Engineering progress
Repair the actual path. Keep the original failure.
The examples below show specific changes to failure observation and evidence handling. Each retains its tested source and population. Local verification is separate from operating adoption.
Stable identity for evaluation and selection receipts
IMPLEMENTED NOT RUNTIME VERIFIED
What changed
Entry-code fingerprints record an explicit serialization format. Selection replay refuses unknown formats and retains the historical comparison path.
Why it matters
Keeps evaluation and selection observations tied to the entry code that produced them while preserving uncertainty.
Implemented
Pinned evaluation-input and forward-selection receipt modules. Tracker integration and natural runtime adoption are outside this local test population.
Implementation status
IMPLEMENTED_NOT_RUNTIME_VERIFIED
Observed
The supplied local test record contains 8 passed cases out of 8. Natural runtime adoption is not established.
Environment / scope
Isolated Python stdlib modules with synthetic pure callables Evaluation entry-code identity and selected replay counterexamples; excludes tracker integration and natural runtime
Evidence ID
evo-67b239088702e7c706f9
Source commit
4e9700fdfb8a9dcfc6049d74b0157437d19af3fc
Original observation date
2026-10-09T11:47:53.594795+00:00
Canonical record SHA-256
67b239088702e7c706f9ddaaf755b260fb77e65074934f9cfd3db60491d61958
Remaining limitations
No natural runtime or live adoption observed. Entry-code fingerprints do not attest imported dependencies or globals. Pre/post equality does not rule out change-and-restore races. Selection replay does not reproduce statistics, queue effects or execution authority. The pipeline validates recorded identity and consistency; producer authenticity and implementation meaning require human review.
Run artifact SHA-256
6d0a3fbe23089d0750dbc334bacd39e798f2d0a7f26481655b1bcd71e93380e0
Public source and test records →
Unevaluated inputs no longer count as clean evidence
IMPLEMENTED NOT RUNTIME VERIFIED
What changed
Freshness rates count only explicitly evaluated states. Soak readiness requires a recognized comparison-record kind instead of accepting any JSON object.
Why it matters
Prevents missing evaluation from being reported as a clean measurement or as evidence that a comparison ran.
Implemented
Receipt quality calculation, displayed measurement scope and comparison-observation recognition at the pinned source commit. Original input rows, schemas and readiness thresholds are preserved.
Implementation status
IMPLEMENTED_NOT_RUNTIME_VERIFIED
Observed
Source inspection only; natural runtime is not established.
Environment / scope
Pinned source inspection of receipt quality and comparison-observation functions build_receipt, render and soak_maturity only; operating scan adoption is outside this record
Evidence ID
evo-4a2d40d3a7638c025c39
Source commit
8eb4eee4118d8d037db06a1941708e509d6aa0a0
Original observation date
2026-10-22T11:46:26.993Z
Canonical record SHA-256
4a2d40d3a7638c025c3930c8a4a7504aeb1567f44a493cd9b1d47662e5c31e64
Remaining limitations
No natural operating scan adoption or current collector freshness was verified. Recognizing a comparison kind does not authenticate its producer or payload. This correction does not restore older ledger records, establish live trading coverage or enable external anchoring. The pipeline validates recorded identity and consistency; producer authenticity and implementation meaning require human review.
Public source and test records →
Recorded scoped candidates
Initialization failure and committed risk verdicts.
The complete before/after populations are kept on one verification page. They are examples within the wider evidence index, not the total validation scope.
Read the source, cases and runtime limits →
Later source and publication reconciliation →
Earlier evidence-handling review
Health, provenance and recovery need interpretable evidence.
Local counterexamples expose false positive paths: an unreadable heartbeat, a reused report with a new apparent timestamp, and a recovery claim with no direct target observation.
Implementation and local behavior
Locally verified
Pinned changes, actual producer/consumer counterexamples and separate internal source review. Required local repository checks passed on every exact revision; platform skips remain unevaluated. Draft publication does not establish operating adoption.
Operating adoption and effect
Not verified
No installed-version receipt or natural-run before/after effect was obtained for these revisions.
External reproducibility and attestation
Not verified
Reviewers are internal agents. Private source references and public file hashes do not constitute independent attestation.
A file touch cannot establish operating health
Dated local review
Before
A running synthetic process with a freshly touched but malformed or timestamp-less state file was classified HEALTHY.
After
Uninterpretable heartbeat evidence provides no age. A running target is UNKNOWN unless an existing degradation priority applies; unknown targets stay in the total population but outside healthy and operational numerators.
Six base failures became 17 passing heartbeat tests. The internal reviewer independently repeated 17; a further isolated run of the focused and adjacent selection passed 88 tests.
Existing staleness and classification priorities remain. PID presence and a parsed timestamp do not establish installed code identity or successful application work.
Historical publication state: Reviewed Draft published at the pinned revision. Required local repository selection: 922 passed and 60 skipped, exit code 0. The platform-specific skips are not evaluated. No verified deployment or natural operating effect.
Carried data keeps its original observation time
Dated local review
Before
Partial unreadable input could be treated as a complete record set, while carried report values acquired a fresh apparent observation time.
After
Partial unreadable input follows the existing degraded path. Carried values retain degraded status, carry markers and their original observation time across repeated publication; legacy records acquire no fabricated provenance.
The actual producer-to-dashboard base failed three cases. The final combined selection passed 109 tests, including 47 focused tests; the internal reviewer independently repeated 47. A pre-integration schema suite passed 39 tests.
The tested producer and dashboard preserve degraded status, carry markers and original observation time. Arithmetic and authority behavior were unchanged; live reconciliation and operating adoption are not established.
Historical publication state: Reviewed Draft published at the pinned revision. Required local repository selection: 922 passed and 60 skipped, exit code 0. The platform-specific skips are not evaluated. No verified deployment or natural operating effect.
Recovery requires a direct observation of the affected target
Dated local review
Before
A successful recovery attempt could be counted effective when no health history existed, the observation was outside the window, or only other targets were measured.
After
The producer emits untruncated direct HEALTHY targets. The consumer requires target evidence in the configured window, gives direct unhealthy observations priority, and retains unknowns outside verified and effective counts.
Seven initial base failures and five further review counterexamples were reproduced. Final 69 passed, including the actual producer-to-consumer path; the internal independent reviewer repeated 69.
HEALTHY remains process/log health, not proof of completed domain work or durable restoration. Naive health timestamps, malformed target evidence and legacy absence remain unknown.
Historical publication state: Reviewed Draft published at the pinned revision. Required local repository selection: 922 passed and 60 skipped, exit code 0. The platform-specific skips are not evaluated. No verified deployment or natural operating effect.
Original inputs, outcomes and sources →
Original field meanings and disclosure limits remain in the fixed record. This technical summary changes the presentation, not the historical result. Reproduction of the original evaluations requires access to the private source and tests.
Earlier local engineering review
Input validity and identity remain explicit.
Incomplete input cannot become ready
Dated local review
Before
Two timestamp-qualified rows plus 598 invalid timestamp rows could satisfy a 600-record readiness threshold; malformed tails were ignored and mixed naive/aware times could fail the cycle.
After
Only timestamp-qualified object rows contribute records, depth and diversity. Partial invalid evidence becomes unknown with explicit diagnostics, visible in the report and dashboard.
83 tests passed on the final frozen tree. The independent reviewer repeated 41 measurement and 3 actual dashboard-consumer tests. Base counterexamples failed before the fix.
All absent inputs remain missing; partial invalid inputs become unknown. Completeness covers only listed/discovered files, and older reports acquire no invented unknown count. Freshness, PIT correctness and research quality are not certified.
Historical publication state: Reviewed Draft #4916 updated; exact-head CI was queued or running at this review.
Unknown observations cannot inflate the clean rate
Dated local review
Before
Three UNKNOWN or UNPARSED observations could produce a clean 3/3 result, presented as 100%.
After
Only evaluable states enter the rate. Unknown, unparsed and unprobed rows remain visible and excluded from both sides; no evaluable observations yield 0/0 with no rate.
302 tests and 462 subtests passed. The independent reviewer repeated 15 regressions; the base failed 8 of those cases.
This rate describes an evaluable subset, not whole-fleet coverage. Original rows, states and receipt schema are preserved. Historical operating receipts were not rewritten.
Historical publication state: Reviewed Draft #1500; exact-head CI was queued or running at this review.
Missing decision identity cannot establish lineage
Dated local review
Before
Two executed orders sharing a blank native decision reference could inherit one signal and falsely count as linked: 2/2 in both executed and strict synthetic metrics.
After
Blank native decision references retain raw evidence but cannot establish counted or trace links. Both synthetic metrics become 0/2 while both orders stay in the denominators.
275 full lineage tests and 97 adjacent tests passed. The independent reviewer repeated 14 identity regressions through the real schema, deserialization, trace and rule chain.
Direct nonstring native decision references are rejected. Normal opaque IDs, multiple children and content identities remain compatible. Already stringified producer IDs cannot reveal their original type; missing live DCA identities remain unresolved.
Historical publication state: Reviewed Draft #4923; exact-head CI was queued or running at this review.
Historical source and operating observations
Earlier identity changes, source installation observations and linkage measurements remain in the original records. Their populations are not added to the local test populations above. No new operating effect is established by this presentation.
Neurobrick · systems, authority and evidence
Last updated · 22 October 2026