reasoning-checks
Skill109 starsChanged 39 days agoMITView on GitHub
Shared reasoning-discipline contract - four pre-conclusion checks (epistemic inversion / pre-mortem, attractor-state check against the live core fingerprint, provenance weighting, conditional reference-class anchor) with required output blocks. Referenced by /advise, /canon-advise, and both crystallization paths before a conclusion is finalized; also directly invocable on any draft conclusion or recommendation.
In this repository61 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- adviseSkill
- ai-crystallizeSkill
- analyze-kbSkill
- auto-discoverySkill
- benchmark-memorySkill
- brain-mergeSkill
- changelog-protocolSkill
- coherence-sweepSkill
- compute-lifecycleSkill
- create-articleSkill
- create-explanatory-imageSkill
- decideSkill
- deep-researchSkill
- detect-tensionsSkill
- dialecticSkill
- domain-watchSkill
- elicitation-techniquesSkill
- epistemic-classificationSkill
- epub-chapter-extractorSkill
- extract-document-insightsSkill
- extract-insightsSkill
- find-connectionsSkill
- first-runSkill
- get-perspective-onSkill
- get-youtube-transcriptSkill
- git-commit-pushSkill
- graduate-insightsSkill
- incubation-loopSkill
- ingest-sourceSkill
- insight-capture-formatSkill
- insight-interviewSkill
- integrate-recent-notesSkill
- learn-new-thingsSkill
- manage-reference-dataSkill
- manage-thinking-topicsSkill
- manage-watching-domainsSkill
- multi-format-book-extractorSkill
- nano-banana-image-generatorSkill
- propagate-changeSkill
- quick-searchSkill
- recallSkill
- ref-auditSkill
- ref-bridgeSkill
- ref-ingestSkill
- ref-querySkill
- ref-reconcileSkill
- ref-refreshSkill
- refresh-indexSkill
- ref-supersedeSkill
- scheduled-runSkill
- scope-mountSkill
- search-vaultSkill
- self-diagnosticSkill
- start-hereSkill
- synthesize-insightsSkill
- talkSkill
- test-memory-systemSkill
- think-about-itSkill
- update-changelogSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 4 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 42
When: always, before any recommendation or conclusion is finalized.
- Distinctive: line 57
When: always on advice surfaces.
- Distinctive: line 66
If 2 or more are top-10 hubs or bridges → declare an attractor state and generate ONE alternative framing that uses zero of the top-10.
- Distinctive: line 78
When: always when the conclusion rests on KB notes.
- Very specific: line 101
Direction A hook: when the operational-memory belief store ships (TARGET-ARCHITECTURE.md → Direction A), the base rate must *adjust the …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Reasoning Checks
ℹ️ First, set expectations: print one line with this skill's version and its most recent change - the top of
metadata.changelog- e.g.reasoning-checks v1.0 — recent: initial four-check contract. Then proceed. (Skip the banner when applied silently inside a consuming skill.)
Purpose
The calibration counterpart to the Thinking Geometries. The 2026-07-31 self-audit found the reasoning repertoire strong on discovery and synthesis but thin on calibration - everything runs from the inside view, and a graph this connected (50k+ edges, spreading activation) is structurally prone to falling in love with its own semantic connections. These checks close the gap at the two moments that matter: before advice is given and before a conclusion crosses the endorsement boundary.
Contract rule (the SOURCE-AUTHORITY.md pattern): this file is the single source of truth for the check definitions. Consuming skills reference it and apply the checks; they never copy the definitions inline. Change a check here, not in consumers.
State Dependencies
| Source | Location | Read | Write |
| Core fingerprint (live) | resources/local-brain-search/run_connections.sh --hubs --json and --bridges --json | ✓ | |
| Core fingerprint (fallback) | knowledge-base-analysis.md → "Core hubs" / "Core bridges" tables (note the analysis date) | ✓ | |
| Cited notes' frontmatter | Brain/**/*.md (provenance: field) | ✓ |
Never writes. Output blocks are embedded in the consuming skill's own artifact or answer. All reads are lookups, not retrievals - nothing here trains q-values.
The Four Checks
Check 1 - Epistemic Inversion (pre-mortem)
When: always, before any recommendation or conclusion is finalized.
Assume the conclusion is wrong and work backwards. Required output block:
### Epistemic Inversion
- **Failure assumption:** it is 12 months out and this [recommendation | conclusion] proved wrong.
- **Causal autopsy:** most likely reason it failed: [specific mechanism]
- **Specific falsifier:** we will know it failed if [observable event / metric crossing a threshold / dated milestone]The anti-theater rule (load-bearing): the falsifier must name a specific observable - an event that either happens or doesn't, a metric with a threshold, a dated milestone. Generic hedges ("market conditions shift", "priorities change", "new information emerges") FAIL the check. On human-gated surfaces, a generic falsifier means redo the inversion before presenting. On autonomous surfaces, stamp falsifier: weak and flag it in the run log - never block the run (No-Gates Rule).
Check 2 - Attractor-State Check
When: always on advice surfaces; on crystallization. The system's dominant frameworks are measured - the core fingerprint names its own gravity wells - so this check is semi-mechanical rather than introspective.
Procedure:
- Fetch the current top-10 core hubs + top-10 bridges (live call below; fall back to the
knowledge-base-analysis.mdtables, noting their date). In consuming skills that already run a parallel search batch, fetch hubs in that same batch - no extra round.resources/local-brain-search/run_connections.sh --hubs --json 2>/dev/null resources/local-brain-search/run_connections.sh --bridges --json 2>/dev/null - List the frameworks/notes that are load-bearing in the draft (the ones the conclusion would collapse without).
- If 2 or more are top-10 hubs or bridges → declare an attractor state and generate ONE alternative framing that uses zero of the top-10. Compare honestly; keep the survivor or present both.
### Attractor Check
- load_bearing_frameworks: [[A]], [[B]], ...
- top10_overlap: N → [clear | ATTRACTOR]
- alternative_framing: [only when ATTRACTOR: 1 paragraph using no top-10 hub] → [kept | original stands because ...]Check 3 - Provenance Weighting
When: always when the conclusion rests on KB notes.
Tally the provenance: of the load-bearing cited notes (visible in frontmatter of notes already read; grep -m1 '^provenance:' for any others). A conclusion resting mainly on ai-inferred/encountered notes is unendorsed synthesis - it must say so and must never be phrased as "your view" / "your framework". reference records are facts, not endorsements. Notes with no provenance: field count as unmarked (legacy) - report them honestly, don't assume.
### Provenance Base
originated/endorsed: N · encountered: M · ai-inferred: K · unmarked: J · reference: R
→ [rests on endorsed thinking | rests substantially on unendorsed synthesis - weighted accordingly]Check 4 - Reference-Class Anchor
When: CONDITIONAL - only when the conclusion asserts a forecast, probability, magnitude, timeline, or success likelihood. Never force it onto purely conceptual syntheses (that is exactly how checks become checkbox theater).
Name the reference class, state the base rate (or state honestly that none is known - itself a calibration signal), and justify any deviation. The block is deliberately parseable:
### Reference Class
- class: [what population of cases this belongs to]
- base_rate: [X% | unknown]
- case_vs_base: [above | below | at] - [why the deviation is justified, or "no deviation claimed"]Direction A hook: when the operational-memory belief store ships (TARGET-ARCHITECTURE.md → Direction A), the base rate must adjust the stated confidence before any beliefs.db write - not ride along as metadata.
Applicability Matrix
| Surface | Inversion | Attractor | Provenance | Ref-class |
/advise | required | required | required | conditional |
/canon-advise ([MY READ] layer ONLY) | required | required | required | conditional, bet surfaces only |
/decide | required | required (situation framing only - decision-science hubs are its tools, not attractor pull) | required | required in practice (decisions embed forecasts) |
manage-thinking-topics crystallize | required - embedded in the artifact | required | required | conditional |
ai-crystallize (autonomous) | required - embedded; weak falsifier → flagged, never blocks | skip (stays search-free by design) | skip | skip |
The Canon Guard (do not violate)
These checks apply to the agent's own reasoning layers only. Never invert, score, base-rate, or generate "alternative framings" for a canon decision or spec - a decision has a status, not a probability. On /canon-advise, the checks run on the [MY READ]/dissent layer exclusively; bet is the only scoreable canon surface (per the Authority Contract in that skill).
Autonomous-Surface Rule
On any autonomous consumer (currently ai-crystallize): check presence is required, check quality is flagged (falsifier: weak in the run log), and nothing ever blocks or gates the scheduled run.
Direct Invocation
/reasoning-checks [material] - run the applicable checks on a pasted conclusion, a note path, or (with no argument) the current conversation's most recent recommendation/conclusion. Output the blocks; make no edits.
Deferred Wirings (documented so nothing wires early)
- Incubation moves #7-8 (Reference Class Forecast; Causal Graph Sketch) - add to the rotating move set when the incubation loop is re-enabled.
/think-about-it- deliberately unwired; divergent exploration stays unconstrained.
(Decision-theoretic framing shipped 2026-07-31 as the standalone /decide skill - see the applicability matrix.)
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

