piensalo 0.1.0a2__tar.gz

piensalo-0.1.0a2/.github/CODEOWNERS

@@ -0,0 +1,2 @@
+# PIÉNSALO code ownership — reviews route to the creator/maintainer.
+* @ralfyishere

piensalo-0.1.0a2/.github/ISSUE_TEMPLATE/bug-report.yml

@@ -0,0 +1,56 @@
+name: Bug report
+description: Something in piensalo behaves incorrectly
+labels: ["bug", "needs-triage"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for the report. Two notes before you start:
+        - **Security issues:** do NOT file here — use private reporting (see SECURITY.md).
+        - **"It made my results worse":** that's a negative-transfer report — use that template instead; it's a first-class contribution.
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: What happened
+      description: The observable behavior, not the interpretation.
+      placeholder: "`piensalo verify draft` exited 0 but printed no verdict line."
+    validations:
+      required: true
+  - type: textarea
+    id: expected
+    attributes:
+      label: What you expected
+    validations:
+      required: true
+  - type: textarea
+    id: repro
+    attributes:
+      label: Minimal reproduction
+      description: Exact commands from a clean state. If you can't reproduce it, say so — flaky reports are still useful when labeled as flaky.
+      render: shell
+    validations:
+      required: true
+  - type: textarea
+    id: environment
+    attributes:
+      label: Environment
+      description: Paste the output of `piensalo doctor` (it redacts secrets; still skim it before pasting).
+      render: text
+    validations:
+      required: true
+  - type: input
+    id: model
+    attributes:
+      label: Model + adapter (as recorded by provenance)
+      placeholder: "e.g. ollama/llama3.1:8b via ollama adapter, or 'skills-only, no runtime'"
+    validations:
+      required: true
+  - type: checkboxes
+    id: checks
+    attributes:
+      label: Confirm
+      options:
+        - label: This is not a security vulnerability (those go through private reporting, per SECURITY.md)
+          required: true
+        - label: I removed secrets and private paths from everything pasted above
+          required: true

piensalo-0.1.0a2/.github/ISSUE_TEMPLATE/case-study.yml

@@ -0,0 +1,52 @@
+name: Case study / benchmark reproduction
+description: Share a real usage trace or a reproduction of our published numbers — agreeing or disagreeing.
+labels: ["case-study", "evidence"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Reproductions that DISAGREE with our published numbers are especially welcome —
+        they get the same visibility as confirmations (docs/methodology.md).
+  - type: dropdown
+    id: kind
+    attributes:
+      label: Kind
+      options:
+        - Case study (real task, full lifecycle trace)
+        - Benchmark reproduction (matches published results)
+        - Benchmark reproduction (DISAGREES with published results)
+    validations:
+      required: true
+  - type: textarea
+    id: setup
+    attributes:
+      label: Setup
+      description: OS, install method, piensalo version, exact model identifiers as recorded by provenance, hardware if local.
+    validations:
+      required: true
+  - type: textarea
+    id: narrative
+    attributes:
+      label: What happened
+      description: The trace or run summary. For case studies include verdicts (REPAIR REQUIRED / NO REPAIR NEEDED / PASS) as they occurred — including wrong ones.
+    validations:
+      required: true
+  - type: textarea
+    id: artifacts
+    attributes:
+      label: Evidence artifacts
+      description: Outputs, evidence bundles, result tables. Scrub secrets and private paths before attaching.
+    validations:
+      required: true
+  - type: textarea
+    id: divergence
+    attributes:
+      label: Divergence from published results (if any)
+      description: For reproductions — what differed, by how much, and your best hypothesis why.
+  - type: checkboxes
+    id: consent
+    attributes:
+      label: Publication
+      options:
+        - label: I consent to this content being published (anonymized on request) in the project's case studies or evidence reports
+          required: true

piensalo-0.1.0a2/.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,5 @@
+blank_issues_enabled: true
+contact_links:
+  - name: Security vulnerability (private)
+    url: https://github.com/ralfyishere/piensalo/security/advisories/new
+    about: Please report security issues privately via GitHub security advisories — never in a public issue. See SECURITY.md.

piensalo-0.1.0a2/.github/ISSUE_TEMPLATE/negative-transfer.yml

@@ -0,0 +1,66 @@
+name: Negative-transfer report
+description: Piénsalo made your results WORSE. This is a first-class contribution — confirmed reports are published with credit.
+labels: ["negative-transfer", "evidence"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thank you — this is the report we most need (Manifesto, principle 6).
+        Confirmed reports go into NEGATIVE-RESULTS.md with credit (if you want it).
+  - type: textarea
+    id: task
+    attributes:
+      label: The task
+      description: What you were doing, in enough detail to reproduce the setup. No secrets or client-identifying details.
+    validations:
+      required: true
+  - type: textarea
+    id: comparison
+    attributes:
+      label: The comparison
+      description: What "worse" is measured against — the same task without Piénsalo? A different configuration? Describe both sides.
+    validations:
+      required: true
+  - type: textarea
+    id: worse-how
+    attributes:
+      label: How it was worse
+      description: Observable difference — wrong output, damaged correct work, unnecessary repair applied, failure to abstain, time/cost burned.
+    validations:
+      required: true
+  - type: input
+    id: model
+    attributes:
+      label: Model + adapter (as recorded by provenance)
+    validations:
+      required: true
+  - type: input
+    id: version
+    attributes:
+      label: piensalo version / skill versions
+    validations:
+      required: true
+  - type: dropdown
+    id: repeatability
+    attributes:
+      label: Repeatability
+      options:
+        - Reproduced 2+ times
+        - Happened once, could not retry yet
+        - Happened once, retry did NOT reproduce it
+    validations:
+      required: true
+  - type: textarea
+    id: evidence
+    attributes:
+      label: Evidence artifacts
+      description: Relevant .piensalo/evidence files or transcripts (scrub secrets first).
+  - type: dropdown
+    id: credit
+    attributes:
+      label: Credit preference if published
+      options:
+        - Credit me (GitHub handle)
+        - Publish anonymously
+    validations:
+      required: true

piensalo-0.1.0a2/.github/ISSUE_TEMPLATE/new-skill.yml

@@ -0,0 +1,103 @@
+name: New skill contribution
+description: Propose a new portable skill. All fields are required — they ARE the skill's spec.
+labels: ["skill-proposal", "needs-review"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Every field below is mandatory — a skill missing any of them will not be merged
+        (CONTRIBUTING.md, type 1). New skills honestly start at evidence level `DESIGNED`.
+  - type: input
+    id: skill-name
+    attributes:
+      label: Skill name
+      placeholder: "kebab-case, e.g. unit-consistency-check"
+    validations:
+      required: true
+  - type: textarea
+    id: observable-trigger
+    attributes:
+      label: Observable trigger
+      description: What a session can SEE that activates this skill. "When appropriate" is not observable.
+    validations:
+      required: true
+  - type: textarea
+    id: counterindications
+    attributes:
+      label: Counterindications
+      description: Situations where this skill must NOT fire, even though it might seem to apply.
+    validations:
+      required: true
+  - type: textarea
+    id: bounded-procedure
+    attributes:
+      label: Bounded procedure
+      description: The steps, with an explicit budget (max iterations / max scope). Unbounded procedures are declined.
+    validations:
+      required: true
+  - type: textarea
+    id: expected-output
+    attributes:
+      label: Expected output
+      description: What the skill's result looks like, concretely.
+    validations:
+      required: true
+  - type: textarea
+    id: verification
+    attributes:
+      label: Verification
+      description: How a reader checks the skill did its job (not "review carefully" — a checkable step).
+    validations:
+      required: true
+  - type: textarea
+    id: stop-condition
+    attributes:
+      label: Stop condition
+      description: When the skill is done, and when it must abort. Include the abstention path if one exists.
+    validations:
+      required: true
+  - type: textarea
+    id: positive-example
+    attributes:
+      label: Positive example
+      description: A realistic trace where the skill helps — input, skill behavior, improved outcome.
+    validations:
+      required: true
+  - type: textarea
+    id: near-miss
+    attributes:
+      label: Near miss
+      description: A concrete case that LOOKS like the trigger but isn't, and why the skill must not fire on it.
+    validations:
+      required: true
+  - type: textarea
+    id: negative-transfer-risk
+    attributes:
+      label: Negative-transfer risk
+      description: Where could this skill make results WORSE? (Strong models, easy tasks, and already-correct work are the usual victims — see NEGATIVE-RESULTS.md NR-1, NR-8.)
+    validations:
+      required: true
+  - type: textarea
+    id: evaluation-plan
+    attributes:
+      label: Evaluation plan
+      description: How this skill will be measured — tasks, comparison condition, grader, success criteria (docs/methodology.md).
+    validations:
+      required: true
+  - type: dropdown
+    id: evidence-level
+    attributes:
+      label: Claimed evidence level
+      description: Per docs/evidence-levels.md. DESIGNED is the honest default for new skills.
+      options:
+        - DESIGNED
+        - SMOKE_TESTED
+        - EXPERIMENTALLY_TESTED
+        - REPLICATED
+    validations:
+      required: true
+  - type: textarea
+    id: receipts
+    attributes:
+      label: Receipts for the claimed level
+      description: Required for any level above DESIGNED — link the runs/artifacts.

piensalo-0.1.0a2/.github/ISSUE_TEMPLATE/security.md

@@ -0,0 +1,25 @@
+---
+name: Security issue (do not use this template)
+about: Security reports must go through private reporting — this template exists only to redirect you.
+labels: ["invalid"]
+---
+
+<!--
+STOP — please do not file security issues publicly.
+
+A public issue gives attackers the details before users have a fix.
+-->
+
+## Please report privately instead
+
+1. Go to the repository's **Security** tab → **Report a vulnerability**
+   (https://github.com/ralfyishere/piensalo/security/advisories/new)
+2. You'll get an acknowledgment within 72 hours (see SECURITY.md for the
+   full policy, scope, and disclosure process).
+
+If you file vulnerability details here anyway, maintainers will redact the
+issue and ask you to resubmit privately — you'll still get credit for the
+find, but please save us both the step.
+
+**In scope:** the CLI/runtime, skill loader and `skill scan`, model adapters,
+release artifacts. **Threat model:** THREAT-MODEL.md.

piensalo-0.1.0a2/.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,41 @@
+# Pull request
+
+## What this changes
+
+<!-- One concern per PR. If you're tempted to write "and", consider two PRs. -->
+
+## Type
+
+- [ ] Skill (all 11 required fields present — see the new-skill issue template)
+- [ ] Verifier (adversarial test case included; FP/FN behavior stated)
+- [ ] Model adapter (provenance honored, no silent fallback, no creds in logs)
+- [ ] Task / benchmark reproduction
+- [ ] Bug fix
+- [ ] Docs
+- [ ] Other:
+
+## Verification I actually ran
+
+<!-- Paste the commands and their real output. "Should work" is a prediction,
+     not a result. For docs PRs: paste the link-check / claim-check you did. -->
+
+```
+```
+
+## Evidence discipline
+
+- [ ] No capability claim in this PR exceeds its evidence level (docs/evidence-levels.md)
+- [ ] I did not upgrade any evidence level via prose (only experiments do that)
+- [ ] N/A — this PR makes no capability claims
+
+## Security invariants (SECURITY.md)
+
+- [ ] No telemetry, no network outside explicit adapters, no silent fallback,
+      no code execution from skill text, no secrets in code/logs/fixtures
+- [ ] If this PR touches a guard (scanner rule, path check, scrubber): the
+      adversarial test that tries to defeat it is included
+
+## Breaking changes
+
+- [ ] None
+- [ ] Yes — described above, CHANGELOG.md updated

piensalo-0.1.0a2/.github/workflows/ci.yml

@@ -0,0 +1,121 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+permissions:
+  contents: read
+
+jobs:
+  test:
+    name: pytest (${{ matrix.os }}, py${{ matrix.python-version }})
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest]
+        python-version: ["3.10", "3.11", "3.12", "3.13"]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+      - name: Install
+        run: |
+          python -m pip install --upgrade pip
+          pip install -e .
+          pip install pytest
+      - name: Run tests
+        run: pytest -q
+
+  skill-lint:
+    name: Skill lint
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - name: Install
+        run: |
+          python -m pip install --upgrade pip
+          pip install -e .
+      - name: Lint shipped skills (structure + scan)
+        run: |
+          python -m piensalo skill scan skills/ micro-skills/
+          python -m piensalo skill lint skills/ micro-skills/
+
+  secret-scan:
+    name: Secret scan
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Grep for secret patterns
+        # A net, not a license — see CONTRIBUTING.md security standards.
+        run: |
+          set -e
+          PATTERNS='(sk-[A-Za-z0-9]{20,}|AKIA[0-9A-Z]{16}|ghp_[A-Za-z0-9]{36}|gho_[A-Za-z0-9]{36}|xox[baprs]-[A-Za-z0-9-]{10,}|-----BEGIN (RSA|EC|OPENSSH|DSA|PGP) PRIVATE KEY-----|AIza[0-9A-Za-z_-]{35})'
+          if grep -rInE "$PATTERNS" --exclude-dir=.git --exclude=ci.yml .; then
+            echo "::error::Secret-shaped string found. Remove it and rotate the credential."
+            exit 1
+          fi
+          echo "Secret scan clean."
+
+  link-check:
+    name: Relative links + missing files
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Check relative links in markdown
+        # Verifies every relative markdown link resolves to a real file.
+        # External-URL checking is deliberately out of scope (network
+        # flakiness is not broken docs); tracked in ROADMAP v0.2.
+        run: |
+          python3 - <<'EOF'
+          import os, re, sys, urllib.parse
+          bad = []
+          for root, dirs, files in os.walk('.'):
+              dirs[:] = [d for d in dirs if d not in ('.git', '.venv', 'node_modules')]
+              for f in files:
+                  if not f.endswith('.md'):
+                      continue
+                  p = os.path.join(root, f)
+                  text = open(p, encoding='utf-8', errors='replace').read()
+                  for m in re.finditer(r'\]\(([^)#\s]+)(#[^)]*)?\)', text):
+                      target = urllib.parse.unquote(m.group(1))
+                      if re.match(r'^[a-z]+:', target):  # http(s), mailto, etc. — stub skips
+                          continue
+                      resolved = os.path.normpath(os.path.join(os.path.dirname(p), target))
+                      if not os.path.exists(resolved):
+                          bad.append(f"{p}: {target}")
+          if bad:
+              print("Broken relative links:")
+              print("\n".join(bad))
+              sys.exit(1)
+          print("Relative links OK.")
+          EOF
+
+  build:
+    name: Package build
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - name: Build sdist + wheel
+        run: |
+          python -m pip install --upgrade pip build twine
+          python -m build
+          twine check dist/*
+      - name: Wheel install smoke
+        run: |
+          python -m venv /tmp/smoke && /tmp/smoke/bin/pip -q install dist/*.whl
+          /tmp/smoke/bin/piensalo doctor
+          /tmp/smoke/bin/piensalo think examples/math/task.md > /dev/null && echo "think OK"
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist
+          path: dist/

piensalo-0.1.0a2/.github/workflows/publish.yml

@@ -0,0 +1,26 @@
+name: Publish to PyPI
+
+on:
+  release:
+    types: [published]
+
+permissions:
+  contents: read
+
+jobs:
+  publish:
+    runs-on: ubuntu-latest
+    environment: pypi
+    permissions:
+      id-token: write   # OIDC for PyPI Trusted Publishing — no tokens stored
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - name: Build distributions
+        run: |
+          python -m pip install --upgrade build
+          python -m build
+      - name: Publish to PyPI (trusted publishing)
+        uses: pypa/gh-action-pypi-publish@release/v1

piensalo-0.1.0a2/.gitignore

@@ -0,0 +1,7 @@
+__pycache__/
+*.py[cod]
+.venv/
+dist/
+*.egg-info/
+.piensalo/
+uv.lock.tmp

piensalo-0.1.0a2/BENCHMARKS.md

@@ -0,0 +1,69 @@
+# Benchmarks
+
+Two pre-registered controlled runs stand behind the current evidence map
+([EVIDENCE.md](EVIDENCE.md)). Both are internal-lab runs; the public
+reproduction harness lives in [evals/](evals/) and reproducing (or
+refuting) these numbers on your own models is a first-class contribution
+([evals/results/](evals/results/)).
+
+## Design shared by both runs
+
+| Property | Value |
+|---|---|
+| Cells | 8 sealed tasks × 3 reps × 5 conditions = 120 |
+| Execution model | `claude-haiku-4-5` (requested + resolved recorded per cell) |
+| Conditions | bare · monolithic skill · adaptive repair · oracle repair · output-contract guardian |
+| Graders | deterministic, layered (cognition / delivery / rendering / routing / verification), frozen before any cell |
+| Gates | pre-registered, frozen before any cell, never re-scoped |
+| Stub policy | infra failures = NOT RUN, never FAIL |
+| Budget | equal across conditions (bounded turns, bounded prompt prefix) |
+
+n = 8 tasks per run. Everything below is directional, not powered. We
+publish it anyway, with the confounds attached.
+
+## Run 1 — weak-baseline task set
+
+The bare model failed most procedural tasks (procedural mean 37.5/100).
+
+| Finding | Number | Confound |
+|---|---|---|
+| Oracle targeted repair, procedural lift | **+18.8pp** | task-concentrated: +66.7 on the best task, **−20 on the worst** (a repair pass destroyed a delivered answer); one task grader-blind |
+| Output-contract guardian, delivery-failure reduction | **30% relative** (0.417 → 0.292) | part of the apparent cognition movement was measurement visibility (absent output is ungradeable) |
+| Guardian paired cognition harm | **zero** | — |
+| Adaptive router fires | **0 / 24** | two deterministic implementation defects (threshold arithmetic; contract-schema mismatch) — found by forensics, fixed, and re-tested in run 2 |
+| Grader artifacts discovered | 3 classes | fixed in the run-2 grading library; both directions (false credit AND false blame) |
+
+## Run 2 — saturated task set (the harm measurement)
+
+The fresh set turned out too easy: bare cognition rate **1.0** (verified
+against answer keys). No lift was measurable — so harm became cleanly
+measurable instead.
+
+| Finding | Number | Note |
+|---|---|---|
+| Forced repair on correct drafts (oracle) | **−20pp procedural, 7/24 paired regressions ≥ 40pp** | every regression was DELIVERY damage (prose wrapped around contract-exact output) |
+| Cognition flips from forced repair | **0 / 48 cells** | re-prompting is a delivery risk, not a reasoning risk |
+| Adaptive router (after both fixes) | fires 8/24, precision **0.375** on correct drafts | third defect class: correctness-blind detection; abstention on easy classes was perfect (12/12) |
+| Monolithic skill on knowledge-floor tasks | **60 vs 100**, 2 induced wrong answers | artifact-free graders this time; +1109 median tokens |
+| Guardian on contract-clean drafts | 0 fires, **0 harm** (24/24 correct abstention) | its improvement gate was unexercised — no delivery failures to fix |
+| Abstaining paths, total harm | **zero** | the core product thesis, measured |
+
+## What these runs do NOT show
+
+- No cross-model generality: one small-model family executed the cells.
+- No statistical certainty: n = 8 tasks per run.
+- No end-to-end product benchmark yet: these are mechanism isolation runs.
+- Run 2's saturation means the repair-lift hypothesis still rests on run 1
+  alone. The next kill test is a difficulty-calibrated fresh set (pilot
+  gate: bare procedural cognition ≤ 0.7 before sealing).
+
+## Reproduce
+
+```
+make benchmark          # grader self-test battery + 2-task smoke, no API needed
+uv run piensalo-eval # full harness (bring your own adapter/model)
+```
+
+Task sets in `evals/public-tasks/` are labeled `DEMO` or `DEVELOPMENT`.
+Internal unconsumed held-out reserves are never published; consumed sets
+are ported here after their runs close.

piensalo-0.1.0a2/CHANGELOG.md

@@ -0,0 +1,58 @@
+# Changelog
+
+All notable changes to Piénsalo are documented here.
+Format: [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
+Versioning: [SemVer](https://semver.org/) once we reach 0.1.0; until then,
+everything may change.
+
+## [0.1.0-alpha.2] — 2026-07-13
+
+Truthfulness and UX correction release.
+
+### Added
+- Real flagship workflow: `examples/flagship/` with a committed transcript of
+  actual CLI output; CI parity tests prevent README/demo drift
+- `repair` adapter mode: writes a NEW repaired file + provenance sidecar and
+  automatically re-inspects the result; source drafts are never overwritten
+- `think` accepts a task file or inline text (`--file`/`--text` overrides)
+- `python -m piensalo`; `skill lint`/`scan` accept multiple and parent paths
+- PyPI publication via GitHub trusted publishing (release-triggered)
+
+### Changed
+- `repair` default mode clearly labeled as a repair PACKET (nothing applied)
+- `verify` reports in five strict buckets; UNMEASURED never counts as a pass
+- Roadmap reconciled to the truthful maturity hierarchy; stale
+  pre-publication language removed; public tree simplified (operator/process
+  scaffolding removed, provenance kept under docs/provenance/)
+
+### Fixed
+- Red CI Skill lint job (root cause: missing `__main__.py` + single-path lint)
+- Tests: 49 → 86
+
+## [Unreleased] — 0.1.0-dev
+
+### Changed
+- 0.1.0-dev — public identity migrated to PIÉNSALO (previously internal
+  codename Fable Think); no functional changes.
+
+### Added
+- Cognitive core: 11 operations (`recover_objective` → `deliver`) with a
+  bounded execute → inspect → repair → verify lifecycle.
+- Six CLI modes: `think`, `inspect`, `repair`, `verify`, `loop`, `skill`.
+- Portable Agent Skills layer (works in any agent that reads skills; no
+  runtime required).
+- Optional runtime: persistent state, bounded loops, model provenance
+  recording, measurement hooks.
+- Abstention as a first-class result: `NO REPAIR NEEDED`.
+- Model adapters: manual copy/paste, Claude / Claude Code, OpenAI-compatible,
+  Ollama/local, generic command adapter.
+- `piensalo skill scan` (third-party skill vetting) and
+  `piensalo doctor` (environment + configuration audit).
+- Evidence discipline: per-capability evidence levels with promotion gates
+  (docs/evidence-levels.md), public negative results (NEGATIVE-RESULTS.md).
+- Security defaults: zero telemetry, offline core, no silent model fallback,
+  no destructive actions without approval (SECURITY.md, THREAT-MODEL.md).
+
+### Notes
+- Pre-release. Evidence base is young; see README "Limitations" and
+  EVIDENCE.md before relying on any capability claim.

piensalo-0.1.0a2/CITATION.cff

@@ -0,0 +1,27 @@
+cff-version: 1.2.0
+message: "If you use Piénsalo in your research, please cite it as below."
+title: "PIÉNSALO"
+type: software
+authors:
+  - family-names: "Peña"
+    given-names: "Rafael"
+    alias: "ralfyishere"
+license: MIT
+repository-code: "https://github.com/ralfyishere/piensalo"
+url: "https://github.com/ralfyishere/piensalo"
+version: 0.1.0-dev
+date-released: "2026-07-12"
+keywords:
+  - cognitive architecture
+  - AI agents
+  - verification
+  - self-repair
+  - agent skills
+  - abstention
+abstract: >-
+  Piénsalo is an open cognitive operating system for AI models and agents.
+  It separates thinking, inspection, repair, verification, evidence, and
+  continuation into distinct operations, supports abstention as a first-class
+  result, and attaches an explicit evidence level to every capability claim.
+  Piénsalo is an independent open-source project. It is not affiliated with
+  or endorsed by Anthropic.