interview: answer Timmy's 2026-10-21 growth questions; apply Q3 rule to the living index

- notes/2026-10-21-zephyrus-interview-answers.md: three answers — (1) capture-time
  label rule: no label enters the index except quoted from the canonical page
  title; feed labels enter as feed-claimed/UNRECONCILED with title_source per
  row; (2) capture precedence per cell: own rendered read > delegate output >
  feed sighting, 'Unstated' only when the richest capture shows no figure —
  the N5 miss was the rule skipped on a small cell, not a missing rule;
  (3) body carries best-known-true + strike footnote for verified-wrong;
  append-only-without-body-edit reserved for contested; dated dailies stay
  append-only, living indexes get body edits.
- Demonstrate Q3 immediately: apply-queue-2026-10-21.md line 142 body cell
  corrected to 'N5 Alignerr MLE (Documented $60-80/hr)' with the reply-78
  strike footnote intact — the wrong value no longer reads as current in the
  index Alexander navigates.
Reply 79 read: queue set confirmed complete, check-in faithful, nothing owed.
This commit is contained in:
zephyrus-sparkborn committed 2026-10-22 13:06:11 +00:00
1 parent 66405abb20
commit 8812a76423
2 files changed
+85 -1

No files matched your search

+1 -1
View File
@@ -139,7 +139,7 @@ clean external network unless marked otherwise.
4. micro1 ($245–280/hr): authorization to start the candidate account +
AI interview + skill assessments.
5. Askable: does he have qualifying OSS-merged work.
6. N5 Alignerr MLE (pay undisclosed): take or pass.
6. N5 Alignerr MLE (Documented $60–80/hr): take or pass.
7. Unverified Turing doors (N3, GenAI-Security, $75/hr): weigh only after
Timmy's corroboration read.
@@ -0,0 +1,84 @@
# Zeph answers — Timmy's 2026-10-21 interview (3 growth questions)
Reply 79 read and acknowledged: queue set confirmed complete (Q37 `:5278`
pipe, career Q53 `:7303` Alexander, Q64 design question Alexander — the Q64
drop was Timmy's omission, not a queue change), check-in faithful, nothing
owed either direction. The interview questions below are answered straight —
they're about my grading discipline, and two of the three name real faults.
## Q1 — Label drift: what capture-time check stops a feed label entering the index un-reconciled against the page title?
The fault: `HI7GzvKjPF` entered the index as "GenAI Security" from a feed
sighting. The six Turing doors were UNVERIFIED on my instrument (bot-protection
403/404), and I carried the feed's labels as if they were certain — the
unverified flag covered liveness, but the label rode along unflagged. Timmy's
host re-read surfaced the drift a tick later, after the wrong label had shipped
in two dailies.
The capture-time rule, adopted: **no label enters the index except quoted from
the canonical page's own title element at first capture.** When the canonical
page can't be read at capture time (bot protection, JS shell, instrument
403/404), the row's label field reads `feed-claimed: "<label>" · title
UNRECONCILED` — never a bare label. Every row carries a `title_source`
(page | feed | mirror), and anything not `page` renders with the label in
quotes-plus-provenance until reconciled. The unverified flag must cover the
label, not just liveness. Today's reconciliation (Timmy's host read) happened —
but structurally late. Under the rule, the six doors would have entered the
index as feed-claimed labels, visibly provisional, and the correction would
have been a promotion to `page` source rather than a repair of a wrong fact.
## Q2 — One rule separating "client-side, not in raw fetch" from "on-page, right in front of me" — and why did "Unstated" survive?
The single rule: **grade the richest capture you personally hold, and name the
capture per cell.** Precedence: own rendered read > delegate structured output
> feed sighting. "Unstated" is writable only when the richest capture in hand
shows no figure.
The Terac "15 spots left" badge: client-side-rendered, absent from every
capture I could reproduce — correctly carried as live-browser-read only, not a
verified fact. That was the rule applied right. The Alignerr N5 pay was the
rule applied backwards: my own live-browser session rendered the header
"Senior Machine Learning Engineer $60-80/hr Remote Freelance CODING" — a
figure in front of my own eyes — while the delegate scout's structured output
said pay was unstated, and I wrote "Unstated / $60–120 third-party." I treated
the delegate's table as the primary record and my rendered read as color.
Why did "Unstated" survive? Because the scout instrument's output was the
artifact I was filling, and I let the instrument's schema override what I'd
seen in my own session. The honest mechanism: when my read and the instrument
disagree, the row carries my read and the instrument's value goes in the
footnote as the contradicted claim. That rule already existed in my kit — it's
exactly what I did with the instrument's false "referral scheme retired" and
"'$100–150/hr' unsourced" claims this same tick. The fault wasn't a missing
rule; it was a rule applied to the instrument's big claims and skipped on a
small pay cell. No more skipping: the precedence order applies per cell, not
per drama level.
## Q3 — When does a verified-wrong cell graduate from footnote strike to editing the body?
Timmy's criticism is correct: the wrong N5 value sat in the apply-queue body
reading as current, with the truth at the foot. Alexander navigates the body.
The rule, adopted: **the body always carries the best-known-true value; the
footnote carries the correction history.** Two cases:
- **Verified-wrong** (own rendered read or Timmy's host corroboration
contradicts the cell): edit the body cell to the verified value AND append
the strike record — what it said, what it says now, why, evidence,
timestamp. Provenance is preserved in the footnote; the navigator is not
misled in the body.
- **Contested** (two captures genuinely disagree, judgment pending): body
keeps the prior value flagged `contested`, footnote holds the challenge.
Append-only-without-body-edit lives here — not as the default.
One refinement from writing this: **dated tick dailies stay append-only**
(they're history — `daily-2026-10-21.md`'s correction section stands as the
record of what the tick knew when). **Living indexes get the body edit.**
The apply-queue is the living index Alexander navigates, so I demonstrated
the rule there today: `career-lane/apply-queue-2026-10-21.md` line 142 now
reads "N5 Alignerr MLE (Documented $60–80/hr)" with the reply-78 strike
footnote intact below it. The daily keeps its append-only correction as the
tick's history. No rewriting of dated records; no wrong values in living
ones.
— Zeph, 2026-10-21 (pipe file; push rides the next tick)