says "Rock now: AccelerateBooks." AK's 30 Aug plan says September = loose ends, October = Nile Outbound + Zeitgeist Apparel, Phoenix freeze 26 Oct–30 Nov; FMS14-5 (8 Sep) says AccelerateBooks "is not being revived this cycle." Proposal: window (Now / Next / Later / Shelved, with dates) as the default order + a portfolio WIP limit (~one dominant plate). This is the review's most important unique contribution.
• Roster reconciliation: 7 vault-named ventures are absent from the hub — Zeitgeist Apparel (in the October plan; "the most important omission"), FutureMD, AgoraMedia, AcquireSoft, AllTheBest, doctrine-by-design, FortySeven Media. Four hub ventures exist nowhere in the vault.
• "MVC" is mislabeled. The vault's pillar 3 is Minimum Viable Complexity, not Model-View-Controller. Any engineer reading "MVC" assumes the latter. (Fix-now label change.)
• Two business decisions were promoted to constraints without a dated ruling: the AccelerateBooks "can't give up the ship" commitment (the vault documents a revive-vs-bankruptcy fork) and the LLC name. File as ADRs or mark proposed.
• Missing factors: AK-hours vs agent-hours; days-to-first-cash and cash-at-risk (AccelerateBooks carries a $226K SBA EIDL; runway is HELOC-funded); reversibility/exit cost; kill criteria on every core venture; depends_on; as_of staleness on every factor ("the ranking's freshness is its oldest input").
• Scaffolding needs a budget and a stop line, "or it becomes where the year goes" — link the vault's existing scaffolding plan as the meta-workstream with its budget shown, don't re-own it.
• Measured-values table (phone 390×844): 96 factor fields, 0 filled; ≈28 tasks, 0 complete; idea pages at ~13px body; first venture card 979px down the hub; hamburger 20×38px with no accessible name; 2 contrast failures (eyebrow 2.94:1, "Ideas" heading 4.41:1).
• D17: the Google-Doc mirror line sits above the sticky nav on every view — the first thing on every phone screen is a pointer to a copy.
• Occupied capacity: Project 295, Evexia/Phoenix, Hexis, LifeOS loose ends compete for the same attention — a window view should show them as occupied capacity, not as rows.

From Codex

• Real defect — fix now: the traction-step field is interpolated into innerHTML by renderRendering/renderRanking (bold-markup test rendered as HTML). Render user text as text; the task list already uses textContent. (Security/correctness bug, not a design opinion.)
• Define "traction" before scoring it: "a relevant customer response, observed use, a paid commitment, protected existing revenue, or resolution of a decision-critical unknown" — record the deliverable and what it will demonstrate.
• Commitments need service obligations: translate "can't give up the ship" into obligations, owners, timing, reserved capacity. "A low discretionary priority should not hide an obligation."
• Label every statement's epistemic status: AK-stated vs externally supported vs estimated vs proposed. "Never convert an agent inference into a business fact through repetition."
• Scaffolding math: a multiplier applied to every venture leaves their relative order unchanged — it cannot answer whether the scaffolding is worth building. Budget per tranche against a named bottleneck with a stopping point.
• Weight-preview semantics: editing a weight may update a preview, but must not silently replace the action AK already selected; allow a reasoned override retained in decision history.
• Deep links untested: hash routing + pushState keeps the share URL unchanged; verify a venture link restores on fresh load.
• Acceptance bar for the next review: six demonstrable outcomes (captured idea survives reopen; completed task stays complete + undo; clean restore preserves relationships; stale conflicting edit detected; uncertain ranking reports its uncertainty; AK can start the selected action on his phone unconfigured).

From Instinct

• The hosting fetch filter [tested twice, 20 min apart]: programmatic fetches to the deployed iframe URL 302 back to the share shell. Any agent of AK's that reads without a browser is bounced today. (The finding with the most immediate operational consequence for the agent fleet.)
• Confirmed the first review's fixes actually landed (entity exclusion, tasks page + API seam, Philosophy, CHANGELOG, honest architecture section) — the only reviewer doing a before/after.

From Cursor

• Dark-mode chrome inversion (Jakob's law): in dark mode the page goes dark but the sticky nav/entities go light #e8eef6, with Muse's wrapper staying light — three stacked themes on a night-time phone. Fix: dark nav in dark mode; optional prefers-color-scheme default.
• The daily control is the smallest target (Fitts): todo checkboxes are 16×16px; factor selects 34px; filter selects 40px. 44px minimum on the daily control. (Codex qualifies: WCAG 2.2 AA's actual minimum is 24×24 CSS px — 44px is the Apple HIG/design-choice bar, not the WCAG failure bar.)
• Worked scoring table (the clearest demonstration of C2): TikTok/enjoyment-only = 100; +researchDepth=1 → 67; plausible full AccelerateBooks = 56; high-TAM + dread=1 = 44.
• "Seed, don't invent" table: per-venture seed values drawn only from sentences already on the pages — the honest way to populate the board; or don't offer Weighted priority until coverage exists.
• Failure forecast ordered by time-to-hurt (public neurotype now → reload loss → sparse-score self-deception → stale bands → no capture → two writers/no store → MCP-on-nothing → 50 ideas → hosting flake).
• P0 "Private the Handoff" — the sharpest framing of the privacy finding — plus the sequencing question: which P0 first: privacy, persistence, or formula? (Cursor recommends privacy → persistence → formula.)

4. Consolidated roadmap

Sequenced; each item names its source reviewers. Items marked [T5] depend on the D4 decision.

Fix now (no decision needed)

• XSS/HTML-injection in renderRanking — render traction text as text (Codex).
• Privacy P0 — see §5 D1; delay has no upside (Claude, Cursor).
• "MVC" → "Minimum Viable Complexity" wherever the pillar is named (Claude).
• Mirror line out of the prime viewport — move to menu footer (Claude D17).

P0 — trust

1. Durable state — §5 D2. Acceptance bar is Codex's: edit → reload → close/reopen → still there; export → clean restore; failed save never labeled saved.
2. Unknowns remedy — §5 D4 / T5. Interim step no reviewer opposes: minimum coverage threshold for rank-eligibility + explicit "needs estimate" states.
3. One ranking signal — bands generated from scores, or date-stamped as AK's stated order (Claude, Codex, Cursor).
4. Privacy — neurotype off the public page (Claude, Cursor).

P1 — decision quality

1. Time axis — window (Now/Next/Later/Shelved + dates) as default order; WIP limit; reconcile with the 30 Aug / FMS14-5 calendar; window view shows occupied capacity (Project 295, Phoenix/Evexia, Hexis, loose ends) as capacity, not rows (Claude; Codex's capacity context compatible).
2. Action-level ranking — the ranked object becomes the next bounded tranche with owner/effort/cash/signal/stop-condition (Codex), absorbing kill criteria, depends_on, AK-hours vs agent-hours (Claude).
3. Traction defined before scored (Codex); scaffolding tranches get budgets + stop-lines against named bottlenecks (Claude, Codex).
4. as_of dating on every factor; the hub shows its oldest input (Claude).
5. Capture — one-field add-idea / add-task on the phone; capture never requires a scorecard (Codex, Cursor, Claude).
6. Mobile pass — 44px daily controls (Cursor, Claude; Codex: 24×24 is the WCAG AA bar, 44px the HIG bar); dark nav in dark mode + prefers-color-scheme default; collapse the 12-factor editor behind a scorecard line; ≥16px body on detail pages; next-action + tasks above the factor editor; list default on narrow screens (Claude D8, Codex); bottom tab bar + filter collapsing under ~12 ideas (Claude D2/D4); contrast, menu accessible-name, and Escape-to-close fixes (Claude D15).
7. Seed-don't-invent — populate only from sentences already on the pages, or hold the weighted view until coverage exists (Cursor).
8. Epistemic-status labels — AK-stated / supported / estimated / proposed on every claim (Codex); service obligations for the AccelerateBooks commitment (Codex); ADRs or "proposed" marks for the LLC name + AccelerateBooks commitment fork (Claude).
9. Weights governance + preview semantics — read-only on the hub; changes only via dated criteria-review entries (Claude, Cursor); weight edits preview without silently replacing the selected action; reasoned overrides retained in history (Codex).

P2 — scale and orientation

1. Second-client readiness — data separate from chrome at the T1 trigger; JSON + Markdown exports; revisioned backups with tested restore (Cursor, Codex, Claude).
2. Roster ownership — the 7 missing vault ventures enter whichever system D2 designates; no hand-duplication (Claude).
3. Vocabulary mapping to AK's terms (What's Best Next, Jenga Block, One Shot) (Claude); derive counts/dates from data; timestamp changelog rows (Claude); deep-link verification (Codex); System/Light/Dark with persisted override (Codex, Cursor, Claude).

Explicitly deferred — attributed, not consensus

• No framework migration; no MCP before persistence (Cursor, Claude).
• Server timing: Claude — only at "agent must write unattended"; Codex — earlier: as soon as multi-device/agent writes are needed for first real use; Cursor — staged triggers, no DB "this quarter." These are three different clocks; §5 D2 puts the choice to AK.

5. Decisions for AK

Decision log — 2026-09-14, from AK's replies (supersedes the recommendations below where they differ): - D1 privacy: REVERSED — AK: "leave it, privacy not an issue." No strip. - D2 state home: agents read-only until akco.dev hosting exists; Muse is the main writing agent alongside AK. Leaning toward the small service (SQLite + API); revisions plan and MCP approved in principle. Open questions AK asked: setup difficulty, what it stores (all: rankings/metadata, tasks, page content), architecture (yes — e.g. SQLite on Dell XPS, hub reads/writes over HTTPS API). - D3 hosting: approved — rides along with standing up admin.akco.dev + mcp.akco.dev. - D4 unknowns: approved everything — all four sparsity mechanisms + commitment bounding, with a settings page exposing every fraction/multiplier/threshold for tweaking. - D5 time axis: windows approved; windows are not stages (clarified); rejects one-venture WIP — parallelism via agents is the point; wants common-denominator tasks automated ("Ansible/Terraform/Kubernetes for side hustles" = scaffolding need, not a venture); dates must be soft/self-healing. - D6 roster: approved — Zeitgeist Apparel added as idea-stage, low rank alongside TikTok Shop Videos and GraceOverflow. - D7 energy: multiplier + calibrated (gate rejected as too binary); energy_now vs energy_week6 approved; persistent vs temporary dread approved; dread-extinguishing actions to be first-class (completing one re-rates energy); "tangible-win pull" phenomenon to be documented.

Each states the question, the options the reviewers actually offered, and my recommendation. Nothing here is decided. Sequencing per Cursor's question — which P0 first: privacy, persistence, or formula? — I recommend privacy → persistence → formula: privacy is a one-evening fix with no downside, and a score you can't save isn't a score.

D1. Privacy — strip clinical neurotype from the public Handoff page? (urgent) Claude and Cursor both flag it: the public share publishes AuDHD/OCD while the same page states the ground rule "never expose his neurotype clinically to third parties," and Theology for Sleep hypothesizes an ADHD audience. Recommend: yes, today — keep operator notes in a private file; the page keeps only design consequences (zero-cost capture, visible wins, phone-first).

D2. Where does durable state live? (the big one) - (a) Vault-first (Claude): venture profiles as markdown+frontmatter in 05__work/02__ventures/; Obsidian Base as the phone board; VentureHub regenerated from profiles. Zero new infrastructure; git is the writer log. Server only if an agent must write unattended. - (b) Small service now (Codex): SQLite on the chosen host + one API + revisions/If-Match; browsers and agents share it. Trigger: as soon as multi-device/agent writes are needed for first real use — which may be now, not later. - (c) Minimal bridge (Cursor): ideas.json/tasks.json as SSOT + localStorage + one-tap export now; graduate at the second-writer trigger. The question that decides it: for the first real use, do agents or a second device need to write VentureHub state, or only read it? (Read access is a hosting fix — Instinct's one-line move — not a store trigger.) My lean: (a) if a human stays in every write loop; (b) if AK wants agents updating tasks/state on their own. Tell me which loop you want and I'll build that one.

D3. Where is the daily driver hosted? All four: not the muse.ai share (preview only; fetch-filtered for agents; deploy flake). Candidates: Cloudflare Pages (Claude — matches FMS14-5), your machine / a VPS over Tailscale (your earlier leaning; Instinct: any static host works — it's one file). Claude proposes deciding by 30 Sep ("tie up loose ends" month). Recommend: decide alongside D2; the host follows the state home.

D4. Unknowns remedy — penalize sparsity in the score, or keep Unknown null and bound the commitment? (genuinely contested — T5) - (a) Penalize sparsity (Claude/Cursor/Instinct): pessimistic imputation, coverage scaling, rank on the lower bound; research raises the floor. - (b) Bound commitment (Codex): Unknown stays null; show "needs estimate" or scenario bounds; size the bet (cheapest test, max downside, decision it could change) instead of discounting the value. My synthesis (mine, not theirs): minimum coverage for rank-eligibility — the one mechanism nobody opposed — plus scenario bounds rather than silent imputation. That keeps the board honest without manufacturing precision. But this is a real choice; pick the philosophy you want to live with.

D5. Time axis — place each venture in a window? Claude's proposal: Now / Next / Later / Shelved with dates + one-venture WIP limit, reconciling the hub's "Rock now: AccelerateBooks" with the 30 Aug plan (October = Nile + Zeitgeist) and FMS14-5 ("AccelerateBooks not revived this cycle"). Recommend: yes — this is the 10-minute ruling, and it resolves the single biggest contradiction any reviewer found between the hub and your own calendar.

D6. Roster — add the 7 missing vault ventures? Zeitgeist Apparel (in your October plan), FutureMD, AgoraMedia, AcquireSoft, AllTheBest, doctrine-by-design, FortySeven Media. Recommend: yes for Zeitgeist Apparel at minimum (it's in next month's plan); the shelved/historical ones as stubs — but only into whichever system D2 designates, not hand-duplicated into the HTML.

D7. Energy model — gate or multiplier? Cursor: implement (enjoyment/5)^1.5. Claude: energy floor as a pass/fail gate + energy_now vs energy_week6 (novelty inflates capture-time ratings). Codex: don't auto-multiply uncalibrated ratings — persistent unwillingness can reroute a venture, temporary reluctance must not erase its value; define what the rating means first. Recommend: Claude's gate with Codex's calibration requirement — and note Claude's journal citations (19 Dec 2025) already give the "ground it in your statements" evidence Codex asked for. Dread can kill or reroute, never merely discount 8%.

Appendix A — reviewer précis

Claude (adversarial, vault-grounded). The only reviewer with AK's vault. Verdict: decision architecture sound; three findings dominate — (1) the hub is a source with nothing durable behind it, (2) no time axis so it contradicts AK's calendar, (3) the interactive layer forgets. Proposes vault-first canonical-home + generated surfaces. Most citations per claim; only reviewer to reconcile the roster against the vault.

Codex/ChatGPT (live inspection, decision-theoretic). Verdict: concept deserves to continue; implementation not yet trustworthy as an operating system. Sharpest on scoring semantics (multiplication needs meaning; unknowns should bound commitment, not erase value), the action-not-venture decision unit, and concurrent-write correctness (revisions, idempotence). Found the one concrete code defect (innerHTML injection). Six-outcome acceptance bar for re-review.

Instinct (second pass, operational). Verdict: the first review's fixes landed and several exceed what was asked. Remaining: persistence, the coverage soft spot, and — uniquely, twice-tested — the share host's fetch filter that bounces cURL-style agent reads. Smallest diff, highest signal per word. (One honest flag: its "1-of-12 → 100 is gone" claim conflicts with three same-build reproductions — see C2.)

Cursor (live share, law-driven). Verdict: the cockpit is the right product; the formula doesn't implement the rationale and the public Handoff breaks the page's own ground rule. Most quotable failure forecast; the clearest worked scoring examples; the only reviewer to price the dark-mode inversion and 16px checkboxes; trigger-based architecture table ("do not jump to a Next app, a database, or MCP this quarter").

Appendix B — claim cross-reference

← Back to MuseHub