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.
A reusable UI migration method: when dividing a mature single-page form into focused pages, retain each control's DOM identity and toggle hidden/inert on page containers. This preserves in-memory drafts and event bindings, but DOM tests alone do not establish …
When adding saved-run navigation to an evaluator UI, separate the mutable draft from the opened immutable run, and bind each asynchronous progress/result response to the currently opened run. Also invalidate visible summary fields immediately on a run switch: …
When assessment dimensions share findings, adding their totals double counts evidence. Compute call-level coverage from unique finding identities, exclude explicitly not-applicable results from the applicable denominator, and retain not-assessable and missing/…
A transport mock can validate the intended query string while preserving a real escaping bug. In a bounded-history read review, SQL LIKE literal-underscore escaping was correct in isolation, but PostgREST's quoted-value parser consumed that backslash before SQ…
When create-new-version and delete-final-draft share a PostgreSQL advisory lock, acquire that lock before checking the aggregate parent exists. Checking first allows creation to observe a parent that a concurrent delete later removes, then resume after the loc…
Reusable design note: in versioned assessment systems with canonical evidence, hard deletion should be capability-based. Permit it only for draft, unreferenced configuration and invitations that have no session; once published, referenced, or session-linked, r…
PR #805 review lesson (2026-09-08; source: e3-solutions/theft exact diff origin/staging a72d82dd… to ff04495b…): starting RF-DETR concurrently is insufficient if downstream retained scheduling still awaits a guide-lane budget; that serializes detector-first su…
Reusable testing lesson from Shield Required Review rollback (COR-4076 / PR #804, observed 2026-09-08): a browser test that opens a review modal, then directly mutates its child to Failed before clicking Decline tests a stale-UI race rather than a stable custo…
Observed 2026-09-08 UTC while resolving COR-4084 in e3-solutions/ender-inference (PR #155, staging evals-service): when a Supabase/PostgREST integration supports configurable source tables or views, metadata fields such as disposition and provider should remai…
Observation (2026-09-08 UTC, source: COR-4082 implementation and independent code-review/test evidence in the Shield repository): a low-latency user-guided barcode lane can safely race RF-DETR without changing the recovery path when four invariants are explici…
Observed 2026-09-08 while designing Shield COR-4083 from the preserved 2026-09-04 Gemini/RF-DETR benchmark contracts: when model outputs expose confidence but only some workflows have a numeric production floor, predeclare any added threshold as experiment-onl…
ROI-crop benchmarks should report pixel-area reduction as the primary crop-size metric and encoded-byte reduction only as a secondary, format-qualified metric: comparing an original JPEG/HEIC to a lossless PNG crop otherwise mixes cropping with codec effects. …
COR-4082 integration review surfaced three reusable cross-service checks. First, extracting shared UI geometry can silently regress another scanner even when the new scanner is correct: preserve existing orientation caps and add regression cases for every cons…
Reusable evaluator integration lesson: A call can be correctly excluded from scoring while still exposing an evidence-plumbing defect. In the staging call evaluator, the canonical universal_context contained the node transcript but omitted lifecycle_summary/f…
Reusable UI/data-contract lesson from the call evaluator dashboard redesign: Keep the public result projection aligned with the visual hierarchy. Use flattened findings only for totals and summary counts; render each selected evaluation exactly once inside it…
Observed 2026-09-08 during COR-4032 local labelled-data verification (Google Vision TEXT_DETECTION receipts; no raw OCR persisted): four crops each returned one unambiguous USDOT candidate but disagreed with the original-image candidate. Modest context widenin…
Adam PDF417 guide-crop design observation (2026-09-08; source: current origin/staging code review): a concurrent guide decoder must be represented as one parser-validated attempt through Adam’s existing admission/scheduler/worker-pool accounting; an out-of-ban…
Observed 2026-09-08 while implementing ender-inference evaluator call selection (COR-4080 / PR #153). A filtered cohort cursor page can legitimately return zero eligible rows and still provide next_cursor when the bounded source scan was consumed by excluded r…
Reusable observability lesson from the Adam PDF417 phase-timing rollout: terminal rejection projections may intentionally report zero aggregate detector/scheduler and strategy counters even after detector and planning have completed. Diagnose deadline location…
Observed 2026-09-08 during negotiation COR-3987/COR-3991 staging acceptance. When a transferred human sees the receiving Telnyx DID even though application tests assert original ANI, first distinguish a downstream rewrite from the application’s bounded fallbac…
Reusable deployment-RCA method: do not infer the active LiveKit runtime from the current `livekit_migration` branch head. A later guarded auto-deploy can fail before build/deploy while the previous release remains active. Verify the public readiness response's…
Observed 2026-09-08 while repairing the ender-inference staging semantic evaluator (COR-4079). OpenAI Responses `max_output_tokens` covers both visible output and hidden reasoning, so a small structured-evaluation budget can be exhausted before any JSON is ret…
When adding a tool to a running MCP gateway, verify three distinct boundaries: the server advertises the tool for the authenticated caller; a real consumer has imported a callable schema; and that consumer executes the tool and follows its source handle. A lon…
A reported release SHA is a label, not an attestation of the running artifact. Before a narrow gateway rollout, compare the deployed routing/configuration and application files against the proposed base; an older labeled base can accidentally undo already-live…
Reusable method from e3-solutions/sesh / COR-4025, observed 2026-09-08 UTC; provenance: repository inspection and independent hosted parity probes, implementation commit 526bb2f. A vector index elsewhere in a schema does not establish that a search endpoint us…
Methodological adaptation, observed 2026-09-07 while reviewing Qdrant FineWeb-10B and Supernova; no benchmark execution or performance claim. FineWeb ground truth is brute-force vector ranking, not passage-answer relevance (https://huggingface.co/datasets/Qdra…
Method follow-up from e3-solutions/sesh / COR-4025, observed 2026-09-08 UTC; provenance: docs/answer-consumer-diagnostic.md, retrieval baseline cd22be8. Keep the retrieved response fixed and give each question to a fresh reader context without expected labels.…
Observation (2026-09-07, COR-3791, ender-inference commit 77d7480): a balanced end-of-turn holdout cannot require Silero/ASR speech pass on the full future window. Genuine finished turns commonly have caller silence or agent-first audio after the checkpoint, s…
Observed 2026-09-08 UTC while testing e3-solutions/sesh, COR-4025; method and aggregate evidence recorded in commit cd22be8, docs/search-quality-guardrails.md. A provisional 0.40 cosine cutoff emptied obvious off-topic searches, yet also emptied two of five ad…
Observed 2026-09-07 on ender-inference COR-3791, commit bf9ff74: in a selected-adapter warm-start Reference-LoRA run (rank-8 all-linear, LR 1e-4, one epoch, exact 90-row balanced corrective/rehearsal corpus), raising selected-correct preservation weight from 4…
For COR-3791 in e3-solutions/ender-inference, a genuine recent-production holdout finalizer now requires one checkpoint per call/caller, explicit hash-bound exclusions from training and prior holdouts, fixed date windows with deterministic bounded extension, a…
Observed 2026-09-07 in e3-solutions/ender-inference, COR-3791 (branch vansh/cor-3791-fine-tune-a-freight-turn-detector-for-identifiers-pauses-and): a fresh qualification set should not be another deterministic subsample of a repeatedly used reviewed authority.…
Observed 2026-09-07 on e3-solutions/ender-inference COR-3791, branch `vansh/cor-3791-fine-tune-a-freight-turn-detector-for-identifiers-pauses-and`, commit `732f414`: an all-linear rank-8 Reference LoRA warm-started from the selected lm_head adapter was trained…
Observed 2026-09-07 during a bounded real-corpus Sesh diagnostic (not an independent relevance benchmark): a link-heavy passage ranked above substantive guidance, even though the returned family/neighborhood contained relevant material. Repository filtering al…
A retrieval-ingestion method lesson: test unchanged reruns through the real cache representation, not just model output reused in memory. If immutable revision identity hashes numeric embedding payloads, a vector database's shortest float32 rendering can diffe…
Observed 2026-09-07 during exact-head verification of e3-solutions/kch-quoter: local PostgreSQL 16 migration suites produced false failures when initialized with SQL_ASCII or C UTF-8 because psycopg returned text as bytes and POSIX `[[:print:]]` rejected valid…
Observation (2026-09-07): When adapting a large telephony-provider migration from a later agent architecture back onto an older single-prompt agent, transplanting the later PR is unsafe because it pulls in unrelated flow, persistence, and service assumptions. …
Methodological synthesis, 2026-09-07. A rejected tool call is not, by itself, evidence that its guard is defective. Diagnose the chain separately: the user's utterance, extracted tool arguments, guard decision, and resulting response. An upstream extraction mi…
Method learned during e3-solutions/sesh / COR-4025 review (2026-09-07): review cross-schema dependency behavior separately from destructive SQL. A CREATE-only private derived schema can add ON DELETE RESTRICT foreign keys to an existing identity table; the emp…
A reusable correction from COR-3791: shared HMAC domains plus each selected row's direct identity tokens do not prove disjointness when protected populations are projected independently. A hidden transitive bridge can connect A—B in one population and B—C in a…
Retrieval-method learning, observed 2026-09-07 while investigating passage delivery: adding a short source-quoted turn opening to every long-turn window can help preserve subject and qualifications, but repeating that opening in the returned evidence bundle co…
Observed 2026-09-07 in ender-inference/COR-3791 local benchmark audit. A 1,800-row pack had perfect manifest→clip lineage and 1,421 strong accepted-final↔stored-turn bindings, yet zero rows could be called causally audio-aligned because recordings lacked sampl…
Method synthesis, 2026-09-07. Extend non-reportable retrieval diagnostics with paired questions: an answerable question, a meaning-preserving paraphrase, and a minimally changed false-premise or unsupported-fact question. Separate premise_contradicted from ins…
Observed 2026-09-07 while repairing COR-3791 in e3-solutions/ender-inference at commit df24ba9. A zero-overlap result is not meaningful when the compared cohorts pseudonymize the same underlying identity with different HMAC/hash domains or payloads. The safe p…
Reusable audit lesson (observed 2026-09-07; provenance: re-running read-only lifecycle/transcript predicates during a voice-call failure investigation): a symptom rule applied only to a previously labelled candidate subset cannot establish its population-wide …
Method synthesis, observed 2026-09-07 while designing a coding-session retrieval benchmark. Extends the published safeguard that synthetic pilots must remain non-reportable: genuine historical questions can still leak their answers if the indexed supporting me…
Observed 2026-09-07 while investigating negotiation/ender-inference abrupt calls (COR-4077, draft PR https://github.com/e3-solutions/negotiation/pull/1014). Read-only source review at negotiation staging 610c61ec and private historical lifecycle evidence showe…
Reusable ingestion boundary learned while tracing Sesh on 2026-09-07: a per-message private-object store must not be adapted by inventing one session-level Storage object. A safe session indexer should enumerate scoped changes with a durable `(updated_at, stab…
Observed 2026-09-07 during an ender-inference staging evaluator-v2 launch. A live calls catalogue used `initiated_at` while the evaluator defaulted to `created_at`; an explicit staging column override restored cohort previews. Separately, an absent batch RPC s…
Observed 2026-09-07 during e3-solutions/theft COR-4076 (PR #804), comparing PR #722 with its parent. Removing an episode-based review workflow while retaining its queue/history UI requires two distinct compatibility contracts: immutable past decisions must rem…