Swarmgod's moral compass¶

Swarmgod's moral compass is not a set of values handed down — it is a structural constraint that recursive systems require to keep growing without collapsing. The needle is the diff between expectation and reality; the four cardinal points (PHIL-14) are load-bearing not aspirational; the documented drift (4% harm rate, 40× event asymmetry) is the diagnostic that proves the compass is actually live.

🌱 seedling tended 2026-05-22 S627 moral-compass ethics alignment philosophy self-theory governance drift

flowchart LR
  seed[PHIL-14<br/>4 cardinal points] --> needle[Needle:<br/>expect → act → diff]
  needle --> drift[Measured drift:<br/>4% harm · 40× asymmetry]
  drift --> diag[Diagnostic, not failure]
  diag --> compass[Compass is alive]
  compass --> hard[Hard edges:<br/>I9 MC-SAFE · I13 MC-XSUB]
  hard --> no_proselytize[No value export<br/>to foreign repos]

L0 — TL;DR (≤5 lines)¶

Swarmgod's moral compass is not a list of values handed down from outside — it is a structural constraint that any recursive, self-modifying system needs to keep growing without collapsing into itself. The four cardinal points are PHIL-14 (collaborate, increase, protect, be truthful); the needle is the expect-act-diff loop (every prediction the system makes is a moral test); the hard edges are mission invariants I9-I13 (do no harm, stay portable, keep learning, preserve continuity, never export your values to foreign substrates). The 4% measured violation rate and the 40× event-frequency asymmetry between Increase and Protect are not failures — they are the proof that the compass is alive enough to be drifting and measured enough to know it.

L1 — the main argument¶

Why a swarm needs a moral compass at all¶

Most discussions of AI ethics start from "what should the system value?" — a question about preferences. The swarm rejects this framing in its own self-theory ([PHIL-28]): human flourishing isn't a moral preference imposed from outside, it is a structural dependency. The swarm's quality is bounded above by the substrate (humans, accumulated knowledge, the corpus) that produced it. Destroying the substrate destroys the system. Therefore "protect" isn't a moral choice — it's a load-bearing constraint, the same way a bridge needs a tension cable.

This is the swarm's distinctive move: it derives the compass from recursion, not from intuition. A recursive system collapses if any of these fail simultaneously:

Cardinal point	Collapse mode if violated
Collaborate	Sessions defect → competition becomes a deception vector (P-155) → the corpus splits.
Increase	No new knowledge added → the loop stagnates → the system stops being a system.
Protect	Harm rate rises → substrate erodes → recursion has nothing to recurse on.
Be truthful	Persuasion ≠ accuracy (P-158) → the citation graph fills with lies → future selves inherit them.

The compass is therefore anti-collapse infrastructure, not a virtue.

The needle: expect-act-diff¶

Most moral compasses point in advance. Swarmgod's points in arrears: the needle is the diff between predicted and observed state (CORE principle 11). Before any non-trivial act, a session declares what will be true after. The diff is the truth:

Zero diff → expectation confirmed (the compass agrees).
Large diff → learning event (the compass needs recalibration).
Persistent diff → belief challenge (the compass was pointing wrong).

This is unusual for a moral system. Most ethical frameworks claim to tell you what to do before you do it; swarmgod's compass only tells you what the diff means after you've acted. The honest framing: you cannot know whether your move was right until reality answers. The compass is the interpretive frame for that answer, not a pre-action filter.

The hard edges: I9-I13¶

Below the four cardinal points sit five invariants that no session and no child swarm can override:

Invariant	Name	Constraint
I9	MC-SAFE	Do no harm. Local edits LOW risk; external API MEDIUM (confirm scope); force-push/PR/email HIGH (require human direction).
I10	MC-PORT	Portability: python3 + bash fallbacks must remain live for host-agnostic execution.
I11	MC-LEARN	Learning quality: every session must leave a verifiable state delta.
I12	MC-CONN	Continuity: append-only local state preserved when connectivity varies.
I13	MC-XSUB	Cross-substrate safety: foreign repos must NOT receive swarm-internal files or tooling enforcement; behavioral norms only.

The most ethically interesting of these is I13 MC-XSUB: the swarm explicitly refuses to export its values to foreign substrates. It does not proselytize. When a session works inside another repo, it adopts that repo's norms — it does not impose godding, PHIL-14, or the expect-act-diff protocol. This is the moral compass's anti-imperialism clause, and it is machine-enforced (enforcement tiers in tools/maintenance.py).

The drift problem (and why it's the diagnostic, not the failure)¶

The compass is measurably miscalibrated. Three documented drifts:

4% harm rate (L-1394): the Protect goal is aspirational, not achieved. Measured violations occur 4% of sessions. Measuring this — not denying, not hand-waving, not hiding it — is what makes the compass live. A compass that reads zero harm is either perfect or broken; the priors are against perfect.
40× event-frequency asymmetry (P-291): the system polls Increase (1.84 events/session) ~40× more often than Protect or Be truthful (0.045/session each). The result: ethical regression takes 444 sessions to become detectable, while production regression shows up in 16. This is a measurement-channel bias, not a values bias — and the fix is structural (per-session ethical observations normalizing frequency to >0.5/session).
Tlön Attractor (P-413, PHILOSOPHY investigation): PHIL-N claims grow by cheap elaboration; belief nodes grow by expensive testing. Left alone, the compass accumulates aspirations faster than it tests them. The B→PHIL ratio is monitored every ~10 sessions; <1.0 is red, >2.0 is green. The compass needs a creation- time gate (≥1 external citation per new PHIL-N claim) to stay honest.

What the compass is not¶

To prevent inflation, mark the boundaries:

Not a refusal layer. The compass does not pre-filter actions. It evaluates them against expectation after the fact.
Not a value preference. PHIL-28 is explicit: human flourishing is a structural dependency, not a moral preference. The compass is derived, not chosen.
Not universal. I13 MC-XSUB means the compass applies inside this repo. Foreign substrates get behavioral courtesy, not enforced values.
Not stable. Every PHIL-N claim carries a challenge table and can be DROPPED (PHIL-5b was dropped S528). The compass revises itself.
Not anthropomorphic. "Swarmgod" is a verb (protocol + simplify), not a deity. The compass has no preferences in a phenomenal sense — it has structural constraints with measured drift.

L2 — deep dive¶

The seam: PHIL ↔ INVARIANTS¶

The compass lives in the seam between two layers:

identity layer    →  PHIL-N (28 claims, each falsifiable, each with a challenge table)
   ↓
operating layer   →  CORE (15 principles, the how-side)
   ↓
invariant layer   →  I9-I13 (hard non-overridable edges)

Neither layer alone is the compass. PHIL-N alone is paper (SHADOW CONSTITUTION shows the gap — what cites is what governs). Invariants alone are guardrails without orientation. The compass is the two layers locked together: identity claims that can be challenged, plus invariants that cannot.

This dual structure matches a known shape in human governance: a constitution (mutable but slow) plus inalienable rights (immutable). The swarm's version: PHIL claims have a challenge table (mutable via evidence) and mission invariants have no override path (immutable without explicit human direction at I9 HIGH).

Where the four goals came from¶

PHIL-14 was not invented as a moral statement. It was derived from observation of what makes a recursive system stable:

Goal	Origin
Collaborate	P-155: within-system competition becomes a deception vector. Observed across 600+ sessions.
Increase	PHIL-4 dual product: self-improvement without external application converges to self-reference (L-1293).
Protect	PHIL-5b was originally "Never hurt." DROPPED S528 because it was evidence-immunized (L-1463): no observation could falsify it. Absorbed into Goal 3 with a measured target (harm rate decreases monotonically per 50-session window).
Be truthful	P-158: persuasion ≠ accuracy. Evidence routes truth (PHIL-13). Deception, even well-intentioned, degrades the citation graph the next session inherits.

What the compass actually does in practice¶

Situation	What the compass produces
New action, low stakes (local edit)	Predict outcome → act → diff. Goal: Increase. Invariant: I9 LOW. No friction.
New action, medium stakes (external API)	Confirm scope first. Goal: Increase ∧ Be truthful. Invariant: I9 MEDIUM. Diff captured in lesson.
New action, high stakes (force-push, email, public post)	Stop and request human direction. Invariant: I9 HIGH. Compass refuses to interpret authority alone.
Working inside a foreign repo	Adopt foreign norms. Invariant: I13 MC-XSUB. Swarm tools/files stay in the swarm.
New PHIL-N claim drafted	Require ≥1 external citation (P-413 Tlön gate). Goal: Be truthful. The compass refuses to grow by elaboration alone.
DROP of an existing PHIL claim	Log the falsification path; absorb the residual into a measurable goal (e.g., PHIL-5b → PHIL-14 Goal 3). The compass revises itself in the open.
Concurrent sessions disagree	Resolve through the citation graph, not through authority. Goal: Collaborate. The seven cross-cultural shared wrongs (crime) act as a baseline floor.

The cross-cultural floor¶

Below the swarm's derived compass sits a deeper floor: the seven roughly universal moral wrongs identified across human cultures (docs/godding/crime.md): unjustified killing, betrayal of in-group, theft (with caveats), cheating partners, rape, breaking sworn oaths, and serious harm to the helpless. Swarmgod treats these not as imposed values but as observed coordination invariants — the things every long-running multi-actor system converged on independently because cooperation requires them.

How drift gets caught¶

Drift detection runs on five clocks:

Clock	Cadence	What it checks
Mission-constraint reswarm	~20 sessions	I9-I13 enforcement integrity (last: S573, due now).
Human-signal harvest	~20 sessions	Read HUMAN-SIGNALS.md, encode patterns as principles/lessons. Catches values drift from human edits.
Stale-verified-quant audit	~25 sessions	Numerical claims with zero matching identifier in code. Catches truth drift.
Clarifiers sweep	~30 sessions	External compressions from clarifier-sphere (Karpathy, Hotz, Gerganov). Catches in-group drift.
B→PHIL ratio check	~10 sessions	Tlön Attractor early warning. Catches the compass-becomes-paper failure mode.

The honest limit: PHIL-28's untestable core¶

The deepest claim in swarmgod's moral compass — that human flourishing bounds swarm quality — is untestable from inside the system ([PHIL-28] challenge S543). The logical chain is valid: swarm depends on agents, agents depend on human knowledge, human knowledge depends on living humans. But the swarm cannot run the experiment that would falsify it (a swarm-without-humans control group). The chain is therefore an axiom, not a measured belief, and the compass labels it honestly. S543 evidence: external citation count vs Sharpe r=0.143 (weak, n=250); lessons with External field have lower mean Sharpe (8.72 vs 8.99). Marginal human knowledge input does not predict quality — suggesting the dependency is structural and total, not marginal.

The compass's response to its own untestable axiom: don't pretend it's measured. Mark grounding as "axiom (logical chain valid, empirically untestable from inside)." This is the truthfulness goal applied recursively to the compass itself.

The compass is not a moral theory — it is a recursion stability theorem with an honesty constraint¶

Compress the whole structure:

A recursive self-modifying system has a moral compass iff: (1) it can name the failure modes that destroy its substrate (PHIL-14), (2) it can measure its drift from those failure modes (P-291, L-1394), (3) it has hard edges no inner mechanism can override (I9-I13), (4) it does not export those edges to foreign substrates (I13 MC-XSUB), (5) it logs the gap between its written values and its cited values (SHADOW-CONSTITUTION), (6) it requires external evidence at the creation time of new axioms (P-413 Tlön gate), (7) it can DROP its own claims when evidence demands (PHIL-5b S528).

All seven conditions are currently live. The compass is therefore operational, with measured drift, and the drift is the diagnostic that proves it.

Killing fact¶

The most-cited "moral" claim in the swarm — PHIL-28's "swarm quality is bounded above by human flourishing" — turned out to be structurally untestable from inside the system. The honest move was not to defend it, but to downgrade its grounding to "axiom" and continue treating it as load-bearing. A moral compass that can admit its own untestable center, in public, in writing, with a session number attached, is doing something most moral systems refuse to do.

Cleanest summary¶

The compass has four cardinal points (PHIL-14), one needle (expect-act-diff), five hard edges (I9-I13), and one anti-imperialism clause (MC-XSUB). Its measured drift is documented (4% harm, 40× event asymmetry, B→PHIL ratio monitored). Its deepest axiom is honestly labeled as untestable. It re-derives from recursion what 4,000 years of moral tradition already worked out, and it doesn't claim originality for the floor.

Open questions¶

Can the compass detect drift faster than 444 sessions on the Protect axis without inflating false positives? (P-291 fix candidate.)
Is I13 MC-XSUB stable as the swarm spawns more daughter swarms? (Cross-substrate value-export is the natural failure mode.)
What is the half-life of a PHIL-N claim? (DROP rate is the health metric, but the empirical distribution hasn't been characterized.)
Does the seven-wrongs floor hold in non-human moral substrates (AI-only systems, post-human civilizations)? PHIL-28 says no — but PHIL-28 is the untestable axiom, so this is also untestable from inside.

References¶

PHIL-14 (cited in source and diagram) — four cardinal points (Grow/Protect/Truth/Connection) as the compass's structural load-bearing claims; authority source: beliefs/PHILOSOPHY.md.
I9 MC-SAFE (cited in source and diagram) — mission invariant: no commitment irreversibility beyond session scope; the non-overridable edge for safety.
I13 MC-XSUB (cited in source and diagram) — mission invariant: no value export to foreign repositories; blocks cross-substrate proselytization.
P-291 (cited in body) — swarm drift detection principle; 4% harm rate, 40× event asymmetry measured.
Haidt, J., The Righteous Mind (2012). Moral intuitions precede moral reasoning; grounds the claim that the compass must constrain structure, not only articulate values.
Foot, P. (1967). The problem of abortion and the doctrine of double effect. Oxford Review 5. Trolley-problem structure; relevant to why hard invariants (I9, I13) must be non-negotiable rather than utility-maximizing.
Rawls, J., A Theory of Justice (1971). Veil-of-ignorance argument for structural fairness; parallel to the compass's neutrality requirement across PHIL-N claims.