EBO v1.2.0

EBO 1.2.0

Capture

• Devin CLI harness (ebo devin run): EBO owns a pinned devin acp child
  (Devin CLI 3000.11.3) over stdio and retains every Agent Client Protocol frame
  as native session evidence, with OTLP/HTTP protobuf telemetry received by an
  EBO-owned endpoint. See the Devin CLI guide.
• Retained readback accepts every released harness pin: each version-gated
  adapter keeps a retained-version list that pin bumps append to, and the
  manifest version must match the version the capture's single native
  composition record names (Pi, DeepSeek).

Normalization (changed in place)

• Per-request token usage events (resourceSemantics: increment) for the
  Claude Agent SDK, Codex and DeepSeek Harness, with native request identities.
• Records without an origin timestamp use the callback or receipt clock,
  labeled nativeTimeSource; delegation records carry task identity, status and
  content references; Pi history records share identities with stream events.
• Tool events carry toolName, inputDigest, isError and exitCode where
  the harness exposes them; PostToolBatch callbacks are not tool operations.
• globalEventKey is <runId>/<attemptId>/<eventId>, percent-encoded.

Normalized datasets differ from 1.1.x for the same bundles. Observation sets and
behavior assertions bound to 1.1.x dataset digests no longer validate:
regenerate observations and re-judge before aggregating or building an Atlas.

Evaluation

• Occurrences (structural extractor 1.1.0): failure responses, validation
  runs, source changes, repeated operations, compactions and delegations, each
  citing only its own events, with unavailable types reported as such. See
  structural observations.
• Occurrence ratings (ebo occurrences rate): typed questions per occurrence
  answered by a System One decision model (TypeSafe Jev or the Fireworks
  decisions API), retained as decision records with an explicit acceptance
  policy.
• Judge preparation (ebo judge prepare): requests show the judge every
  occurrence of the selected types as a ledger row with its ratings, choose full
  records by a recorded policy, and accept citations that name an occurrence.
  See occurrence ratings.

Exports

One redaction implementation decides what portable exports, Atlas reports and
evidence packets redact and what the final secret scan rejects. Secret-named
assignments are classified by context: shell values are scanned as shell words
(literal parts redacted, complete environment references kept), and code
expressions keep syntactic references while their literals are redacted. Atlas
report.json lists every scan match with a resolvable JSON pointer.

Verification

npm run acceptance checks the source, tests, documentation links and two
byte-identical package builds. Normalization, occurrences, ratings and judge
preparation were also exercised on 48 retained attempts across four harnesses
(native content stays local). reproducibility.json
records the pinned runtimes and fixture digests.