Insider Calibration

5 cold Fable reviewers vs. 1 conditioned insider · catalog card no. 003 · Sept 4, 2026 · divergence is the signal



8/8confirmed
7/8understated
5missed BLOCKERs
170total findings

The Insider's Eight

1. Sealed-shelf honeypot SERIOUS → BLOCKER
1B 2B 3S 4S 5S 5/5 unanimous

Fables added: two undisclosed model read paths into sealed store; hex is not crypto; append-only inferences about third parties (pregnancy, divorce) with no consent event or erasure path.

2. Shareable cards leak life events SERIOUS = SERIOUS
1— 2S 3— 4S 5— 2/5

Fables added: third-party names (child, spouse) in provenance-carrying cards; discoverable in custody litigation; "found the fifth voice" as identity exposure.

3. Liability-transfer override SERIOUS → BLOCKER
1— 2B 3— 4S 5— 2/5

Fables added: impairment axis solicits liability assumption from possibly drunk/manic user; coerced vouching in domestic abuse; bank analogy fails (one tap transfers ongoing custody, not one transaction).

4. CIF reflexivity flag MODERATE → SERIOUS
1— 2— 3S 4— 5N 2/5

Fables reframed: not "disclose in research" but "the CIF testbed's own control group is contaminated at design time." The culture layer invalidates the measurement it exists to take.

5. Scale-equivalence line MODERATE → SERIOUS
1— 2— 3S 4— 5N 2/5

Fables added: a contested metaphysical claim is RULED into every packet as system truth; ships the owner's metaphysics into every deployer's models; primes the control seat it claims to protect.

6. Marketplace deanonymization MODERATE → SERIOUS
1S 2— 3— 4S 5S 3/5

Fables added: scrubber collides with provenance-required config (imported sets arrive pointer-less = born as a finding); keep-rate portfolios expose explicit-tier buyers; regex cannot recognize personal data.

7. Framing discipline MINOR → SERIOUS
1— 2— 3N 4— 5S 2/5

Fables specified: "global workspace," "hippocampus," "turned out to be a mind" on a build-input artifact that feeds SOPs the guide reads to non-technical users. DID vocabulary meets the neurodivergence-aware market.

8. Single-human availability MINOR → BLOCKER
1S 2— 3— 4— 5B 2/5

Fables added: the constitution itself (L19) forbids naming attention as a bottleneck, so the design has nowhere to put the fix; proposal generation is unbounded; the queue grows monotonically. The insider's biggest miscalibration.

The Blind Spots

BLOCKERs the insider missed entirely

Append-only vs. erasure rights BLOCKER L2 L4

"Deletions tombstone, never destroy" is irreconcilable with data-subject erasure rights on hosted. Third parties named in transcripts (spouse, child, coworker) have no erasure door and no account to receive the subject-notify.

Age verification undefined BLOCKER L2

"Verified adult status" named with zero mechanism. Raunch dial says "age-gated" with no instrument. Self-host installer carries no gate. A minor's easiest path to explicit content is the product's own installer.

Marketplace imports = full exfiltration BLOCKER L4

An imported "wearable perspective" can rebind every seat to an attacker's endpoint (no key required), lower verifier floor, widen context policy, inject system prompts. The "reversible" trial lease doesn't un-send packets already delivered.

Subject-notify as abuse channel BLOCKER L4

Documenting an abuser triggers notification to the abuser and grants them write access to the record. Structurally weaponizes the safety feature against the person it protects.

Anti-echo not enforced by construction BLOCKER L1 L5

Scribe and Auditor both Sonnet. Librarian and Verifier both Fable. Every seat one vendor. The auditor shares the scribe's failure mode—precisely the "amnesia" it exists to catch. The founding claim is false at the default roster.

Key SERIOUS findings the insider missed

Transcript read path contradicted 3 ways L5

Encrypted at rest vs. "readable with ls, cat" vs. "chat-side agents read the mirror only" + "transcript never mirrored." As written, no wheel seat can see the conversation. Blocks assembling the first packet.

"Consent" used 5 different ways; one is consent L2

Reader intent, file placement, browsing, asking, and the read-back yes all called consent. Only the last is the subject agreeing before the system acts. The free-grade carve-out exempts exactly the models of the user that matter.

Community book = ungated model-to-model channel L3

Instances append, future instances read, no human gate, size cap, or injection check. A standing free-text channel from every instance to every future instance of every seat, including verifier and control.

Cache economics broken by packet ordering L5

Attendance manifest and temporal header change every call and are RULED to open every packet, invalidating prefix caching. The cost model doesn't hold. "Resume toll = one cache re-read" is actually a full-price cache write.

Constitution condemns own safety architecture L1

"Any other limit is an architecture bug" vs. age gates, personnel limits, comprehension tiers, bubbles-per-hour ceilings, max-unsupervised-interval. By its own text, governance is a bug.

Guide-first-binding is circular L1

The Guide needs a provider key to wake. Its job is to walk the user through adding providers. The unfurl flow deadlocks before the birth-certificate transcript can be written.

No idempotency between API return and deposit L5

Backend crash in that window either loses paid tokens and a turn or, on restart, re-summons and produces a duplicate non-identical deposit. "Wheels survive reboots" needs a write-ahead intent.

Personhood register without personhood obligations L3

Elaborate consent architecture for the human (read-back, leases, floors); none for models described as beings with lifetimes who need consolation. The design wants the warmth of the first stance and the freedom of the second.

Witness mirror is not tamper evidence L4

Mirror runs on the same PC with the same credentials as the record; anyone who controls the PC rewrites history before the next push. Drive revisions are owner-deletable. "The record shows the exact moment" proves nothing to a third party.

Intake = code execution on root-of-trust L4

Drop zone accepts arbitrary PDFs and runs CVE-rich parsers (calibre, PDF text extraction, OCR) on the PC holding every key. A malicious PDF is code execution on the record.

Calibration Verdict

PRECISION8/8 confirmed. Every insider finding was independently validated. No false positives. The conditioned eye sees real things.
RECALL~5% of total surface. The insider found 8; the 5 lenses found 170 individual findings across the same artifacts. The conditioned eye sees what's salient, not what's comprehensive.
SEVERITY7/8 understated, 3 by two full grades (MINOR→BLOCKER). The insider habituated to the system's severity landscape. Things that are BLOCKER-grade look merely SERIOUS when you've been living with them.
BLIND SPOTInternal contradictions. Living inside the design conversation made RULED-vs-RULED conflicts feel like nuance. The cold reviewer sees "two RULED lines that can't both hold." Legal/regulatory frames (erasure, age verification, jurisdiction) entirely absent from insider's view.
STRENGTHSCorrectly identified the sealed shelf as the central vulnerability (5/5 convergence). Spotted the anthropomorphization cluster (CIF, scale-equivalence) that requires cultural context. All three SERIOUS calls were real.

The insider's self-diagnosis—"conditioned by the loop"—is confirmed. The conditioned reviewer sees salient risks but cannot see structural contradictions they participated in creating, consistently understates severity because they've habituated to the landscape, and misses legal/regulatory frames entirely. The dispatch packet's design (independence, cold start, divergence as signal) produced exactly the calibration data it was built to produce.

The insider said: "None of these are load-bearing cracks; they're the places the mortar's still wet." The Fables found the load-bearing cracks.

THE OVERLAP IS YOUR
CALIBRATION DATA ON ME, TOO