SIFT v0.11.2

SIFT v0.11.2

Version 0.11.2 fixes slow-model timeouts in the Create-to-Build workflow, especially when using Claude Opus 4.6 through OpenRouter.

Slow-model recovery

• Gives each Idea Forge pass its own bounded time budget and reserves time for the final hypothesis pass.
• Uses low-effort reasoning for Claude 4.6 automated workflow calls without applying Claude-specific controls to other models.
• Replaces the full-size standard-generator retry with one smaller recovery request using the same selected model.
• Reports whether Idea Forge, standard generation, or evaluation actually timed out.

Resume instead of restart

• Checkpoints completed idea generation and public research for the current session.
• A retry after a later evaluation timeout resumes from the completed work instead of regenerating the idea slate.
• Clears checkpoints whenever the project, prompt, or model configuration changes.

Regression coverage

• Adds timeout-chain, stage-budget, checkpoint, and cross-model compatibility tests.
• Validates the embedded Python worker and packaged Windows application.

Windows and macOS packages are currently unsigned. Verify downloads with SHA256SUMS.txt from the release before bypassing operating-system warnings.

SIFT v0.11.1

SIFT v0.11.1

Version 0.11.1 is a focused reliability hotfix for the Create-to-Build Idea Forge workflow.

Idea Forge contract repair

• Accepts and validates the bounded usage metadata produced by the embedded Python worker instead of rejecting an otherwise valid idea slate.
• Aligns private-profile generation so every returned idea contains a numeric personal-fit estimate, while neutral generation continues to require personalFit: null.
• Makes the Python stage parser tolerate one safely extractable JSON object and harmless extra response keys without weakening evidence or protocol-consistency checks.

One-shot recovery repair

• Distinguishes an Idea Forge failure from a later standard-generator failure, so the final message identifies the step that actually needs attention.
• Guards progress updates against cancelled or superseded runs and resets the busy state when connection setup is interrupted.
• Describes the standard generator as a recovery attempt instead of announcing completion before that attempt finishes.

Regression coverage

• Adds real-shaped worker-result tests including usage metadata and private-profile scoring.
• Adds malformed-output parser tests and fallback/run-state regression checks.

Windows and macOS packages are currently unsigned. Verify downloads with SHA256SUMS.txt from the release before bypassing operating-system warnings.

SIFT v0.11.0

SIFT v0.11.0

Version 0.11.0 delivers the complete Create-to-Build handoff, clearer AI failure recovery, and a quieter SIFT interface.

Create to build

• One primary action now generates and critiques ideas, selects the strongest candidate, researches public context when available, runs the deterministic thesis screen, calculates the decision, and prepares a route-aware build brief.
• The result offers Start building or Not now. Signing, spending, leasing, and deployment still require explicit approval.
• Xahau, Evernode, hybrid, and conventional routes receive distinct build handoffs without treating public research as customer validation.

More reliable AI runs

• Recoverable Python worker stops, timeouts, malformed model output, and schema failures now use one bounded fallback generator instead of ending with a generic stopped message.
• Authentication, credit, rate-limit, and model-rejection errors retain safe error codes and show a specific corrective message.
• The Python intelligence worker remains advisory; TypeScript validation and deterministic scoring remain authoritative.

Simpler interface and packaging

• Appearance and the single Clear local data action now live only in Settings.
• Removed the saved-status label, sidebar data-controls drawer, redundant landing disclaimer, and the extra start menu.
• Product branding now appears simply as SIFT while Xahau and Evernode remain available as functional technology routes.
• The portable Windows download is now SIFT.exe; the versioned Windows installer remains available for normal installation.
• Removed unused ChatGPT sign-in and database starter scaffolding from the source repository.

Distribution notes

Windows and macOS packages are currently unsigned. Verify downloads with SHA256SUMS.txt from the release before bypassing operating-system warnings.

SIFT v0.9.0

SIFT v0.9.0

Version 0.9.0 replaces the one-pass idea prompt with a contract-first,
three-pass Idea Forge while preserving SIFT's deterministic evidence and
decision boundaries.

Simpler product journey

• Home now presents one primary path: connect a model, generate and screen a
  fresh idea, then start real-world validation only when the thesis earns it.
• The full one-click flow reports clear progress and keeps the final thesis
  decision available after the run instead of forcing users back through the
  workflow.
• Advanced review controls and the 51-claim rubric remain available through
  progressive disclosure without dominating the default experience.
• Model setup, connection failures, recovery actions, evidence boundaries, and
  build-tool readiness are explained in plain language at the point of use.
• Updated SIFT tornado artwork is used consistently in the app header and
  workflow status surfaces in both light and dark themes.

Better idea generation

• Python now frames opportunity mechanisms before proposing solutions,
  generates a wider raw slate, and independently critiques and rewrites the
  strongest candidates before returning them.
• Final ideas preserve the actor, trigger, current workflow, material
  consequence, buyer, why-now hypothesis, distribution wedge, adoption
  friction, protocol role, conventional counterfactual, failure case, critical
  assumption, and structured 14-day experiment.
• Every experiment includes an explicit target, artifact, metric, pass
  threshold, and kill threshold.
• Xahau and Evernode prompts use a bounded capability manifest. Both requires
  separate protocol responsibilities, while Neither yet remains available
  when conventional software is the stronger counterfactual.
• The model may not invent customer evidence, traction, payments, production
  behavior, audits, market statistics, or changing protocol facts.

Local quality contract

• TypeScript measures thesis construction independently of model-provided
  exploration estimates.
• Generic actors, missing triggers, vague alternatives, marketing language,
  unsupported validation claims, incomplete experiments, contradictory routes,
  and unjustified protocol usage are rejected or flagged.
• Near-duplicate candidates are removed before the requested slate is returned.
• Eligibility is checked before weighted profile priority, so a vague idea
  cannot win merely by assigning itself high scores.
• This local quality value is not evidence, a market score, or a probability of
  success.

Architecture and compatibility

• The user-facing one-click flow remains one action; its internal gates are now
  explicit and ordered.
• Windows and macOS packages embed the supervised Python intelligence worker;
  users do not need to install Python or Docker.
• The Python worker remains isolated, bounded, cancellable, and unable to write
  project state or alter SIFT's deterministic score.
• TypeScript still owns project mutation, evidence integrity, official scoring,
  stage gates, and decisions.
• Existing projects import safely. Older ideas receive blank values for the new
  fields until edited or regenerated.
• The deterministic 51-claim venture scoring engine and evidence policies are
  unchanged.

SIFT v0.7.0

SIFT v0.7.0

Version 0.7.0 adds an honest one-click path from a blank page to a deterministic thesis decision while keeping real-world validation separate.

Highlights

• New Generate & Screen workflow: SIFT generates four fresh ideas, selects the strongest profile match, screens the winner, saves the complete result atomically, and provides one-step undo.
• New deterministic thesis decisions: Advance to validation, Revise & rescreen, Park this idea, or Screen incomplete.
• Newly generated ideas are no longer penalized for having zero customer interviews, payments, production usage, or audits. Those records cannot exist before the idea does.
• The thesis screen evaluates hypothesis specificity, coherence, falsifiability, and the G1, G2, and G7 controls. Evidence multipliers, validation-stage caps, and evidence floors do not affect it.
• Deterministic engine v3-powershell-parity/1.1.0 records the new stage-aware scoring semantics while retaining the existing canonical rubric manifest.
• OpenRouter can add cited public market, competitor, regulatory, and protocol context. Public context can inform the thesis assessment but is never written into the direct-validation ledger.
• Ollama, LM Studio, and other OpenAI-compatible providers can complete Generate & Screen without public web context.
• Start validation opens Discovery with an intentionally empty evidence ledger and resets AI-owned thesis gates. Strict evidence requirements begin from Discovery onward.
• Switching to another idea starts a clean thesis review so claims, gates, interviews, payments, and other evidence cannot leak across ideas.
• Thesis mode now hides evidence grades, production-stage controls, and evidence tools until validation explicitly begins.
• Updated high-resolution tornado artwork is used consistently for the Windows, macOS, and web application icons.

What one click means

Generate & Screen automates everything that can honestly happen immediately: idea generation, selection, optional public context, hypothesis assessment, and the decision about whether an idea deserves validation.

It does not fabricate future customer validation. Interviews, direct tests, commitments, payments, production behavior, and audits must come from later real-world work.

Upgrading

• Installing SIFT 0.7.0 over 0.6.0 preserves the existing installation identity, local projects, and protected model settings.
• The desktop app remains local-first and requires no SIFT or ChatGPT account.
• Windows and macOS packages remain unsigned. Verify downloads against SHA256SUMS.txt on the GitHub release page.

Important boundaries

• AI proposes and explains thesis inputs; SIFT's locked local formula calculates the thesis decision.
• Public research remains desk context and cannot substitute for direct customer or product evidence.
• SIFT does not sign transactions, accept wallet secrets, spend funds, acquire leases, or deploy contracts automatically.

SIFT v0.6.0

SIFT v0.6.0

Version 0.6.0 turns SIFT into a fuller local Xahau/Evernode venture workspace while preserving the deterministic review boundary.

Highlights

• New Research & Run flow: generate or select an idea, draft a shadow evaluation, search public sources through OpenRouter, validate provider-returned citation excerpts, and calculate a provisional outcome with one consolidated source approval.
• Public web findings are permanently capped at DeskResearch / E1; AI cannot invent customer interviews, commitments, payments, audits, production behavior, or reviewer verification.
• Citation-only provenance: model-authored URLs are ignored, excerpts must match provider annotation content exactly, and the ledger retains source URL, title, retrieval time, excerpt hash, model, and search provider.
• Atomic approval: the live project remains unchanged until the final packet confirmation; malformed or stale sources fail the entire batch without partial writes.
• New guarded Build workspace for Evernode MCP, Xahau MCP, XAHC, and XAHC Prover with fixed, unsigned, allowlisted operations.
• Complete SIFT visual identity, updated tornado assets, polished light/dark themes, responsive layouts, and friendlier OpenRouter model search.
• Windows installer and portable packages use the SIFT product name. macOS Intel and Apple-silicon packages are built through the release workflow.

Upgrading to SIFT

• Existing desktop projects and protected model settings remain available. SIFT keeps the established installer, bundle, and user-data identities internally so an upgrade does not create a second Windows installation or abandon encrypted credentials.
• Browser project data and unfinished personality drafts move automatically to SIFT storage keys on first use.
• On macOS, drag SIFT.app into Applications. After confirming the new app opens and your project is present, remove the pre-SIFT app bundle if it is still shown separately. Do not delete its Application Support data.
• The release files, shortcuts, window titles, executable metadata, and all current documentation now use the SIFT name.

Important boundaries

• Research & Run currently requires a user-supplied OpenRouter API key and credits. Model-token and web-search charges apply.
• OpenRouter web search is a beta provider feature. SIFT fails closed when attributable citation excerpts are unavailable.
• A researched outcome is still provisional. E1 public research does not replace direct customer evidence, human gate decisions, independent security review, or a final investment/launch decision.
• SIFT does not sign or submit transactions, accept wallet secrets, acquire leases, spend funds, or deploy contracts automatically.

SIFT v0.5.0

SIFT v0.5.0

Version 0.5.0 adds an optional research-based Big Five profile and an AI one-click outcome preview while preserving the locked, local review engine.

What changed

• Added the public-domain Johnson (2014) IPIP-NEO-120 as an optional profile-building path. The 120-item assessment typically takes 15–20 minutes and reports five broad domains plus 30 facets.
• Assessment results use 0–100 response-scale positions, not population percentiles, and are labeled for self-reflection and idea personalization—not diagnosis, hiring, or other consequential decisions.
• Added AI one-click preview. A connected model can generate ideas when needed and propose values for unanswered claims and unresolved gates; a local profile-priority formula selects among saved exploration estimates when the user has not already chosen an idea.
• The locked local engine calculates those proposals in an isolated review copy and displays a clearly provisional outcome. The model does not calculate or control the deterministic formula.
• Kept Guided review as an alternative for users who want to move through and approve each checkpoint themselves.
• Fixed the Quick Run progress connector so lines no longer cross step labels or markers.

Integrity and privacy

• A one-click preview never mutates the live review, overwrites existing human decisions, fabricates or verifies evidence, or turns an evidence-free result into an official decision.
• If the review route is unresolved, the preview may derive it from the chosen idea’s declared Xahau/Evernode route. The preview shows that provenance, and the derived route exists only in the isolated copy.
• The scoring rubric, 51 claims, weights, caps, floors, evidence validator, and gates are unchanged. Personality results affect idea personalization, not deterministic review scores.
• Unfinished questionnaire answers exist only in the current session's sessionStorage. Raw answers are not saved to the project, included in exports, or sent to an LLM as profile context.
• Only the derived assessment result is saved locally after the user applies it. Exact derived scores are included in connected-model prompts only when the user explicitly opts in; a cloud provider receives them only under that opt-in.
• Cloud Quick Run confirms before sending its operation-specific context. Ollama and LM Studio continue to run on the local machine.

Downloads

• Windows 64-bit installer and portable executable
• macOS 12+ DMG and ZIP for Apple silicon (arm64)
• macOS 12+ DMG and ZIP for Intel (x64)
• SHA256SUMS.txt for release verification

The Windows and macOS packages are currently unsigned. Verify the checksum before opening them. Windows may show an Unknown publisher warning, and macOS may require right-clicking the app and choosing Open.

SIFT v0.4.1

SIFT v0.4.1

Version 0.4.1 adds an optional guided Quick Run and fixes AI connection state after local data is cleared.

What changed

• Added Quick Run to the welcome screen and Home. One button starts a guided path through idea generation, evaluation, evidence, gates, and the deterministic decision.
• Quick Run generates four editable hypotheses when needed, then pauses for the user to choose the direction.
• The connected model can draft unanswered merits, organize source text, and refresh gate recommendations against the current evidence.
• Gate-only refreshes cannot rewrite existing claim ratings.
• Already-reviewed projects resume at the Evidence checkpoint instead of skipping it.
• Added Clear project, which keeps the saved AI connection, and Clear everything, which also forgets the provider, model, and protected API key.
• Fixed provider switching after a clear or reload. Returning to a previously saved provider restores its saved model and recognizes its protected key only at the exact saved endpoint.
• Late model loads, searches, saves, and Quick Run responses can no longer overwrite a newer user choice.

Human control and privacy

• Quick Run applies nothing automatically. Idea choice, merit ratings, evidence records, and gates still require explicit user approval.
• Evidence-free results remain provisional, and AI never calculates or chooses the deterministic outcome.
• Exact excerpt matching, evidence ceilings, reviewer verification, conflict disclosure, dates, contradiction handling, and all other validators remain enforced.
• Cloud-model steps disclose their data boundary and confirm before sending the selected idea, review notes, or stored evidence excerpts.
• A local-data clear immediately blocks connector saves and AI work until the protected configuration has been removed.
• The scoring rubric, 51 claims, weights, caps, floors, and gates are unchanged.

Downloads

• Windows 64-bit installer and portable executable
• macOS 12+ DMG and ZIP for Apple silicon (arm64)
• macOS 12+ DMG and ZIP for Intel (x64)
• SHA256SUMS.txt for release verification

The Windows and macOS packages are currently unsigned. Verify the checksum before opening them. Windows may show an Unknown publisher warning, and macOS may require right-clicking the app and choosing Open.

SIFT v0.4.0

SIFT v0.4.0

Version 0.4.0 makes the desktop app faster to understand, much easier to connect to OpenRouter, and able to assist with evaluation and evidence preparation while preserving the locked deterministic scoring engine.

What changed

• Added live OpenRouter model typeahead. Enter a name, version, or partial canonical ID—such as 4.8 or opus.4.8—and SIFT queries OpenRouter automatically, ranks the response, and shows the best matches in a keyboard-accessible dropdown.
• Added punctuation-tolerant model matching, so dots, hyphens, underscores, and slashes behave naturally without hard-coded model names.
• Simplified the primary workflow to Home → Ideas → Evaluate → Evidence → Decision.
• Moved personalization, AI connection, and import/export tools into a compact Settings & data drawer.
• Reworked the start screen, idea-generation controls, candidate cards, profile editor, and model setup around progressive disclosure and plain-language actions.
• Combined model saving and connection testing into one Save & connect action.
• Added an optional desktop Draft evaluation workflow for connected local or cloud models. It proposes merits and rationales for unanswered claims plus gate recommendations, but applies nothing until the user approves it. Gate proposals are applied one at a time.
• Added an optional desktop Organize evidence workflow. It converts pasted source material into reviewable artifact proposals and locally confirms that every proposed excerpt appears in the supplied text.
• Added one-click undo for approved AI-assisted evaluation and evidence changes.
• Added the new SIFT logo, application icon, and transparent mark to the web interface and native Windows/macOS build resources.
• Preserved manual model-ID entry and full-catalog browsing as advanced fallbacks.

Integrity and security

• The scoring rubric, 51 claims, caps, floors, and gates are unchanged.
• Evidence dated after the review cutoff is rejected and excluded, and expiry dates can no longer predate the observation.
• Invalid, duplicated, expired, future-dated, or incompletely reviewed evidence contributes zero. Typed caps and floors now require one artifact to satisfy both the required grade and evidence type.
• Reprocessing the same source reuses its source family, while duplicate excerpts are rejected instead of being counted as new evidence.
• AI output remains a staged draft. It never supplies rubric weights, totals, final scores, or advancement decisions, and it cannot change the deterministic calculator.
• The AI cannot verify its own evidence. E2 or higher requires an explicit human verification action, a reviewer name, and a conflict disclosure before approval.
• The full pasted evidence source is not persisted. Approved artifacts retain only the accepted excerpt, a source fingerprint, and model provenance for auditability.
• OpenRouter remains pinned to https://openrouter.ai/api/v1; typeahead text is URL-encoded and API keys remain header-only.
• Other remote compatible endpoints must use HTTPS, and saved API keys are bound to both their provider and exact endpoint.
• Credentials remain protected by the operating system and are excluded from projects, exports, and browser storage.
• Ollama and LM Studio keep prompts on the local model server. OpenRouter and other remote compatible providers receive the operation-specific idea, review, or pasted evidence context only when the user invokes that assistant.

Downloads

• Windows 64-bit installer and portable executable
• macOS 12+ DMG and ZIP for Apple silicon (arm64)
• macOS 12+ DMG and ZIP for Intel (x64)
• SHA256SUMS.txt for release verification

The Windows and macOS packages are currently unsigned. Verify the checksum before opening them. Windows may show an Unknown publisher warning, and macOS may require right-clicking the app and choosing Open.

SIFT v0.3.0

SIFT v0.3.0

SIFT now runs on macOS and makes large model catalogs much easier to navigate.

What changed

• Search models by friendly name, version, provider, or exact ID. Typing 4.8, for example, surfaces matching entries such as Anthropic Claude Opus 4.8 when the connected provider exposes them.
• Choose a result with the mouse or keyboard while retaining manual model-ID entry as a fallback.
• Discover up to 2,000 unique models instead of truncating the catalog at 500.
• Download native macOS packages for Apple silicon (arm64) and Intel (x64). Both architectures are built, inspected, and launch-smoke-tested on matching GitHub-hosted Mac hardware.
• Continue using the Windows x64 installer or portable executable.

Downloads

Computer	Recommended file	Alternative
Apple silicon Mac (M1 or newer)	Idea-Foundry-0.3.0-macOS-arm64.dmg	Idea-Foundry-0.3.0-macOS-arm64.zip
Intel Mac	Idea-Foundry-0.3.0-macOS-x64.dmg	Idea-Foundry-0.3.0-macOS-x64.zip
Windows x64	Idea-Foundry-Setup-0.3.0-x64.exe	Idea-Foundry-Portable-0.3.0-x64.exe

macOS 12 Monterey or newer is required. Open a DMG and drag SIFT into Applications. ZIP files are provided for manual deployment.

Signing notice

These first macOS packages are not yet signed or notarized. After verifying the download against SHA256SUMS.txt, macOS may require you to right-click SIFT, choose Open, and confirm, or approve it under System Settings → Privacy & Security. Do not disable Gatekeeper globally.

The Windows packages also remain unsigned and may show a Microsoft SmartScreen warning.

Privacy and model connections

• SIFT itself requires no account.
• OpenRouter requires the user's own account, API key, and credits.
• API keys remain encrypted by the operating system, bound to their selected provider, and excluded from project files, exports, and browser storage.
• OpenRouter stays pinned to https://openrouter.ai/api/v1.
• When OpenRouter is selected, the displayed prompt is sent to OpenRouter and the selected upstream model provider.