Most expensive
Silent behaviour change
A rewrite alters an edge case nobody knew existed. It is discovered in production months later, usually by a customer or a regulator.
Never change what you do not understand.
ATLAS models your legacy system from its own evidence, then carries the work through to merged, verified code.
SurfaceAuto
StructureAuto
SemanticsAsked
RuntimeObserved
Reads what your system already produces
Why modernization fails
Three failure modes, in order of what they cost.
Costly
Analysis paralysis
Twelve weeks of discovery produces a slide deck. No code moves, and the assessment goes stale before execution starts.
Corrosive
Plausible fiction
Tools generate confident documentation, plans and code from inference. Everything reads well. Some of it is wrong, and there is no way to tell which parts.
Static analysis maps structure but never observes runtime. Observability sees runtime but not intent. Code assistants generate changes with no model of the system at all. Nothing connects the four links as one chain. That chain is the product.
One continuous chain, with nothing inferred between the links.
Evidence
Source, history, schema, live data, servers, logs, metrics and traces, pulled read-only and redacted at the boundary.
Understanding
One typed graph where every node carries its stratum, its sources and its confidence. Gaps become questions, not guesses.
Planned work
Tasks generated from the graph, each shipped as a self-contained envelope a person or an agent can execute without research.
Verified change
Production transactions replayed through old and new. Every difference ruled on by a named person before merge.
What the platform does
Six capabilities, each writing back to the same model.
F1
Ingest everything the system knows about itself
Ten connector families, read-only by default. No customer has all of them, so the platform states precisely what each missing source costs.
Evidence and the modelF2
Reconcile it into one typed graph
Static structure joined to runtime observation. Reachability, hot paths, data reality, failure semantics.
Read moreF3
Ask instead of inferring
Below the confidence floor, the analyzer raises a question rather than a claim.
ClarificationF4 to F7
Ship work as context envelopes
A budgeted graph traversal, not a file dump. The same envelope drives a person, an agent, or an external assistant.
context_slice:
callers: [QuoteServlet, RenewalBatch]
rules: [RATE-0042, RATE-0044]
runtime: branch 218 true in 0.4% of calls
F8
Prove the behaviour did not change
Characterization tests pinned to live paths, then differential replay of real captured transactions through both implementations.
VerificationF9
Record the forks as binding constraints
Choosing an option rewrites every affected envelope and binds every agent dispatched afterwards. Not a chat message.
DecisionsOne gap. Three vocabularies.
When evidence runs out, the platform does not fill the space with plausibility. It writes a question, routes it to whoever actually knows, and renders it in that person's language.
Choosing beats typing. Options are synthesised from competing readings, observed production values and precedent. Free text stays available as an escape, never as the default path.
Branch at PremiumCalculator:218 fires in 0.4% of calls and was last hit six days ago. Live requirement or never-removed fallback?
trace:otel, 14d window, 47 executions ยท vcs:commit a4f2e91
None of these
I do not know
Ask someone else
Answering this unblocks 23 tasks and decides whether 2 test cases are treated as correct behaviour.
Policies written before January 2014 in New York get a different surcharge rate. Do you remember why that cutover exists?
Same gap, rendered for the person most likely to know. Routing followed commit history and defect history.
None of these
I do not know
Ask someone else
Answering this confirms rule RATE-0042 and releases the surcharge work for the whole quarter.
Should a New York policy issued in 2013 still be charged the old surcharge rate when it renews today?
No code, no identifiers, no jargon. An underwriter can rule on this without opening a repository.
None of these
I do not know
Ask someone else
Your answer becomes a recorded rule. Every task and agent downstream inherits it as a constraint.
Differences are grouped by cause, and someone has to sign for each one.
Captured production transactions run through both implementations. The output is not a wall of failing cases. It is a short list of reasons, each needing a ruling before the merge is allowed.
| Cause | Count | Example | Ruling |
|---|---|---|---|
| Rounding on half-cent | 201 | 1284.125 to .12 against .13 | Accepted Matches the filed rate manual |
| Credit ordering | 21 | Credit applied before the surcharge rather than after | Accepted Signed off against RATE-0042 |
| New York pre-2014 rate | 2 | 0.0725 against 0.0680 | Merge blocked Unexplained |
0
Unexplained differences permitted to reach a merge. This is a gate, not advice.
Live paths
Coverage is measured against paths runtime evidence proves are used. Coverage of dead code is worthless.
15:1
Target clustering ratio, so one answer retires many questions instead of one.
How it lands
Read-only first. Nothing writes to your systems until the model is trusted.
Assess
F1, F2, F3
Ingestion, the system model and clarification for one language pair. No code is written and no write access is needed.
System map, risk register, confirmed rules, a populated task board
Prove
F8 test generation
Characterization tests for untested legacy code. Still purely additive, so nothing changes behaviour.
Behaviour pinned before anything is allowed to move
Execute
F5, F6, F7
Low-risk task types only: dependency upgrades, dead code removal, test backfill. Agent-authored pull requests, human review.
Differential replay before every merge
Decide
F9
Migration decisions with constraint propagation. Only valuable once the model underneath has earned trust.
Architecture decision records that bind downstream work
Transform
Full scope
Agent execution on service extraction, schema migration and framework replacement.
The parts everyone wanted on day one, attempted last
Evidence or a question. Never an inference.
Every fact in the model traces to source code, schema, a runtime observation, or a named human's answer. Where the platform cannot establish something, it blocks and asks. It recommends, explains and refuses. It never decides.
Start with an assessment that cannot lie to you.
Read-only, on your infrastructure, one language pair. You get a system map and a task board, or a precise account of what was missing and why.