Skip to content

getaero-io/gtm-eng-skills / engineering/warm-intro-scoring

warm-intro-scoring

Verified

FileKindSize
.gitignoreother121 B
GUIDE.mdmarkdown4.7 KB
LICENSEother1.1 KB
README.mdmarkdown8.0 KB
SKILL.mdskill md327 B
TESTING.mdmarkdown8.9 KB
UPSTREAM.mdmarkdown1.3 KB
assets/company-research.play.tsscript1021 B
assets/customer-db.sqlother2.7 KB
assets/evaluation-example.jsonconfig4.6 KB
assets/example.jsonconfig6.9 KB
assets/portfolio-focus.weights.jsonconfig32 B
assets/review.htmlother7.4 KB
assets/tuning-example.jsonconfig1.2 KB
assets/tuning.htmlother30.1 KB
check_all.pyscript1.4 KB
package-manifest.jsonconfig9.8 KB
plays/README.mdmarkdown18.2 KB
plays/check-report.pyscript1.8 KB
plays/check_all.pyscript1.6 KB
plays/check_parity.pyscript6.0 KB
plays/collect.play.tsscript4.0 KB
plays/collection.tsscript1.9 KB
plays/core.tsscript44.2 KB
plays/evaluation-contract.mdmarkdown3.4 KB
plays/evaluation.tsscript13.8 KB
plays/features.play.tsscript4.7 KB
plays/features.tsscript22.6 KB
plays/generatedtemplate.tsscript38.1 KB
plays/guardrails.tsscript948 B
plays/legacy.tsscript9.7 KB
plays/merge-payloads.pyscript8.6 KB
plays/parity-driver.tsscript995 B
plays/parity-evaluation.pyscript8.5 KB
plays/parity-reports.pyscript2.3 KB
plays/prepare-inputs.pyscript6.5 KB
plays/quality.tsscript3.8 KB
plays/read-sources.play.tsscript3.4 KB
plays/report-cli.tsscript590 B
plays/report.play.tsscript2.9 KB
plays/report.tsscript3.1 KB
plays/reports.tsscript5.7 KB
plays/research.play.tsscript1.6 KB
plays/review-uncertainty.play.tsscript5.5 KB
plays/score.play.tsscript2.3 KB
plays/snapshots.play.tsscript4.2 KB
plays/strict-json.tsscript1.8 KB
plays/sync-template.pyscript849 B
plays/uncertainty.tsscript4.9 KB
plays/validation.play.tsscript3.6 KB
references/adversarial-review.mdmarkdown7.0 KB
references/batch-collection.mdmarkdown5.1 KB
references/company-network.mdmarkdown5.8 KB
references/competitive-review.mdmarkdown8.8 KB
references/evaluation.mdmarkdown8.5 KB
references/evaluator-validation.mdmarkdown3.4 KB
references/evaluator.mdmarkdown7.9 KB
references/external-data.mdmarkdown4.0 KB
references/factor-test-matrix.mdmarkdown4.6 KB
references/factors.mdmarkdown6.1 KB
references/intro-drafts.mdmarkdown1.6 KB
references/pipeline.mdmarkdown7.6 KB
references/readable-output.mdmarkdown5.4 KB
references/research-brief.mdmarkdown2.4 KB
references/sources.mdmarkdown7.5 KB
references/tuning.mdmarkdown14.6 KB
scripts/audit_history.pyscript5.1 KB
scripts/check.pyscript5.1 KB
scripts/compare_versions.pyscript1.9 KB
scripts/evaluate.pyscript22.5 KB
scripts/package-skill.pyscript1.9 KB
scripts/portfolio.pyscript3.7 KB
scripts/quality.pyscript5.3 KB
scripts/recipe-demo.tsscript2.6 KB
scripts/render.pyscript1.6 KB
scripts/score.pyscript9.6 KB
scripts/temporal.pyscript2.2 KB
scripts/test_ask_handoff.pyscript2.1 KB
scripts/test_audit_history.pyscript2.7 KB
scripts/test_handoff_evaluate.pyscript9.1 KB
scripts/test_portfolio.pyscript3.6 KB
scripts/test_quality.pyscript3.1 KB
scripts/test_temporal.pyscript1.3 KB
scripts/test_tuning.pyscript14.1 KB
scripts/test_tuning_browser.cjsscript7.7 KB
scripts/test_version_parity.pyscript1.0 KB
scripts/tuning.pyscript10.9 KB
tests/adversarial.test.tsscript8.2 KB
tests/browser_e2e.cjsscript1.9 KB
tests/cloud_e2e.pyscript6.9 KB
tests/e2e-fixture.jsonconfig3.6 KB
tests/feature_parity.pyscript2.4 KB
tests/features.test.tsscript8.3 KB
tests/test_prepare_inputs.pyscript8.6 KB
tests/uncertainty.test.tsscript2.5 KB
vendor-manifest.jsonconfig748 B

references/adversarial-review.md

View on GitHub
# Cached pipeline: end-to-end and adversarial review

Reviewed with Claude Code CLI on 2026-09-23/24. This is a functional and adversarial review, not a prediction-accuracy evaluation. No new paid profile or public-post collection was needed.

## Reproduce

Run the offline suite from this package:

```sh
python3 plays/check_all.py
```

The opt-in cloud harness writes fictional test revisions and a private review to the selected Deepline workspace. Run preflight separately and confirm the organization before starting. Use a new output directory; credentials use the caller's CLI scope.

```sh
deepline preflight --json
python3 /absolute/path/to/tests/cloud_e2e.py --out /private/new-test-directory
PLAYWRIGHT_MODULE=/path/to/playwright node tests/browser_e2e.cjs /private/new-test-directory/review.html /private/browser-receipt.json
```

The harness runs cached collection → normalized profiles → immutable DB snapshot → DB read-back → feature extraction → exported payload merge → tuned scoring → cloud report → independent file read-back. It checks all input and output identities. The fixture has five contacts and four paths: investor 120, coworker 80, employer only 20, sparse 0. Increasing work weight to 160 produces 160, 120, 40, 0 and preserves a declined investor path. Browser checks cover the top three, the blocked state, selected weights, mobile layout, script errors, and network requests.

## Corrections made after adversarial review

- Invalid or duplicate connector identities fail closed. Known LinkedIn language/detail URLs are normalized. Member-ID URLs need a reviewed vanity alias; they are rejected until resolved. Stable equal person IDs exclude self-pairs.
- Investor-facing employee titles do not qualify as direct investment roles.
- Investor-employer links require both reviewed investment evidence and a current, dated role at `investor_company_id`. Company names alone cannot produce portfolio points.
- Cached profiles retain the earliest supplied original receipt date. A missing original date fails. Cached collection now produces normalized profiles. `account_domain` is an explicit reviewed account assignment, not inferred from a person's name.
- Duplicate company sizes, reversed dates, invalid months, duplicate JSON keys, nonfinite numbers, and dangling feature-catalog references fail.
- Declines survive missing/unresolved refreshes in the exported review-state ledger. A changed feature set invalidates a prior ready state. Keep this ledger with the next refresh.
- `--source-input` reconciles target export keys against the original input. Without it, completeness is explicitly unverified.
- Reports show unresolved targets and unmatched prior review states. Cloud and local TypeScript reports require an evidence registry. CSV text is escaped against spreadsheet formulas.
- Location matching no longer treats similarly named cities outside California, or ambiguous multi-location text, as one verified Bay Area location.

The added adversarial regression suite contains 20 TypeScript cases; together with the 12 existing feature cases, all 32 pass. Seven Python packaging, refresh-ledger, export-reconciliation, and CSV safety cases pass. The full offline suite also passes.

The original migration matched the frozen Python output. Correctness fixes deliberately change source-derived features. On the reviewed full snapshot, all 311 targets and 91,428 paths remain. 323 baseline scores decreased, none increased; 240 city-overlap values and 81 investor-portfolio values were removed. Some paths have more than one correction. These changes must not be described as unchanged-output parity. Historical source receipts and individual differences remain private.

## Remaining prebuilt release gates

- Live provider contracts, charges, asynchronous actor recovery, and partial profile/posts failures have not been exercised by this cached test. Live collection still requests profiles and posts for supplied rows; it is not an automatic gap-only refresh service.
- Profile age is visible through observation dates but there is no calibrated relationship-decay policy or enforced refresh threshold. Historical coworker evidence can still rank highly; it does not imply a current personal relationship.
- An evidence ID and a source locator do not prove source truth or subject relevance. Arbitrary caller-supplied features need a trusted, reviewed evidence-binding layer. The package does not independently verify every cited claim.
- The prebuilt needs tenant/requester-scoped persistent willingness and deletion handling. The local review ledger is not a substitute for reading the latest DB willingness records. Preserve stable canonical identities across imports.
- Automatic source-table mapping, identity-alias resolution, queue/leases, affected-pair triggers, per-run budget enforcement, and recovery orchestration are not enabled.
- Bad shared source records fail a target batch instead of entering a complete per-record quarantine/recovery workflow. Resolve or isolate invalid inputs before retrying.
- Large reports still use local assembly; cloud output has a 1 MB cap. Database page reads are not transactionally consistent during source updates.

Status: the bounded cached pipeline is testable and passes. A fully automatic, unattended prebuilt is not release-ready yet. No messages, consent updates, or schedules are enabled by these checks.

## PR refresh review — 2026-09-24

Two independent read-only reviews covered linked profiles, report rendering, employee coverage and refresh holds. Reproduced findings and changes:

| Case | Result |
|---|---|
| Old ready state plus new decline | Fixed: the decline wins. |
| Same features and IDs, changed source detail/date | Fixed: readiness hashes include referenced source records. Old hash versions reset readiness. |
| Source path count differs from exported paths | Fixed: merge rejects inconsistent counts. Coverage notes and per-export counts survive. |
| All target keys exported, partial employee roster | The status now says `target_exports_verified`; it makes no employee completeness claim. Roster completeness still requires provider receipts and reconciliation. |
| Connector shards repeat the same target key | Unsupported and rejected. Documentation no longer recommends this merge path. Use local TypeScript extraction when shared source context exceeds 5,000 records. |
| Unsafe profile URL or HTML in a name | URL validation rejects unsafe destinations; names render as text. Linked profiles use a new tab and `noopener noreferrer`. |

Regression cases are in `tests/test_prepare_inputs.py`, `tests/features.test.ts`, `scripts/test_tuning.py` and `scripts/test_tuning_browser.cjs`. No live employee pull or introduction send was performed. No independent source-truth or conversion claim is made.

The reviewed PR head also had a failing repository CI test: `tests/lib/integrations/fullenrich-v2.test.ts`, “keeps UNKNOWN results pending in the synchronous async-job handler” (`initialUsageUsd` was undefined). That file is outside this recipe's diff. Do not describe the entire PR as green on the strength of offline recipe checks.
slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 196kUpdated Claude CodeCodexVerifiedInstall