Goal 3
Complete
100-row calibration independently reviewed
DanubiSoft internal control page · 3 September 2026
What is done, what is blocked, and what should happen next
This page replaces the scattered worker sessions as the single review surface. It focuses on the remaining decisions and contains the exact continuation prompt to annotate.
How to review
Click a text line to suggest a rewrite. Command-click or Control-click any object to leave a comment. The most useful place to annotate is the continuation prompt near the bottom.
Goal 4
Not complete
First 100 evidence rows finished, zero admitted
Source pool
1,307
Unique website-backed candidates before semantic and contact gates
Spend
$6.1914
Actual Goal 4 provider usage, $33.8086 remains under the $40 cap
Executive summary
The method is validated. Production is not.
The key mistake to avoid is treating completed extraction as completed leads. The first 100 are evidence envelopes, not campaign-ready contacts.
Goal 3 calibration
- 100 stable rows, independently reviewed 100 PASS and 0 REVISE.
- 76 GoodFit and 24 unknown or excluded.
- Manifest SHA-256: d5794352952c4045dcb54637519f3a940f519178bf73b5b78102d39c29f53a2c
- Private editable review behavior was restored and browser-tested.
- Goal 4 admission controller passed its earlier security review and the current DanubiSoft suite now reports 360 tests passing.
The actual 1,000-lead deliverable
- No fresh company is admitted yet.
- No first-100 independent semantic review has been completed.
- No named primary contact and verified email package has been completed for the first 100.
- No fresh suppression, final deduplication, and admission pass exists for those rows.
- No editable first-100 production review table or final 1,000-row review table exists.
- No final browser acceptance or client delivery has happened.
First fresh 100
Why the current output is not ready for your review
The extraction job completed successfully, but most unknown rows are technical false negatives that need a controlled correction pass.
100 of 100 complete
All websites were processed and all 100 extraction envelopes were written. The coordinator exited with code 0.
6 GoodFit · 1 NotGoodFit · 93 unknown
These counts are not independently reviewed and must not be used as production yield.
77 validator failures
Many rows became unknown because of strict identity, citation, or retry-format failures. They are not 77 confirmed bad companies.
23 people · 9 website emails
These are raw website extractions only. They still require current-role, exact person-email, verification, suppression, and independent-review gates.
Zero sendable rows
No Instantly upload, campaign mutation, activation, send, public share, client Sheet write, or lead contact occurred.
Usage-limited
The executor hit the shared Codex usage limit after finishing the 100 evidence rows. Its session is now archived with all work preserved.
What to review
Spend your attention on three things
My recommendation is to skip raw technical artifacts now. They are evidence for the executor, not useful decision material for you.
1
Annotate the continuation prompt
Review the scope, sequence, spend rules, and final acceptance criteria. Rewrite any line that does not match how you want the next agent to operate.
2
Confirm the recommended review timing
Do not manually review the current raw 100. Wait for the corrected, independently reviewed, editable first-100 table. That table should contain GoodFit, explanation, personalization, named contact, email evidence, verification, and source links.
3
Review the corrected first 100 when it exists
Check every row for three practical questions: would you contact this company, does the personalization sound natural and specific, and is the named person the right recipient? The executor must stop at this approval gate. If you request changes, it must refine the rules and prepare a different fresh 100-row sample for another review round.
Recommendation
Use this execution sequence
This keeps technical recovery ahead of scale and makes each later batch cheaper, faster, and less error-prone.
01
Preserve and normalize the current devbox state
Resume from the archived executor worktree and existing branch. Recheck the 360-test suite, isolate the 150 task-owned status entries, create one clean checkpoint commit, push it, and bind every continuation to that commit.
Your review: none. This is technical recovery.
02
Close the two technical release gates
Release 3 must reject GoodFit without manufacturing-infrastructure evidence and unquoted invented person or role claims. The new 979-row Maps provenance contract must replace trusted free text with deterministic reason codes and pass a fresh independent attack review.
Your review: none unless a real product-rule decision appears. Do not weaken the GoodFit or evidence rules.
03
Correct and fully enrich the first 100
Run a separate immutable correction pass only on the 77 bounded-invalid rows and preserve every original result. Complete named-contact evidence, finder fallback, email verification, suppression and deduplication, then independently review all 100 final rows and freeze an exact manifest and hash.
Your review: none until the complete editable table is ready.
04
Run the 100-row approval loop
Open the private editable table for your review. If you request changes, the orchestrator must convert your annotations into versioned rule and prompt updates, regression-test them, and prepare a different fresh 100-row sample. Repeat until you explicitly approve the method for scale.
Your review: GoodFit, explanation, personalization, named recipient, email evidence, and any row-level correction. The task stops in WAITING_FOR_APPROVAL after each sample.
05
Scale only after your explicit approval
After you approve a specific version of the rules and prompts, process deterministic, non-overlapping batches with stable IDs and receipts. Measure real GoodFit and verified-contact yield before spending on more source acquisition.
Your review: telemetry at meaningful yield, spend, or quality changes. No further per-100 approval is required once you approve scale.
06
Freeze exactly 1,000 and open the final review table
Every accepted company needs one verified named primary contact, supported personalization, current suppression clearance, unique identity, deterministic QA, and independent review. Open the final private comment-enabled table in your Codex browser and verify edit, comment, reload, export, and restore.
Your review: this is the final 1,000-row acceptance surface. Instantly and campaign actions stay out of scope.
Annotate this
Recommended continuation prompt
Each line is independently editable and commentable. Once you finish, the annotated page is the clean handoff for the next execution session.
Resume the DanubiSoft 1,000-company review goal from the preserved devbox state. Do not redefine success and do not restart completed acquisition.
AUTHORITATIVE STARTING STATE
- Worktree: /home/matt/mirror/ColdEmail-danubisoft-1000-review
- Branch: codex/danubisoft-1000-review-20260902
- Current pushed HEAD: 01e5e505bfbab5270b3118b3b87e2c4e5de0a242
- Run ledger: outputs/loop-runs/20260902T232007Z-danubisoft-1000-review/
- Archived predecessor executor: 01a06471-b741-7d62-bea4-23406a6c164d
- Goal 3 is complete: 100-row calibration, 100 PASS, 0 REVISE, manifest SHA-256 d5794352952c4045dcb54637519f3a940f519178bf73b5b78102d39c29f53a2c.
- Goal 4 is incomplete: source capacity is 1,307 unique website-backed candidates, first-100 scraping and extraction are complete, admitted count is zero.
- Goal 4 spend is USD 6.1914 actual, USD 0 reserved, USD 33.8086 remaining under the existing USD 40 cap.
FIRST RECOVERY CHECKPOINT
- Read the current run ledger, receipts, independent reviews, questions, orchestrator registry, Git state, and provider spend before changing anything.
- Recheck the full DanubiSoft test suite. It last passed 360 tests.
- Preserve unrelated dirty work. Reconcile the current task-owned status, create one clean checkpoint commit, push it, and record exact restore and remote receipts.
- Never silently replace or mutate release-2 evidence. Every correction must be a new immutable revision with preserved originals.
CLOSE THE TWO TECHNICAL GATES
- Release 3 remains blocked until outputs/loop-runs/20260902T232007Z-danubisoft-1000-review/reproduce_release3_integration_gaps.py rejects both cases: GoodFit without manufacturing-infrastructure evidence, and an unquoted invented person or role.
- After fixing those cases, rerun focused tests, the full suite, and a fresh independent attack review. Write a formal hash-bound release receipt before any release-3 evidence batch starts.
- The 979-row Maps provenance package remains closed until trusted free text is removed from machine-consumed provenance. Use fixed reason-code enums with deterministic counts, keep human notes outside the trusted contract, rerun all hidden-token and nested injection attacks, and obtain a fresh independent PASS.
MAKE THE FIRST 100 REVIEWABLE
- Freeze the completed release-2 first-100 input and results. Current provisional counts are 6 GoodFit, 1 NotGoodFit, and 93 unknown. Seventy-seven unknown rows are bounded validator failures, not confirmed bad companies.
- Run a separate immutable release-3 correction pass only for bounded-invalid rows. Preserve every original result and source hash.
- Independently review all 100 final dispositions. No row may inherit a positive decision, person, role, email, or personalization claim that lacks exact source evidence.
- Produce a private editable 100-row review table with stable row ID, company, domain, GoodFit decision and explanation, personalization sentence and variables, named contact, role, exact email evidence, verification result, source links, raw website Markdown path, and review status.
- Open that table in Matt's Codex browser and verify comments, inline rewrites, reload, export, and restore. Stop in WAITING_FOR_APPROVAL. Do not begin the scale phase.
MATT'S 100-ROW APPROVAL LOOP
- Matt reviews GoodFit, explanation, personalization, named recipient, email evidence, and any row-level correction in the editable table.
- If Matt requests changes, the orchestrator converts his annotations into versioned rule and prompt updates, runs regression checks, and prepares a different fresh 100-row sample with the same complete evidence and review table.
- Repeat this review loop and stop in WAITING_FOR_APPROVAL after every sample until Matt explicitly approves a specific rule and prompt version for the remaining scale.
CONTACT AND ADMISSION RULES
- Prefer current named people and exact person-email relationships published on the company website.
- Use an approved finder only for unresolved companies and only when the immutable receipt binds the same current person, company, role, domain, and exact email. Recheck provider access and credits before use.
- Use MillionVerifier only as deliverability evidence. It does not prove identity or role.
- Replay suppression from every mandatory current and historical source immediately before admission. Ordinary outreach performed for another client is not an exclusion, but explicit opt-outs and blocklists are.
- Deduplicate against the Frozen 100, all DanubiSoft history, current Instantly membership, within-run place IDs, domains and emails, and every previously accepted company.
- Keep every row admitted=false and sendable=false until the full admission contract and independent review pass.
SCALE TO EXACTLY 1,000
- Begin only after Matt explicitly approves a specific version of the GoodFit, scraping, personalization, contact, and verification method for scale.
- Process the unused source candidates in deterministic non-overlapping batches using only that approved method. Keep batch hashes, rejections, QA, spend, and independent-review receipts.
- Measure real GoodFit, named-contact, and verified-email yield before any further paid source acquisition. The 1,307 source candidates are capacity, not 1,307 accepted leads.
- If the accepted yield proves insufficient, acquire only the additional attributable Hungarian manufacturing companies needed for the remaining gap, within the existing spend cap and with exact provider settlement receipts.
- Continue until exactly 1,000 unique companies have one verified named primary contact each, evidence-grounded personalization, current suppression clearance, deterministic QA, and independent all-row review.
FINAL ACCEPTANCE
- Freeze the 1,000-row manifest, source and output hashes, rejection history, verification and spend receipts, zero-send proof, QA report, and independent-review receipt.
- Open the final private comment-enabled 1,000-row HTML table in Matt's Codex browser at 1280 by 900 and verify comments, edits, reload, export, restore, source links, and responsive behavior.
- Do not upload to Instantly, create or alter a campaign, activate, schedule, send, contact leads, publicly share, write client-facing Sheets, or deliver to the client.
- Keep the goal active until the final table and every acceptance receipt are verified. If a genuine human-only gate remains after all safe recovery routes are exhausted, stop with the exact missing evidence and resume point.
Housekeeping
Ten related sessions were archived
The current parent task remains the only active discussion surface. Archiving is reversible and no session history was deleted.
Show archived session IDs
Local: 019fccc5-d2ff-7a31-9e50-c11374b92b4a
Local: 019ff7fa-bcde-7e73-b546-c827b79efba0
Local: 01a018fa-aee1-7450-a6c8-7cd38bd8aca2
Local: 01a02189-55a9-7071-9533-d0c3a1661f9e
Devbox: 01a051c0-5aa0-7bb2-a3ee-684724b7c4e4
Devbox: 01a051c1-8d62-7100-9a87-5a1aa6df776f
Devbox: 01a051c2-1c0c-7fd1-8921-1cbf72324f09
Devbox: 01a051c2-91dd-7933-aa01-9cfe5b02307d
Devbox: 01a05689-b62c-73d2-9e14-009221e4d013
Devbox: 01a06471-b741-7d62-bea4-23406a6c164d
Show technical evidence references
Devbox worktree: /home/matt/mirror/ColdEmail-danubisoft-1000-review
Goal run: outputs/loop-runs/20260902T232007Z-danubisoft-1000-review/
First-100 completion: clients/danubisoft/ops/campaign-readiness-2026-09-02/lead-review-1000/acquisition/evidence-maps-balanced-001/coordinator-completed.json
Release-3 blocking proof: outputs/loop-runs/20260902T232007Z-danubisoft-1000-review/executor-release3-gate-v1.json
Maps provenance review: outputs/loop-runs/20260902T232007Z-danubisoft-1000-review/maps-provenance-v2-independent-review-v1.json
Spend summary: outputs/loop-runs/20260902T232007Z-danubisoft-1000-review/provider-receipts/spend-summary.json