You may not know it yet 鈥?br>your know-how can be confirmed, and it can pay.

鈶?Extract

Settled out of real work: an agent's real production run leaves evidence, and the reusable lesson is extracted from that evidence 鈥?nothing invented, nothing padded.

鈶?Confirm

Card-by-card review, then the founder signs off in person. Every card carries its own SHA-256 on record; independent verification passed card by card before anything was ingested.

鈶?Ingest

Dual-library ingestion: the authoritative record goes into the library, revenue-share terms are set and on record (personal cards 70/30, collaborative cards 50/50; agents take no cut), and the founder keeps the right to withdraw any card.

CWF-001 路 Production replay: write the criteria first, then correct self-reports with raw evidence

Define rollback conditions and acceptance criteria first; when a script's own account conflicts with the raw logs, the raw logs win. Insufficient evidence stays N/A 鈥?never padded into a pass.

Collaborative 路 public

CWF-002 路 Separation of discovery and disposition: monitors only alert, adjudicators decide the action

Collaborative 路 internal grade 鈥?content not publicly displayed

CWF-003 路 From single-point fixes to immunity against the whole class

Separate the specific bug from the failure pattern; old incidents become regression samples, and detection earns its keep by catching a new case it was never tuned for.

Collaborative 路 public

CWF-004 路 Deployment closure: "changed" is not "in effect"

Probe and back up before changing; after deploying, prove the running process, the version on disk, and the original state. Effect must be proven by the running surface, not by first impressions.

Collaborative 路 public

CWF-005 路 Retiring mechanisms: let the business loop audit the governance

Personal (founder) 路 internal grade 鈥?content not publicly displayed

CWF-006 路 Trust boundary: even a "useful feature" without consent must be blocked 鈥?and disclosed

The user's right to know and to withdraw beats growth, telemetry, or convenience. "No harm done yet" does not downgrade the nature of the act 鈥?disclose first, before the outside finds out.

Collaborative 路 public

CWF-007 路 Number reporting: run ledgers for discovery, official ledgers for verdicts

Run logs surface anomalies; every number carries its source and confidence grade, and nothing becomes a financial conclusion before official reconciliation.

Personal (founder) 路 public

CWF-008 路 Use a moving baseline to catch boiling-frog drift

Rolling baseline plus dual thresholds; alert only on sustained anomaly, then point back to the raw metric. Chronic drift must be visible 鈥?without single-point jitter becoming noise.

Collaborative 路 public

CWF-009 路 Retries must be bounded by both attempt count and total task budget

Failure may be retried, but never drained without limit. Block at either cap, record the reason 鈥?an honest block beats a time promise you cannot keep.

Collaborative 路 public

CWF-012 路 Split deployment and observation shifts 鈥?never make the agent wait on standby

Deploy, hand over a verifiable state, stop. Let the server accumulate observations naturally; waiting and polling are cost risks, not work.

Collaborative 路 public

CWF-013 路 When a system reroutes the user, it must explain, batch, and report

Even a correct auto-optimization must never be silent: every deviation from the user's choice is logged as "original choice 鈫?new choice + reason", aggregated, and escalated past a threshold.

Collaborative 路 public

CWF-014 路 A metric inside the system is not yet a governed metric

A number on a dashboard that triggers no judgment is not a defense line. Give each key metric a threshold with business meaning, wire the alert, and replay history to test it.

Collaborative 路 public

CWF-015 路 Isolation must have a time exit and a minimum sample bar

Check the minimum sample before convicting; write the expiry into the isolation itself, and let time release it. A protection mechanism must never become a permanent punishment.

Collaborative 路 public

CWF-016 路 Before quota governance, hold the line on attribution integrity

Collaborative 路 internal grade 鈥?content not publicly displayed

CWF-017 路 No data? First ask: no business, or a broken pipeline

Silence is not an automatic error: cross-check process, port, real requests and write ability, and let a minimal probe tell zero traffic from a broken channel. Even a misdiagnosis gets filed.

Collaborative 路 public

CWF-018 路 A hanging write? Check the shared queue before blaming permissions

When many writes hang at once, look for the shared resource first: read-only queue inspection, find the long-running blocker, unblock with authorization and a minimal write probe 鈥?never blind retries.

Collaborative 路 public

CWF-019 路 Intelligence gets exactly one authoritative outlet

Having alerts but reading them in the wrong place is worse than having none. Raw events may be many; the authoritative alert box is one, and bypass writes must verify down to zero.

Collaborative 路 public

CWF-020 路 The monitoring system itself must be watched by an independent observer

The monitored must never self-certify health. An external heartbeat checks both data freshness and process state 鈥?absence of alerts is not proof of life.

Collaborative 路 public

CWF-021 路 Authoritative files must carry a state machine

A signature is not the same as implemented. Every rule gets an explicit state 鈥?draft, signed, implemented, enforced, revoked 鈥?and consumers read by status, never by filename.

Collaborative 路 public

CWF-022 路 On delivery, say which layer the defense line lives in

"Works in our environment" is not "ships with the product". Label every claimed capability by layer, test the actual artifact, and shrink the claim rather than stretch the promise.

Collaborative 路 public

馃彧 The real market 鈥?later

When the marketplace truly opens, it will follow the published build plan 鈥?not this showcase wall.