Scan, deterministically
TraceMap produces the facts with no model in the loop, so the evidence is reproducible and identical run to run.
Give the model something it can cite
Ground your AI reviewer on facts, not vibes.
An LLM asked “what does this change touch?” will answer confidently, and sometimes wrongly. TraceMap already emits the deterministic layer underneath that question: rule-backed facts with a rule ID, evidence tier, coverage label, file span, commit SHA, and explicit gaps. Hand those artifacts to your reviewer or agent as the evidence it must cite — so its claims trace back to a fact instead of a guess.
Public claim level: concept. No public conclusion without evidence. The language model is yours and lives outside TraceMap. The TraceMap scanner and reducer use no LLM calls, embeddings, vector databases, or prompt-based classification. This page describes using existing deterministic artifacts as grounding; it does not describe a shipped TraceMap AI feature, a hosted model, or any judgement of a model’s output.
The gap
An ungrounded model invents impact; a grounded one has to point at a fact.
| Review question | Ungrounded model alone | Model grounded on TraceMap facts | What TraceMap supplies | Boundary |
|---|---|---|---|---|
| What does this contract change touch? | A plausible list assembled from training priors and the diff, with no traceable source. | A list where each item points to a fact ID, rule ID, evidence tier, and file span. | Reducer output and index facts for the changed type, property, endpoint, or surface. | Static references are review input, not proof the change breaks at runtime. |
| How confident should we be in each hit? | A confidence number the model made up. | The evidence tier the fact already carries: semantic, structural, syntax/textual, or unknown. | Per-fact tier and coverage label from the scan. | Tiers describe evidence strength, not runtime certainty or release safety. |
| What did we miss? | Silence — the model rarely volunteers its blind spots. | The explicit AnalysisGap rows and reduced-coverage labels, surfaced as open questions. |
Deterministic gap facts where dynamic dispatch, reflection, or partial coverage stop the analysis. | A gap is an honest unknown to route to an owner, not an all-clear. |
What you hand the model
Feed it artifacts it can read, not prose it can only paraphrase.
facts.ndjsonOne deterministic fact per line with rule ID, evidence tier, file span, and scan identity — a citation source the model can quote by ID.index.sqliteThe queryable fact store, so an agent can look up callers, surfaces, and dependency edges instead of imagining them.export (json / mermaid)Portable relationship and symbol shapes an agent can walk to answer “what reaches this?” from evidence.reduce / release-review outputContract-delta classifications already labeled DefiniteImpact, NeedsReview, or NoEvidence with the coverage that produced them.rule catalogThe documented meaning and limitation of every rule ID, so a cited fact carries its own boundary.How grounding works
TraceMap is the floor; your model is the ceiling.
Require citation
Prompt your reviewer or agent to answer only from the supplied facts and to attach the rule ID, tier, and span for every claim.
Keep the gaps visible
Pass the AnalysisGap and reduced-coverage rows through as open questions so the model reports unknowns instead of papering over them.
Where the boundary stays
Grounding constrains the model; it does not certify it.
The core stays deterministicNo LLM, embedding, vector database, or prompt classification runs in the scanner or reducer. The facts do not depend on a model and never will.
The model stays outside and stays yoursWhatever reviewer, agent, or IDE assistant consumes the facts is operated by you, under your controls and data policies. TraceMap neither ships nor hosts it.
Grounding is not a correctness guaranteeFeeding facts reduces invented impact, but a model can still misread evidence. Its output remains a draft for human review, not an approved conclusion.
Non-claims
Grounded evidence does not become an approved answer.
- TraceMap does not perform AI impact analysis, LLM analysis, prompt-based classification, embedding search, or vector database analysis in the scanner or reducer.
- TraceMap does not prove runtime behavior, production traffic, endpoint reachability, release safety, operational safety, or complete product coverage; supplying facts to a model changes none of that.
- TraceMap does not validate, score, or approve any model’s output; a grounded answer is still a draft for human review.
- TraceMap does not replace code review, tests, service-owner judgment, release approval, or governance.