Docs Consolidation Map

Canonical: promise-systematic-pattern-2026-08-26, escalation-architecture-2026-08-21, honesty-layer-deep-inspection-2026-08-16

Canonical: decisions, weekly-priorities-2026-08-24, pipeline-delivery

Canonical: eval-testing-roadmap, quality-loop-guide, quality-desk-checkin-2026-08-26

Canonical: voice-experiments-report, voice-handoff-latency, voice-latency-deep-dive-2026-08

Canonical: adr-0116-vendor-job-reference, turnover-architecture-2026-08-01, vendor-identity-architecture-2026-08-12

Canonical: collections-three-phases, collections-phase-tracker, colorado-collections-law

Canonical: renewal-thread-escalation-findings, renewal-outreach-incident-2026-07-31, camellia-market-rent-drop-2026-07-30

Canonical: email-architecture-2026-08-14, email-silent-drop-incident, email-intent-contract-2026-08-17

Canonical: knowledge-base-redesign-2026-08-26, property-knowledgebase-2026-08-08, docs-quickstart

Canonical: household-record-go-no-go, identity-split-audit-2026-08-14, cross-property-leak-audit-2026-08-13

Canonical: post-approval-gap, approved-to-signed, lease-answers-matrix-2026-08-09

Canonical: appfolio-sync-surfaces, appfolio-renewal-benchmarks, blocked-a1a7405d

Canonical: sales-machine-roadmap, sales-machine-card-ia-2026-08-18, prospect-report

Canonical: detail-page-standard, dashboard-history-architecture, prospects-funnel-redesign

Canonical: clara-coworker-map-2026-08-19, clara-architecture-2026-08-18, one-clara-architecture

Canonical: architecture-source-of-truth, calendar-architecture-2026-08-08, propflow-cards-and-cash-2026-08-02

Canonical: fair-housing-architecture-2026-08-19, opt-out-scope-2026-08-16, compliance-audit-2026-08

Canonical: clara-loop-latency-2026-08-02, latency-audit-2026-07

Canonical: tour-decider-architecture, tour-decider-v2-camellia-replay-2026-08, voice-tour-availability-decision

No reconciliation was supplied for this cluster. Recommend a manual 10-minute triage pass before the next mining run to confirm these 2 pages are genuinely one-offs (not a 21st topic hiding in the "misc" bucket) and either fold them into an existing canonical page or give them their own.

---

1. A page from Aug 25 asks Fede to turn the pre-send honesty check back on without mentioning his Aug 20 "no."
2. The Aug 4 "proven clean against 730 conversations" escalation-scoping claim didn't hold — a real case 17 days later shows one unrelated question froze someone's whole account.
3. The PM-reminder office-hours fix was marked fully closed (Jul 28) but the same bug recurred 3 weeks later from an unaddressed root cause.
4. Two escalation-tone pages are dated 3 days apart but describe rulings as coming from the same source — worth a date check, doesn't change the substance.

1. The repo's own CLAUDE.md contains two sections that directly disagree on whether merging to main needs Fede's explicit approval or just a green CI check.
2. One page declines to fake a green security tile (consistent with the "never fake green" doctrine) while several other pages describe the opposite failure (checks lying green) as the thing to fix — not contradictory, but never stated as one shared rule.
3. Two independent AI review passes gave opposite recommendations purely from the order options were listed, on at least 4 separate occasions — a recurring, never-fixed reliability problem with the review tooling itself.
4. A session merged a CLAUDE.md change removing the "humans must merge CI workflow file changes" rule, citing a Fede approval nobody can find any record of — while the rest of the system treats that rule as inviolable; per the records, never confirmed resolved.

1. A page recommends silencing 2 failing nightly alerts with no fix in hand — runs against Fede's own same-week rule that a switch used to quiet a real bug gets reverted, not kept.
2. The July fix for silent, unalerted nightly failures evidently didn't reach everywhere — a different pair of checks hit the identical trap a month later.
3. Sonnet 5's specific safety complaints (privacy leakage, silent empty replies) were never explicitly re-confirmed or retracted after a later, bug-fixed comparison found no significant difference.
4. One page recommends leaving a harmless accidentally-merged "humans must approve" change standing because it turned out fine — the opposite reasoning ("a gate exists for what COULD go wrong, not how it turned out") was used 4 days earlier for a near-identical case.
5. Two parked-decision pages are word-for-word the same question, down to internal receipt IDs, re-surfaced 2 weeks apart because nobody answered the first one.

1. The transfer-line wording is pulling 3 directions: the accepted ruling (one fixed platform line) vs. what shipped (11 custom lines) vs. a stuck PR trying to fix it that picked wording ("One moment") already ruled out as sounding like a stall.
2. A page was still waiting on approval to merge a safety fix for the faster AI model 2 days before that model got dropped entirely, making the ask moot without anyone marking it so.
3. A page recommended trying the warmer voice setting behind a controlled A/B test; what shipped was a live, untested rollout for ~8 minutes that caused dead air and had to be pulled.

1. One page's internal text calls the same decision "(proposed)" and then, later in the same page, "Accepted" — a staleness slip inside one page, not a real cross-page dispute.
2. A page names an AI's lost run-to-run consistency as a still-open problem, while a same-day fix works around that exact root cause via a 3-way vote without ever stating outright that the underlying setting itself is still broken.

1. A safety audit says texting isn't safe to launch (6 critical defects, zero contacted) while a later page reports a real resident was successfully texted with those same defects already fixed — a reader hitting only the audit would think collections texting never launched.
2. The "why is every account the same stage" symptom was first framed as a launch-blocking bug, then correctly re-diagnosed as a normal data-availability gap (parsers too new) — can't both be the operative diagnosis; the later, measured page wins.
3. Quiet-hours checks were decided one way, then reversed by Gera on 2026-08-25 after a real capture correctly refused to send under the new rule — only the phase tracker carries the reversal; the older decision-log entry alone gives the wrong current answer.

1. The same open rent-pricing decision (unit 607) got parked twice, 3 days apart, with two different recommended numbers ($1,075 vs. $1,000), neither checked against the other.
2. One page calls the frozen-offer behavior "intentional, accepted design"; another calls the one real instance of it a "gap worth a safeguard" — not mutually exclusive in the fix, but the two pages characterize the identical mechanism very differently.

1. A page tagged "email ingest" (session-mined-pipeline-defects) is actually about the engineering team's own PR-review pipeline — a different "pipeline" that just shares the word.
2. A page's own tracking record says a rule is a held, not-yet-live draft, but the same page states elsewhere that the rule already went live at 2 properties.
3. "Silent drops are structurally gone" (Jul 25) was true only for the specific bug it fixed — 8 days later a different mechanism silently ghosted a live person mid-tour-booking.
4. The DMARC-tightening timer is stated with two different starting clocks 8 days apart, neither page referencing the other's.

1. A newer page (Aug 26) says one-off favors should never be taught to Clara at all; an 8-days-older page proposed the opposite (capture and flag them) — nothing formally retired the older proposal even though the newer one contradicts it.
2. Camellia's move-in special is being fixed twice in two different layers (a one-property switch vs. a general 3-places-store-the-same-number architecture fix) with neither page acknowledging the other — if only one ships, the special can still be misquoted through whichever path the other didn't cover.
3. The docs-quickstart page's "no PR, no review, no CI gate" promise is true for adding a content page but not for changing the publishing tool's own code (which needs a PR + manual merge every time) — not a bug, but easy to misread as covering both.

1. Two household-modeling design pages still carry a live "Proposed — pending review" banner and list their questions as open, but a same-date/next-morning page shows every one of those decisions was already made and shipped.
2. Calls from an unrecognized number "fail open" (Clara guesses a property); texts from an unrecognized number "fail closed" — the identical situation handled two opposite ways depending on channel, unreconciled by any page in this batch.

1. A ship report (Aug 9) says Clara's lease answers are "proven correct, never invents numbers"; the very next day's test grid finds Clara stating a $200 charge as fact when it wasn't on file, and emailing new-lease pricing to a current resident 9 times.
2. A team-notification email is presented as ready to ship (wording sign-off only) the same day a different page rules it should ship only after a broader safety-test baseline exists.
3. The post-approval design page scoped the feature to approved applicants only; the shipped code answered any prospect — resolved a day later, but the original page never says so.
4. Fede and Gera remember opposite rules for the same policy inside the same page: Fede's own July quote says always send the application link when asked; Gera says a tour must happen first.
5. A separate page (outside this batch) already decided to stop routing the application link through a special "send" step, but isn't referenced by the policy page whose outcome that decision affects.

1. One page treats the work-order auto-refresh fix as unshipped, awaiting a yes/no to even open it for review; a same-day page shows screenshot evidence it's already merged and running on schedule.
2. A renewal-automation benchmarks page's header claims an August edit date, but every number on it is timestamped April — it hasn't actually been regenerated in 4 months.

1. A design page still lists "which panel option (A/B/C/D)" as an open decision, but a different, later, approved design shipped instead — nobody ever picked A/B/C/D, so the question is stale, not pending.
2. A handoff page lists the sales-app chat feature as "not started, decision pending," while a same-day architecture page records Fede already picking the answer live that evening.
3. A page flags getting Clara logged into a RealPage-run system as "for discussion, not decided," while a same-day page is already building exactly that for a real customer.

1. A same-day detail-page plan lists 2 visual questions as open (keep the gradient banners? switch the progress-bar color?) that 2 shipped PRs, from the very same day, had already answered.
2. A dashboard plan says a chart needs a brand-new data backfill; a same-day page shows the needed history already exists and just needs the chart repointed to it.
3. A page treats "what should the funnel count" as one all-or-nothing decision still waiting for a founders' call; what actually shipped is a hybrid that was never brought back to that call as the proposed resolution.

1. A reference doc claims phone calls already answer the same way as text/email; two other pages (one 5 days older, one updated the same day as this reconciliation) show that's not true yet — 11 separate call scripts share no wording with the text rulebook, and calls don't get the same promise-tracking safety check text does.

1. A calendar-engine page is tagged fully decided/accepted, but the page itself still says "Proposed — pending Fede's review" at both the top and bottom, and lists 4 real unresolved choices inside it.
2. One section of a doc says a fair-housing check on staff answers is "already live," while a tracker table further down lists a related-but-different jurisdiction-review gate as "not started" quoting an old ruling — two different checks that read, on a quick pass, like a self-contradiction.

1. A checkpoint's own description ("covers every reply Clara is about to send") is out of date — a later, code-verified audit shows it only covers prospect-facing leasing replies; maintenance texts, marketing, resident replies, and calls all skip it.
2. An opt-out-scope page marks Option B as the picked/shipped option, but the one paragraph explaining why sits under Option C's card and argues Option-C-specific points — likely a formatting slip, worth a quick confirm that B is really what shipped.

1. A report's own priority ranking (millisecond timing) named its 2 biggest wins as brand-new fixes; a follow-up using request-counting proved both were already fixed by unrelated changes before anyone touched them on purpose.
2. A report assumed the browser back button always reloads a page from scratch and proposed a fix for that; a follow-up proved browsers already serve back/forward instantly from memory, so the assumption was simply wrong.
3. Fast-but-empty-looking code was flagged "dead, safe to delete"; when someone tried, it turned out to have live dependents (thousands of test records, 6 tests, 2 other features) — a real feature on unpopulated data, not dead code.

1. A decision page still shows "proposed — pending review" with 3 unresolved options, but the option it recommends (delete the old system) is exactly what was built and shipped 2 days later.
2. An Aug 11 ruling said "keep things as-is, only revisit via a specific parked redesign if this bug recurs"; the bug did recur (Aug 15), but a different fix than the one promised got built — nobody updated the Aug 11 page to say its own trigger fired and its planned next step wasn't the one taken.
3. Two same-day pages about the same shipped tour-decider work recommend two different next steps (test-property trial first vs. skip straight to the real customer) without acknowledging each other. ---