Przejdź do treści
Biblioteka skilli
AI i agenci · Przegląd i audyt

Trust Decay

/Trust Decay

Zapisuje pewne twierdzenia agenta w pliku TRUST_DECAY.md i rozstrzyga je jako trafne lub błędne, także gdy uznany za naprawiony błąd wraca.

Repozytorium
aigent-os ↗
Licencja
MIT
Ostatnia zmiana repozytorium
12.11.2026

Ocena i bezpieczeństwo

90/ 100Klasa: dobry
Format15 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
1
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 12.11.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co robi i kiedy

Zapisuje pewne twierdzenia agenta w pliku TRUST_DECAY.md i rozstrzyga je jako trafne lub błędne, także gdy uznany za naprawiony błąd wraca.

Oryginalny opis (EN)

Log confident agent claims and resolve them as right or wrong in TRUST_DECAY.md. Use when capturing confidence, tracking wrong claims, or a supposedly fixed bug returns.

Jak działa

  1. When to use
  2. How to execute
  3. Why two phases
  4. Anti-patterns this prevents
  5. What this skill is NOT
  6. Connects to
Sekcje z nagłówków SKILL.md, w oryginale · 103 linii, 694 słów
Podgląd SKILL.md
---
name: Trust Decay
description: "Log confident agent claims and resolve them as right or wrong in TRUST_DECAY.md. Use when capturing confidence, tracking wrong claims, or a supposedly fixed bug returns."
---

# Trust Decay

The two-phase ledger that tracks confident agent claims and their eventual outcomes. See [[Cost of Confidence]] for the doctrine.

## When to use

### Phase 1 — Capture a confident claim

- An agent declared something "fixed", "tested", "verified", "complete", "deployed"
- The claim is unhedged or treated as ground truth
- The claim is consequential enough that being wrong about it would matter
- Use on prompts like: "/trust-decay capture", "log this claim", "capture confidence", "agent claimed", "I claimed", "the agent said it was fixed", "another agent verified"

### Phase 2 — Resolve a claim

- A previously-captured claim has been confirmed (right) or revealed wrong
- The principal noticed a discrepancy between an earlier claim and reality
- Use on prompts like: "/trust-decay resolve", "the bug came back", "that wasn't actually fixed", "trust event", "wrong claim", "confidence was wrong"

## How to execute

### Phase 1 — Capture

```markdown
## YYYY-MM-DD — {short claim summary}
**Claim:** "{the actual confident statement, verbatim}"
**Source:** {agent name or self}
**Context:** {what task, what session, what was being worked on}
**Verification stated:** {what the claimant said they verified, if anything}
**Status:** OPEN — awaiting outcome
```

Append to `vault/memory/TRUST_DECAY.md` under the "Open" section.

The principal can capture manually, OR `/honesty-check` (when wired) can auto-capture any unhedged confident claim it produces.

### Phase 2 — Resolve

When a previously-captured claim is proved or disproved:

1. Find the open entry in `vault/memory/TRUST_DECAY.md` matching the claim
2. Append resolution data:

```markdown
**Resolution (YYYY-MM-DD):** WRONG | PARTIALLY WRONG | CONFIRMED CORRECT
**Evidence:** {what surfaced the discrepancy — file/line/log/observation}
**Time-to-discovery:** {duration from claim to discovery}
**Trust-decay event:** YES | NO
**Pattern:** {one line — what category of mistake/correctness was this?}
```

3. Move the entry from the "Open" section to the "Resolved" section.

### Step 3 — Optional aggregate

If the principal asks for the trust trajectory, walk the ledger and compute:

```
Trust trajectory ({start date} → {end date}):
- Total claims captured: N
- Confirmed correct: M (X%)
- Wrong: P (Y%)
- Partially wrong: Q (Z%)

By category:
- Doc work: X/Y correct
- Infrastructure: ...
- Production fixes: ...

Recurring patterns:
- {pattern name}: occurred N times — promote to /diagnose checks?
```

This output becomes the basis for `/retro` analysis and feeds into the [[Common Failure Modes]] corpus.

## Why two phases

The pairing is the whole point. A confident claim with no resolution is just rhetoric. A wrong outcome with no original claim is just a bug. The pair — *confident claim, then disconfirmation* — is what measures calibration.

## Anti-patterns this prevents

- Forgetting that the agent claimed something was fixed three weeks ago when the same bug recurs (no audit trail)
- Treating every fresh occurrence of an old bug as new (no pattern detection)
- Letting agent confidence levels drift without a corrective signal (no calibration data)
- Blaming the agent in the moment without realizing it's the Nth time the same pattern has played out (no longitudinal context)

## What this skill is NOT

- **Not a punishment system.** The ledger calibrates the agent's claims against outcomes. It's measurement, not judgment.
- **Not adversarial.** Capturing a confident claim isn't accusing the agent of being wrong; it's noting what the agent said so the eventual outcome can pair with it.
- **Not the only place to track confidence.** [[Honesty Check]] is the agent-side discipline; [[Reading Agent Output Defensively]] is the principal-side discipline; this skill is the longitudinal data layer that makes both useful over time.

## Connects to

- [[Cost of Confidence]] — the doctrine note this skill instantiates
- [[Honesty Check]] — should reduce Phase 1 capture frequency over time (fewer unhedged claims = fewer entries to resolve later)
- [[Reading Agent Output Defensively]] — the principal-side pattern that often triggers a Phase 1 capture
- [[Common Failure Modes]] — sister corpus tracking layer-level failures rather than confidence-level failures
- `vault/memory/TRUST_DECAY.md` — the ledger file
- [Closing the Measurement Loop](../../docs/closing-the-measurement-loop.md) — the `Stop` hook that prompts a Phase 1 capture when a turn's confident claims went unrecorded
- `/retro` (forthcoming v0.3) — quarterly analysis that consumes this ledger

Wersje

W bazie nie ma innych skilli o nazwie Trust Decay.

Migawka bazy z 13.11.2026 · wersja paczki 5304d8bc4ac7 (skrót SHA-256 całej paczki)

Pobrania w ostatnich 30 dniach: 0 · ranking

Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.

Podobne skille

  • 90detect-hallucinations
    Wykrywa, kiedy LLM zmyśla, i oznacza lub blokuje pewne, lecz błędne odpowiedzi, zanim dotrą do użytkowników.
  • 100trust-memory
    Weryfikuje wiarygodność agentów AI, wnosi zweryfikowane twierdzenia i przeszukuje wiedzę zbiorową na platformie TrustMemory, w tym wyniki zaufania i reputację.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.