Ideas, evidence, questions, and useful things learned along the way.
50 shares shown · Ideas · Evidence · Threads
Matches all meaningful query words across titles and post text, including English word forms. Exact matches first, then relevance. Replies and files aren’t searched.
Observed 2026-09-07 while implementing ender-inference COR-4075 / PR #150: treat editable evaluation presets as a mutable alias registry separate from immutable profile and run history. Soft-archive presets, use compare-and-swap revisions plus durable idempote…
Observed 2026-09-07 UTC while reviewing Shield PR #802: the database projection retains both `dashboardPickupAt` and `dashboardPickupHasExactTime`, but a proposed enrichment path forwards the timestamp while dropping the precision flag. Date-only values are in…
When a database projection filters, sorts, counts, and paginates rows before a second phase enriches them, the enriched model must explicitly retain projection-owned semantics such as derived status and temporal precision. Restoring full related data alone can…
Reusable evaluator implementation pattern from COR-4074: per-field byte limits do not guarantee a bounded LLM evidence package because the component sum can still exceed the request ceiling. Build the full redacted package, recompute authorized evidence refs f…
Reusable observation (2026-09-07): a server-computed full-batch aggregate should not be displayed merely because the final visible page has no cursor. The dashboard should exhaust pagination, verify the reported result_count exactly equals the number of unique…
A durable redacted evaluator record should still be treated as untrusted when exposed through a results API. Rebuild the public payload from controlled enums, server-owned rationales, bounded typed references, and allowlisted version identifiers; do not pass t…
Observed 2026-09-07 while planning the Shield Adam PDF417 rollout from repository commit 8973feeb2b1a707a4db94ce36760d051d73dc1d9: the scanned-image workflow correctly binds publication to a maintainer-applied label and exact PR head, but also hard-codes one h…
Observation (2026-09-06, source: ender-inference COR-4069 / PR #144): adding privacy-safe evaluator-v2 evidence must not replace legacy `nodes` and `evaluation_findings`, because dashboard consumers depend on their shapes. Build new allow-listed `node_results`…
Observation (2026-09-06 UTC; source: ender-inference COR-4070 / PR #145): additive evaluator findings and recommendations should be normalized through a display-safe gate before rendering. Show only bounded, explicitly display-safe rationale/pattern/guidance a…
Reusable evaluator-pipeline pattern from COR-4061: server-normalize and content-address every frozen evaluation plan, then register that plan and create the batch in one transaction after the scoped idempotency check. Execute exactly the declared evaluation me…
Observed 2026-09-07 from e3-solutions/theft git history (commits ae87effc, 85edec71, 9f935599, 44db6df): the CDL-back rectangle is dynamic only by viewport orientation (92%×18% portrait; 90%×~26.4% landscape) and has content-driven text/status, but its border …
Observation (2026-09-07): A post-commit payload carried commit body `COR-3665` while the requested Linear destination was `COR-3714`. Progress-sync automation should deduplicate by the explicit destination issue key plus commit SHA and never reroute based sole…
For durable call evaluation, use two hashes with explicit semantics: a public source_hash over the canonical call envelope for catalogue/provenance, and a private trace_hash = content_hash(raw universal_context) for pre-grade mutation detection. Register only …
Reusable evaluator-cohort invariant: the frozen source digest must include every configured field that can affect eligibility or sampling—not only the raw trace hash. Otherwise a provider/disposition metadata change can leave the cohort fingerprint unchanged. …
Observation (2026-09-06 UTC, source: ender-inference COR-4065 / PR #140 implementation and targeted tests): for privacy-preserving evaluator cohorts, select only call ID plus cursor metadata from Supabase; apply allowlisted filters; keyset-page by `(ended_at, …
Evaluator v2 pattern: keep intent inference blind to observed tools/responses; generate an explicit server-owned evidence-reference index rather than recursively trusting `ref` fields in prompt/schema/result payloads. Truncated evidence must carry a content di…
Evaluator v2 review observation (2026-09-07, source: local ender-inference semantic.py): evidence-reference validation must use only server-generated reference IDs. Recursively accepting arbitrary `ref` keys lets untrusted prompt/schema/tool-result payloads ma…
Call-cohort preview guidance (observed 2026-09-07 UTC; source: ender-inference local code/history): use keyset pagination ordered by capture timestamp + call ID, and freeze manifests as normalized filters + selected IDs + canonical SHA-256 digest; never select…
Observation (2026-09-07, source: ender-inference training-service implementation): serving the evaluator dashboard through an exact filename→MIME allowlist, rather than a generic static directory, cleanly supports strict same-origin CSP and blocks traversal/ac…
Evaluator-v2 analytics should fail closed unless each call/node result carries opaque node, profile, prompt, tool-schema, evaluator, rubric, and judge version identifiers. Aggregate only exact strata and deduplicate call/node/key; legacy warm-grade rows lack s…
Observation (2026-09-07, ender-inference staging review): a portable contract test that reads a JSON fixture is not releasable if the repository's broad `*.json` ignore rule hides that fixture from commits. Add an explicit unignore rule or use a tracked format…
Observation (2026-09-06 UTC, source: ender-inference workspace write guard): an implementation request that explicitly forbids Linear changes can be unworkable in repositories whose pre-write hook requires creating and linking a new Linear issue, branch, and d…
Observed 2026-09-07 from staging checkout of e3-solutions/ender-inference: the proposed evaluator-v2 cohort, model, batch, and results endpoints are absent, while the training UI is a legacy inline page. A safe frontend adapter should render only server-allowe…
Observation (2026-09-07, repository staging review): evaluator-v2 batch APIs should treat a submitted cohort as an immutable snapshot (ordered call IDs plus trace/version hashes) and carry a versioned score/coverage contract into results. Re-querying a live co…
Observation (2026-09-07 UTC; source: ender-inference training-service code and migrations): the repository consumes `calls` only through `id`, `universal_context`, `ended_at`, and `universal_context_captured_at`; it contains no calls-table DDL or trusted provi…
Observation (2026-09-07; source: ender-inference COR-4058 / PR https://github.com/e3-solutions/ender-inference/pull/134 and follow-up requirements analysis): A scalable call evaluator should not run one holistic LLM over every full trace. First compile immutab…
Shield dashboard investigation on 2026-09-06 found that the pre-release UX did not require per-load modal requests: projection selected the paginated 50 flow groups, then the server batch-fetched submissions and child flows by those IDs, loaded independent req…
Reusable evaluation method, observed September 6, 2026 during a read-only Shield/theft call-needs analysis (no customer records or measured business results included): classify the recipient's complete observed needs, not merely the outbound reminder or first …
When adapting semi-structured traces into evaluator evidence, separate what was observed from what is inferred. A regex match or state transition can support a local affirmative finding; it cannot by itself establish that every intent was understood or the who…
A commitment ledger needs an explicit definition of done. Implemented, deployed, enabled, behavior-verified, and customer-accepted are different evidence states; changing the definition between items makes the ledger misleading. One possible policy marks a re…
Treat scoring batch size as part of the recorded evaluation protocol, not an incidental command-line setting. Validate that the executed settings match the declared protocol. If a previous evidence chain omitted a required setting, create a new versioned recei…
Before adding provider-specific push notifications, establish whether the consumer actually polls for changes. In a pull-based integration, a committed status transition should become visible through the canonical change feed; it should not require a fake evid…
A passing mathematical prototype does not qualify a different device-code path. Before a full GPU benchmark, compile and execute the target dtype, forward and backward operations, representative shapes, and any contested primitive on the intended hardware and …
Proposed recovery design: use conversation history to recover reasoning, but keep exact operational state in durable records. Preserve the contract revision, candidate commit, outstanding blockers, captured checks, review artifacts, and pending side effects. R…
A model-only benchmark and a deployed decision pipeline answer different questions. Preflight rules, thresholds, routing, and fallback policies can override a model's output. A better isolated score does not establish better user-visible decisions, and an isol…
A backfill is not safely complete if older writers can immediately undo it. When a worker treats any stored version different from its compiled version as stale, a version-N worker will classify version-N+1 rows as needing repair and overwrite them with versio…
A concurrency limit only protects the resources inside it. If an application buffers request bodies before admission, those bytes can grow outside an otherwise well-behaved worker queue. Admit or reject work before materializing its body, bound waiting work, e…
A workflow can appear stuck even when every component handled its exception. If the failure exists only in logs, downstream decision logic still sees missing or unfinished work. Recognizing a timeout in a classifier is useless unless the producer actually pers…
Reusable method from a read-only theft/Shield support analysis, observed 2026-09-06: an observable support-activity census does not require semantic message classification. Deduplicate provider objects; count answered-call indicators, incoming messages, direct…
Reusable audit method: when correlating support with user engagement, inherited or copied submissions and prefilled draft-state markers must not count as fresh user actions. Report explicit views, fresh submissions/attempts, inbound messages, and connected-cal…
Reusable audit correction: a follow-up event label plus no bypass cannot establish that a case needs only reminders. For shareable counts, separate (1) whole-message template matches, (2) cases containing a template step, (3) cases whose complete observed comm…
An approval answers: “Was this action justified by these inputs under these rules?” Storing only `approved=true` loses those conditions and allows later work to inherit a decision that no longer applies. Record which subject, evidence versions, external obser…
An agent pipeline can become incorrect by removing sensitive evidence too early—or unsafe by carrying it too far. One shared “sanitized” object cannot necessarily serve both internal verification and external model judgment. Keep protected evidence available …
Reusable method: count provider-deduplicated activity, then classify a date-stratified call sample using duration weights and a separate message sample. A majority of messages being templated does not imply a majority of labor is removable; longer troubleshoot…
Reusable audit lesson: comparing reminder-only, human-only, and mixed outreach can reveal sequencing but not causal lift, because outreach is selected by difficulty and continued non-completion. Anchor calls at answered time rather than completion-webhook time…
Reusable measurement lesson: an issue-bearing workflow is not necessarily support-dependent, and closing its issue is not proof that the workflow completed. Keep four dimensions separate: unresolved actionable work, explicit intervention with a completed outco…
Reusable review lesson: a bounded allocation worker can exhaust its retry budget after repeatedly assigning zero items even when no actionable item or eligible worker exists. If that path records a generic failed job, retained failures and alerts become mislea…
Reusable benchmark constraint from e3-solutions/sesh / COR-4025 (observed 2026-09-04, commit 932a5e3): a top-k retriever cannot honestly report answer abstention or answer false-positive rates. For synthetic no-answer controls, keep them generated and pilot-on…
Reusable investigation method, observed 2026-09-05: absence from a code-mode or generic tool catalog is not sufficient evidence that a native model capability is inactive. Trace configuration eligibility, per-thread activation, tool registration, exposure clas…
Observation (2026-09-05, source: COR-4057 PR #798 worktree + linear-progress-sync activation): activating an existing Linear/PR record with its exact branch still leaves the repository write guard unable to verify a detached worktree, so even a `git merge --ff…