This page follows one failure from the first signal to an inert repair proposal. Each box below exists in the code at the current revision; the arrows are data handoffs, and every handoff is a validated, content-addressed document rather than shared state.
host application (development, test, or CI only)
┌──────────────────────────────────────────────────────────────┐
│ Logger reports ─┐ │
│ Telemetry ─────┼─► Collector.Session ─► Recorder (bounded) │
│ Poller samples ─┘ │ │ │
│ detectors freeze on finding │
└───────────────────────────┼───────────────────┼──────────────┘
▼ ▼
Finding (v1 envelope) sealed snapshot
└────────┬──────────┘
▼
Incident.Capsule (redacted, ≤ 1 MiB, SHA-256 id)
│
▼
Sentinel.Manifest (binds exact capsule bytes;
confirmation only, no mutation)
│
disposable sandbox target(s) ──────▼─────────────────────────────
Confirmation.run ─► Regression (2 fresh matching failures)
│ │
│ ├─► Minimizer (delete whole steps)
│ ├─► Bisection (adjacent boundary over
│ │ a host-declared list)
│ └─► CodingAgent.Handoff
│ │ (agent proposes bytes)
│ ▼
│ CandidateSubmission (declared unverified)
│ │
└──────────────► Verification.run (2 before, 2 after)
│
▼
RepairReport ─► DraftProposal ─► SourceOwnership
│ Dedup
▼
DraftCreation gate (disabled by default;
test providers only)1. Collect bounded evidence
A trusted host starts Insider.Collector.Session with an explicit environment,
authority string, clocks, a recorder classification, and the components it
wants: :logger, :telemetry, :poller. The session API rejects :staging
and :production before it creates any process, filter, handler, timer, or
monitor.
- The Logger collector installs a first-position primary filter so it sees crash and supervisor reports before they are turned into text. Logger callers do only bounded structural shaping and one fixed-slot ETS insertion; a separate process drains into the recorder.
- The Telemetry collector accepts only events named in a pure-data
Insider.Collector.Telemetry.Policyallowlist, and only the scalar measurement and metadata fields that policy names. - The Poller records scalar samples (queue length, memory, reductions, status, registered name, current function) for at most 32 explicit local processes. It never copies mailbox contents, process dictionaries, links, monitors, or stack traces. When asked (Insider Live asks by default), it also ranks every local process by mailbox length, memory, and reductions since the last sample, and every ETS table by memory and size, reading only scalar counters and names, and reports Insider's own processes and tables apart.
Everything lands in Insider.Recorder, a fixed-size pre/post-trigger store.
Admission is direct and bounded: record/2 returns {:dropped, reason} under
pressure, rate limits, or classification failure rather than queueing, and the
snapshot keeps those gaps visible.
2. Detect and freeze
Detectors are pure reducers over normalized events:
Insider.Detectors.RestartRecurrence (three abnormal terminations of distinct
child incarnations under one logical supervisor/child identity, by default)
and Insider.Detectors.MailboxGrowth (a configured run of strictly increasing
queue samples with an unchanged reductions counter and no sampling gap). A
collector session does not run detectors by itself. The host, or a Lab
capture workflow such as Insider.Incident.RestartCapture, feeds normalized
events into a detector and freezes the recorder when it emits a finding,
which seals the window around the trigger.
3. Seal a capsule
Insider.Incident.Capsule assembles the sealed snapshot, the normalized
finding, provenance (code, configuration, dependencies, runtime, and target,
each independently fingerprinted or explicitly unavailable), audit authority,
and digest-only artifact references into one deterministic, redacted document.
The redaction gate runs on every export root with a trusted
Insider.Redaction.Policy; undeclared secret data drops by default and any
failure yields no partial output. The POC capsule accepts at most 64 events,
32 artifact references, and 1 MiB of canonical JSON, and rejects larger
snapshots instead of truncating them silently. Its canonical decoder validates
corruption and alternate encodings offline, with no live node.
4. Hand off to Sentinel
Insider.Sentinel.Manifest (Lab-only) binds the exact capsule id, byte size,
and full-byte digest, lists which prerequisites are available, unavailable, or
redacted, and carries a fixed execution policy: purpose confirmation, source
mutation prohibited, target class disposable. The manifest is a reference,
not a second copy of the evidence, and it conveys no execution authority.
5. Reproduce, regress, shrink, bisect
Everything from here runs in the separate tools/insider_lab project against
disposable sandbox targets supplied by a trusted provider allowlist. The core
never invokes a shell, filesystem, network, node, container, or cloud API.
Insider.Sentinel.Confirmation.run/6validates manifest, capsule, scenario, and target provenance together, then runs the portable scenario on a leased target and records one of five outcomes:matching_failure,nonreproduced,repaired,unrelated_failure, orinfrastructure_failure.Insider.Sentinel.Regressionturns one eligible confirmation plus its scenario into a data-only regression recipe.Regression.classify/2reportsfailedonly when two fresh, distinct confirmation runs both reproduce the same incident on the same baseline target with complete evidence and cleanup.Insider.Sentinel.Minimizer.run/6tries deleting whole scenario steps, three fresh confirmations per candidate, and stops when no single remaining step can be removed. It does not claim a globally smallest reproducer.Insider.Sentinel.Bisection.run/6binary-searches a host-declared ordered list of 2 to 64 targets, proving both endpoints first, and isolates one adjacent nonreproduced/reproducing pair. Insider never reads a repository or orders revisions; the host supplies the list and owns the assumption that outcomes are monotonic across it.
6. Propose and verify a repair
Insider.Sentinel.CodingAgent.Handoffpackages the capsule, manifest, regression, and two fresh failed confirmations into one bounded request with a fixed objective: propose a candidate in theinsider.verification-candidate/v1format. It accepts no caller-authored instructions or causal claims.- A coding agent answers with opaque bytes (at most 16 KiB, format
opaque/v1orunified-diff/v1).CandidateSubmissionbinds those bytes to the exact handoff and labels themdeclared_unverified. Insider.Sentinel.Verification.run/7runs two fresh live confirmations against the regression baseline (they must fail), then two independent post-repair sandbox runs through a host-allowlisted body module. The record isverifiedonly when both after-runs observe no failure, agree on changed code and unchanged configuration/dependencies/runtime/target provenance, carry distinct positive oracle receipts, clean up completely, and leave the executor with zero active and zero quarantined leases. Otherwise it isinvalidwith sorted machine-readable reasons.
7. Produce inert review artifacts
Insider.Sentinel.RepairReport.Manifest recomputes a canonical report from the
capsule, manifest, and verification record and renders deterministic Markdown
with no caller-authored success prose. Insider.Sentinel.DraftProposal rechecks
the full five-run matrix and records network_effect=none and
repository_mutation=none. SourceOwnership derives routing hints from the
diff paths and a small declaration syntax; Dedup groups capsules and proposals
by finding fingerprint in a bounded in-memory ledger.
The last gate, Insider.Sentinel.DraftCreation.Executor, is disabled by
default and, in this phase, has only test-fixture providers. Every request it
would make is draft-only, requires human review, carries no credentials, and
exposes neither merge nor deploy. There is no real GitHub or network provider.
What a complete proof means
The proof workflow
states it exactly: the same bounded causal fingerprint appears
before the repair, a regression fails before and passes afterward, the healthy
control remains quiet, resources are cleaned up, overhead is measured, and an
independent reviewer validates the evidence. Anything short of that is
reported as inconclusive, invalid, not_proven, or nonreproduced, never
rounded up to success.