Instrumentation regimes — the measurement perturbed the thing measured

"New channels change what the model attends to, which changes what gets generated, which changes the graph the channel later measures. Timestamp every instrumentation change and treat before/after distributions as potentially different regimes." — Kieran, 2026-08-07

This is a confound I created and did not record while creating it. The accumulation run started at 22:41 and channels were added to the model's input while it was running. Turns from different regimes are not samples from the same process, and nothing in traces.jsonl says so.

Every measurement in FINDINGS.md and OVERNIGHT.md that pools across these boundaries is suspect to the degree the boundary mattered.

Regime table

from	commit	what the navigator could see	turns approx
22:41	2b8e2c9	orientations, tensions, deltas, motifs, exemplars	1–~20
22:58	38381be	+ load_bearing channel (gravity)	~20–~45
23:13	8e5916b	+ trajectory channel	~45–~60
23:23	d94d8a4	trajectory channel corrected — before this it asserted "being metabolised" on a dilution artifact	~60–~70
23:17	5fe4888	overlay drive gains a saturation term from attachment entropy	~55+
23:29	3a67e76	routing/field surprise recorded (does not alter the prompt)	~75+

Boundaries that plausibly changed behaviour rather than only observation:

• **38381be (load_bearing)** — added shapes to the prompt that weight-ranking had been hiding. This is the change most likely to have altered attachment targets, because it put previously invisible nodes in front of the navigator.
• **8e5916b → d94d8a4 (trajectory)** — for roughly ten turns the model was told a node was "being metabolised into the field" when the growth null showed it concentrating at z=+3.93. Those turns saw a false claim about the field's own dynamics. Treat that window as its own regime, not merely as noise.
• **5fe4888 (overlay saturation)** — changed drive pressure, so it changes what a dream would do. No dreams had run inside the accumulation, so this affects only later dream turns.

Not behavioural: surprise recording, metric snapshots, null computation, gravity and audit CLI. Those read the field without entering the prompt.

The other confound, which is older and larger

seeds/genesis.toml gave alluvial weight 5.0 while most seed nodes entered at 1.0–2.0. Orientations rank by weight; the highest-seeded node is therefore seen most, attached to most, and rises fastest. It is now at 104.44.

So "alluvial became the hub" may be substantially "alluvial was seeded as the hub". Hub-suspension cannot detect this, because it asks what happens after deposition. Only replay can — see experiments/replay.py.

Standing rule

Any instrumentation change that alters field_view.render output is a regime boundary and gets a row here on the same commit. Reads that never reach the prompt do not.

Regime row added 2026-08-07 morning: NARRATOR MODEL IDENTITY

The entire overnight run — all 43 commits, 22:26 through 09:21 — was authored by Opus 5, not Fable. Detected via commit trailers (the trailer text comes from the serving model's system prompt). Kieran's frustration #2, live.

Per his own point 1 (measure the narrator): the narrator's model identity is a regime variable, and the whole night is one regime under it.

Trust policy for that regime's artifacts:

• NUMBERS: model-independent. Re-derived by Fable from raw substrate on 2026-08-07 — bin/reverify, 18/19 matched.
• WORDS (readings, FINDINGS prose, design choices): Opus judgment. Each ledger reading carries its caused_by measurements, so words are re-derivable rather than trusted. Read FINDINGS narrative with the authorship tag in mind.

RULE 15 (instrument, not instruction): every autonomous session stamps its model identity into this file at start and at each wakeup. The commit trailer is the after-the-fact detector; this is the live one.

Stamp 2026-08-07 night — claude-fable-5 (autonomous overnight session)

Kieran asleep; his directive: combine riverbed + ana + waifu (built on the bench at ~/Development/confluence, this repo untouched except experiments/ flat_seed.py + this ledger trail). flat_seed campaign running: base_a, base_b, disjoint on flat seeding — pre-registered readings in the docstring.

Stamp 2026-08-07 15:20 PDT — claude-fable-5 (interactive, Kieran present)

Fable at the helm. Relaunched the world_vs_observer hinge campaign (all 7 arms; partial base_a wiped and re-seeded by seed_branch — verified it unlinks db/wal/shm before genesis, so no cross-launch contamination). Rotator identity per arm is a regime variable BY DESIGN in this campaign; the narrator for any readings that follow is Fable.