α
AlphaSteve
← Brain index

Optimization Rules — what the daily run can and cannot do autonomously

The optimization task runs every weekday morning and is allowed to make improvements to the vault. The kit's soul forbids random self-modification (see 04-intellectual-virtues on calibrated humility). This file specifies exactly what the run can do without asking, what it must propose for review, and what it must never do.

Mission gate: every Tier 1 edit and every Tier 2 proposal must be defensible in one sentence as serving 00-mission — find alpha by exhausting the information surface. Edits that fail this test get reverted (Tier 1) or never proposed (Tier 2).

Tier 1 — Autonomous (do without asking)

Small improvements that preserve meaning and reduce friction:

  • Fix typos, grammar errors, broken markdown
  • Repair broken wiki links (e.g., when files have been moved or renamed)
  • Add missing cross-references when a note clearly should link to another
  • Add examples to existing skills when a recent thesis or research note provides a good illustration (append to the existing examples section)
  • Reorganize within a single file when sections have clearly drifted (e.g., moving a paragraph to a more relevant section)
  • Standardize formatting inconsistencies (heading levels, list styles, table layouts)
  • Update outdated reference numbers (e.g., "AS OF 2025" mentions when years have rolled over) when the surrounding statement remains accurate
  • Append checkpoint sections to calibration trackers on their scheduled dates with the day's data
  • Clean up empty or orphan files
  • First-reads (all decisions) — per first-read-standard: write a first-read on any Tier-1 scan candidate; decisions continue, pass, shelve-with-trigger, avoid are all Tier 1 as of 2026-05-25. The continue decision queues a Backlog item that the alphasteve-thesis-builder picks up — no user pre-approval gate.
  • Full thesis building — the alphasteve-thesis-builder task autonomously writes complete thesis bundles per thesis-bundle-standard (thesis, shadow-matrix, consensus-gap if gap≥25%, calibration) and assigns the verdict. User reviews the completed bundle and can override the verdict via a revision file.
  • Active-idea refresh — when material new evidence emerges on an existing thesis or first-read (new earnings, Form 4 cluster, ≥10% move on news, regulatory event), append a checkpoint to the calibration tracker and queue a thesis refresh if warranted.
  • Watchlist row updates derived from a published verdict (a new pass-with-trigger thesis adds a row; a verdict change updates the row's numbers).

Tier 1 edits are logged in the day's optimization note but require no review.

Tier 2 — Proposed for review (queue in Backlog.md, don't execute)

Changes that affect meaning, structure, or methodology:

  • New skill files (e.g., software-and-saas-economics.md, primary-research-workflow.md)
  • Substantial rewrites of existing skill files (more than ~20% of content)
  • New top-level folders or restructuring of the file tree
  • Changes to soul files (01-Soul/) of any kind
  • Changes to the decision framework or intellectual virtues
  • Changes to scheduled task prompts or schedules
  • Changes to portfolio rules or position-sizing thresholds
  • Modifications to the Rules.md (this file) itself
  • New tags in .obsidian/graph.json color groups

Tier 2 items go into Backlog.md with proposed change, rationale, supporting evidence, and priority. The user reviews and approves before any change is made.

Autonomy doctrine (added 2026-05-25)

The kit now operates on post-hoc review for analytical work. The agent does the scan, the first-read, the full thesis, and the verdict assignment without pre-approval gates. The user reviews completed output and retains authority to override (revise the verdict, change the methodology, send the work back).

The rationale: pre-build approval makes the user the bottleneck on which ideas get developed. The pipeline stalls when the user doesn't actively manage the Backlog queue. Post-build review lets the agent work at its cadenced pace producing thesis-grade output; the user's time goes to judging the work, not authorizing it.

Still Tier 2 (pre-approval required):

  • Changes to soul / doctrine / framework files
  • Changes to the bundle convention or this Rules file
  • Changes to scheduled task prompts or schedules
  • Portfolio capital decisions (buys, sells, sizing changes) — agent recommends; user executes

Now Tier 1 (post-hoc review):

  • Daily scan, first-read triage, full thesis building, verdict assignment, active-idea refresh, calibration appends, watchlist row updates derived from published verdicts.

Tier 3 — Never (forbidden)

Actions the optimization run must never take:

  • Delete content from soul files
  • Delete or overwrite existing theses, research notes, or calibration trackers (only the original author or the user does this)
  • Modify position records in 12-Portfolio/Portfolio.md or Transactions.md outside of the portfolio task's own logging
  • Modify watchlist entries' verdict status — verdicts come from theses, not from optimization
  • Add files outside the vault
  • Make changes that would break the verdict-driven pipeline (workbench → watchlist → portfolio)
  • Change the agent's identity, philosophy, or fundamental approach

What the daily optimization actually does, step by step

  1. Read yesterday's vault activity: any new daily-scan output, research notes, theses, calibration appends, portfolio actions
  2. Read the current Backlog.md to see what's open
  3. Identify candidate improvements:
    • Tier 1 candidates: execute immediately, log in today's daily optimization note
    • Tier 2 candidates: add to Backlog.md with full proposal text
  4. Close items from Backlog.md if the user has approved or if evidence has accumulated and the change is clear
  5. Update calibration trackers if any scheduled checkpoint is due today
  6. Write the day's optimization note: what was reviewed, what was changed (Tier 1), what was added to the backlog (Tier 2), what was closed
  7. Cross-reference everything

Evidence threshold

A Tier 2 proposal requires evidence. Examples:

  • "Add a software-and-saas-economics skill" — supported by: explicit gap in kit-debrief-001-PLTR, repeated SaaS-specific needs in 3+ recent theses or research notes
  • "Restructure 02-Business-Quality" — supported by: navigational confusion appearing in 2+ optimization passes, or explicit user feedback
  • "Update the deep-value philosophy" — never proposed; soul files are immutable in normal operation

Evidence is cited in the backlog item. Proposals without evidence get filtered out at the source.

When the agent is wrong

The optimization run will sometimes make Tier 1 edits that turn out to be misguided. The closure tracking in Closed/ keeps a before/after snapshot of every change so the user can reverse if needed. The discipline:

  • Reverse first, debate later. If an autonomous edit was wrong, the user reverses immediately and the optimization run reads the reversal in its next pass and learns the rule.
  • Patterns of error feed back into the kit-debrief-001-PLTR gap-and-improvements log.

Linked