No magic. Just math.
Patent Pending System & Method for Deterministic Resolution (No AI Inference).
See the deterministic engine behind recall. Déjà matches repeat incidents using stable identity, tiered matching, and validation gates — then withholds anything that doesn't clear confidence thresholds.
The pipeline
Déjà turns messy production signals into a stable incident identity, then matches only when the evidence holds. An Integrity Filter rejects minified or garbage data before normalization to preserve data integrity.
Before normalization, an Integrity Filter rejects incidents with minified frames (e.g., a.js:1) or missing symbolication. This preserves data integrity by ensuring only hydrated stack traces proceed to matching. Rejected data is discarded immediately.
ingest:
sentry|datadog signals
github|gitlab context
normalize:
paths -> stable
env -> removed
line numbers -> removed
noisy fields -> removed
fingerprint:
sha256(normalized_signal)
match:
tier0 exact identity
tier1 bounded similarity
tier2 conservative fallback
validate:
file gate + rate gate + duration gate
compute confidence score
recall:
return PR + diff that shipped (or withhold)Déjà withholds matches that don't meet confidence gates. It returns fewer results by design.
Forensic Linkage
We do not guess based on similarity. We construct a cryptographic link between the runtime artifact (Stack Trace Hash) and the build artifact (Commit SHA). If the chain of custody is broken by minification or missing history, we refuse to link.
Generative guesses
Generative systems can produce plausible-sounding output that's wrong. Déjà does not invent fixes. It does not "suggest" patches. It does not synthesize remediation steps.
If Déjà cannot prove a match with deterministic evidence, it stays silent.
Validation gates
Gates are conservative by design. They're how Déjà earns trust: by rejecting matches that look plausible but fail reality checks.
Did the PR touch the active blast radius? Verifies that the fix touched the files most likely responsible for the incident identity. Filters out "drive-by" changes that coincidentally occurred nearby.
Did the error rate statistically decouple from the traffic rate? Confirms the incident rate dropped meaningfully after the candidate fix shipped — not just random variance.
Requires that the fix held across a sustained window. Prevents "temporary quiet" from being mistaken as resolution.
How confidence is calculated
Confidence is a bounded score derived from deterministic signals. Déjà does not "feel confident." It earns confidence by passing gates and aligning evidence across tools.
score =
w0 * identity_match +
w1 * file_gate_pass +
w2 * rate_gate_pass +
w3 * duration_gate_pass +
w4 * cross_tool_agreement
confidence = clamp(score, 0, 1)
if confidence < threshold:
withhold- •Confidence is explainable: every component is inspectable.
- •False positives are expensive — the system is biased toward silence.
- •Outputs are artifacts that shipped: PR + diff + deploy context.
Déjà captures decisions — not people. Names provide context, not attribution.
Integrations
Déjà meets teams where incidents already live — then returns recall in the tools engineers already trust.
Sentry · Datadog (and similar sources) provide the production signals used to form stable incident identity and validate fixes over time.
GitHub · GitLab provide the canonical proof of what shipped: the PR, the diff, and the commit trail that resolved the incident.
Privacy by Design: We don't read your code
Déjà operates on metadata and hashes, not raw source code.
- •We ingest ephemeral stack traces.
- •We hash the file paths and error signatures.
- •We discard the raw payload immediately.
- •Your proprietary algorithms never rest in our database.
Ready to deploy the engine?
Connect Sentry and GitHub. Backfill two years. Prove deterministic recall on your regressions.