Project Intelligence model
Project Intelligence is Snipara's outcome-weighted judgment layer for AI agent work. Before an agent edits, it should know what changed, why, what it impacts, what should happen next, and whether evidence says proceed, review, or stop.
Project Intelligence formula
Context
Source-backed project facts, code graph context, workflow state, and freshness metadata.
Decisions
Reviewed rationale, active constraints, confidence, authority status, and stale warnings.
Outcomes
Guard, review, test, workflow, and served-judgment evidence that can later confirm, refute, or stay unevaluated.
Judgment
Advisory next actions with confidence, evidence, counter-evidence, caveats, and calibration limits.
The product is intentionally broader than memory. Snipara gives humans, Claude Code, Codex, Cursor, CI, and custom agents one governed project layer to inspect before acting.
Product boundary
Snipara is not the reasoning model. Claude Code, Codex, Cursor, ChatGPT, and customer agents still reason and execute. Snipara supplies the project-owned structure: context, reviewed memory, source authority, impact, workflow state, coordination, and verification guidance.
Homepage relationship
The homepage leads with what is live today: continuity, coordination, governance, proof, and collaboration. Project Intelligence is the category and depth model behind those surfaces. It explains how Snipara answers the five agent-work questions while keeping emerging behavior and product limits visible.
The five questions
Memory alone is not enough for serious agent work. A useful project layer must answer these questions in a way that a human can inspect and a model can act on.
What changed?
What Changed For Me, Team Sync handoffs, PR Answer Packs, resume context, recent files, workflow journals, and Project Intelligence Brief V2 summarize repository movement before the next session starts.
Why?
Reviewed decisions, ProjectDecision drafts, Decision Graph anchors, issue links, PR context, handoff notes, Why Capture, Answer Capture, and Knowledge Compilation keep rationale attached to source-backed project evidence.
What does it impact?
Code graph context, impact plans, affected symbols, related tests, routes, config facts, local workflow impact gates, and Historical Impact give agents a blast-radius view before editing.
What should happen next?
Start Work Briefs, verification plans, handoff next steps, recommended checks, workflow phase state, Brief V2 next-step answers, outcome-weighted reasoning, and Judgment V0 turn context into a reviewable execution path.
Can I safely proceed?
Confidence Engine profiles, Outcome-weighted Judgment V0, Safe Parallel Coding, collaboration guards, resource leases, stale warnings, and Decision Consistency checks expose weak evidence, overlap, explicit contradictions, or shadow candidates before risky actions.
What is available today
The current product already answers parts of the five-question model through concrete surfaces. Treat these as operational entry points and governed agent surfaces, not separate products.
Measured impact
Public proof should stay measurable and narrow. The current replay evidence reports less rediscovery before planning, stronger surface coverage, and lower search noise. It does not claim universal time saved, regression reduction, or ROI until workflow traces capture those measurements directly.
Inspect the controlled replayRepository continuity
Start Work Briefs, What Changed For Me, PR Answer Packs, handoffs, phase commits, and resume context keep agent work from restarting at zero.
Context authority
Reviewed memory, provenance, source URLs, freshness, confidence, stale warnings, and validation state make context inspectable before agents trust it.
Decision Graph
Project decisions, reviewed memory, context graph edges, and outcome signals are projected into inspectable claims with anchors, provenance, and explicit contradiction edges.
Confidence Engine
Project Intelligence claims carry score, band, state, reason codes, evidence, and caveats so weak, stale, conflicting, or low-sample inputs stay visible.
Code impact
Snipara code tools and companion impact commands give agents affected files, related tests, risk signals, and verification hints before implementation.
Why and Answer Capture
Confirmed rationale and explicitly accepted LLM answers can become pending reviewed-memory candidates and review-pending ProjectDecision drafts. They are not raw transcript storage and not auto-approval.
Knowledge Compilation
Accepted answers are redacted, deduplicated, assigned idempotency keys, and filed as provenance-rich memory candidates instead of raw transcript truth.
Decision Consistency
Approved decision memories can participate in collaboration guard verdicts. Explicit Decision Graph CONTRADICTS evidence can block; shadow contradiction candidates remain non-blocking review signals.
Historical Impact
Changed files, symbols, routes, and diff summaries can be checked against outcome signals and reviewed memory for advisory breakage or regression hints.
Outcome-weighted reasoning
Later workflow evidence is summarized with sample gates and non-causal wording, so outcomes inform review without becoming automatic attribution.
Outcome signals
Guard, review, test, and workflow events can produce passive Outcome Loop metrics. They improve operator visibility today; ranking changes remain gated.
Outcome reliability curve
Project Intelligence calibration now exposes reliabilityCurve with confidence-band bins and reason-code bins for observed precision, confirmed/refuted candidates, and unevaluated ambiguous links. It is observability and promotion evidence, not causal proof.
Safe parallel coding
Team Sync, presence, leases, guard profiles, GitHub checks, and local companion commands reduce conflicting work across humans and agents.
Project Intelligence brief
Project Intelligence Brief V2 composes Decision Graph, Confidence Engine, outcome reasoning, and fuel status into a core response, with heavier sections exposed through include flags.
Decision drift and shadow signals
Decision drift and shadow contradiction detection now share one signal surface with roles for acting advisories and shadow measurement. Rejected contradictions can carry both roles as a single finding.
Outcome-weighted Judgment V0
Project Intelligence Brief V2 now includes an experimental advisory layer with recommendation confidence, evidence summary, counter-evidence, competing explanations, and Project Health signals. It is sample-gated and not autonomous reasoning.
What is still emerging
These are the remaining edges of the model. Snipara exposes shipped behavior conservatively: candidates stay reviewable, historical hints stay advisory, and outcome signals are used as evidence rather than marketing magic.
Broader automatic why extraction
Why Capture exists for confirmed source material. Wider extraction from every commit, pull request, phase commit, and handoff is still being expanded and reviewed.
Outcome-aware ranking
Outcome signals and passive metrics exist. Using those signals to rerank context or change Memory evidence scores remains gated until there is enough validated history.
Advisor-grade confidence
Judgment V0 exposes confidence shape and reliability curves today. Advisor-grade confidence should wait until enough evaluated outcome and reason-code samples exist across projects.
Persisted project pages
Readable briefs can be generated from atomic memory, source refs, and outcome evidence, but persisted synthesized pages are not canonical knowledge yet.
Project Health dashboard
Richer dashboards for knowledge health, decision drift, unverified assumptions, coordination risk, context freshness, and long-history outcome evidence are still emerging.
How to use it in an agent workflow
- Connect the repository with create-snipara or Hosted MCP.
- Start risky or resumed work with snipara-companion brief or a hosted MCP context query.
- File confirmed rationale or accepted answers as reviewed-memory candidates instead of saving whole transcripts, and review captured ProjectDecision drafts before relying on them as current decisions.
- Use code impact, Historical Impact, or symbol cards before changing routes, services, jobs, auth, billing, schema, deployment, or shared behavior.
- Commit phases, hand off the session, and let the next agent resume from project-owned state instead of transcript memory.
What Snipara does not claim
- Snipara does not replace Claude Code, Codex, Cursor, ChatGPT, or your own agent runtime.
- Snipara does not treat every chat transcript as project memory.
- Snipara does not claim every project judgment is fully autonomous today.
- Judgment V0 confidence is not a calibrated probability or advisor-grade certainty; it is an inspectable, sample-gated advisory signal.
- Project Intelligence and outcome-weighted judgment are the public category; memory is a component capability, not the category.
- Snipara keeps internal feedback-loop mechanics out of public workflow contracts unless they become stable product surfaces.
- Project Intelligence briefs are compiled evidence views, not canonical project truth.
- Reliability curves summarize observed outcomes for calibration; they are not causal proof.
- Shadow contradiction candidates are advisory measurement signals; explicit CONTRADICTS evidence is the blocking safety path.
- A cold-start fuel status means matched decision or outcome evidence is thin, not that the system is unavailable.