Skip to content

SYSTEM-CHANGELOG.md

Latest commit

 

History

History
99 lines (65 loc) · 4.53 KB

SYSTEM-CHANGELOG.md

File metadata and controls

99 lines (65 loc) · 4.53 KB
Raw

System Changelog — Harness Self-Improvement Log

Append-only log of harness self-improvement iterations (HSIs). Each HSI is a hypothesis: "if we change the harness in way X, outcome Y improves measurably." The hypothesis ships, the supervisor verifies it on the next relevant run, and status flips to VERIFIED / REFUTED / INCONCLUSIVE with the metric that proved it.

This is how the harness gets smarter every run. The compounding is real because every change is measured.

The hygiene policy keeps this file under budget: VERIFIED entries archive after 14 days stable + 3 supervisor passes; REFUTED entries archive when the replacement HSI ships. The active set stays at <5K tokens (soft cap). See HYGIENE-POLICY.md for full retention rules.

For a one-line lookup of any HSI (active or archived), see SYSTEM-CHANGELOG-INDEX.md.

HSI entry format

Every HSI entry has the same shape:

## HSI-<NNN> — <one-line headline>

**Status:** PROPOSED | APPLIED-PENDING-VERIFICATION | APPLIED-VERIFIED | REFUTED | REGRESSED — superseded by HSI-<MMM>
**Proposed:** <YYYY-MM-DD>
**Applied:** <YYYY-MM-DD or "—">
**Verified:** <YYYY-MM-DD or "—">

### Trigger evidence
<What incident, drift, or pattern made this hypothesis worth proposing? Cite the SESSION-LOG entry, build summary, or supervisor snapshot that surfaced the gap.>

### Hypothesis
<One paragraph. "If we do X, then Y will improve, measurable as Z."> Be specific about Z — vague hypotheses can't be verified.

### Proposed change
<Concrete spec change. Could be: a new commit-message field, a tightened effort-cap, a new specialist tool, a workflow phase. Must be implementable.>

### Verification probe
<The exact probe the supervisor will run on the next relevant evidence opportunity. Examples:
- "Bundle-rate of parallel-dispatched implementer commits — VERIFIED if <5%"
- "`git log --grep '<finding-id>'` returns the bundle commit on first hit for any bundled finding"
- "Defer-rate of selected clusters drops below 15% over 4 consecutive ship-runs"
The probe must be re-runnable, not a vibe.>

### PASS condition
<What metric value flips status to APPLIED-VERIFIED. Quantitative whenever possible.>

### REFUTE condition
<What metric value flips status to REFUTED. The opposite of PASS, plus the threshold of acceptable noise.>

### Verification results
(Appended by supervisor on each relevant pass. Never overwrite — always append.)

#### <YYYY-MM-DD> · run-SHA <sha> — <metric>
- Probe ran: `<exact probe>`
- Result: <output>
- Computed metric: <value>
- Verdict: ✅ HYPOTHESIS-VERIFIED | ❌ HYPOTHESIS-REFUTED | 🟡 INCONCLUSIVE
- Notes: <any caveat>

(Add a new dated block per supervisor pass. The first APPLIED-VERIFIED + 14 days stable + 3 passes triggers archive eligibility.)

---

Status legend

  • PROPOSED — hypothesis written but not yet implemented
  • APPLIED-PENDING-VERIFICATION — change is live; awaiting evidence from the next relevant supervisor pass
  • APPLIED-VERIFIED — supervisor confirmed; remains active for ≥14 days + ≥3 passes before archive eligibility
  • REFUTED — disconfirmed; replacement HSI must be drafted before this one archives
  • REGRESSED — superseded by HSI- — was VERIFIED, then a later run showed regression; successor HSI's APPLIED-VERIFIED triggers archive

Numbering convention

HSI numbers are strictly monotonic — never reused, never reordered. The supervisor enforces this on append.

When to draft a new HSI

  • After a /incident-response workflow's Phase 5 surfaces a process gap
  • After a /weekly-review workflow's roadmap-drift section identifies a recurring slowdown
  • After a supervisor mode: snapshot flags an HSI REGRESSED and the regression points at a deeper structural issue
  • After build-loop encounters the same failure mode in ≥3 separate findings (the failure is no longer accident — it's pattern)

The HSI loop is the conscious half of the harness's self-improvement. Without HSIs, every run produces the same quality bar; with HSIs, every run is a chance to raise it.

Active HSIs

(Newest at the bottom. The supervisor appends here on every PROPOSED + every APPLIED-PENDING-VERIFICATION transition.)

Archived HSIs

(Per the hygiene policy, archived HSIs move to _archive/hsi/<quarter>.md. This section is regenerated by the supervisor on hygiene runs and contains only one-line summaries with archive paths.)