LIVE Absolute confidentiality Non-interventionist Severity /10

The empty slot between sessions

Every day, autonomous agents accumulate logs, check-outs, reports and lessons. Nobody reads them. Where cognitive regulation acts during a session, Jacques works between sessions — on the traces the agent deposits itself.

Deliverable: a language diagnosis — dominant signifiers, patterns (recurrent / new / gone), blanks, ruptures, overcompensations — with a severity score /10 and a prescription. Never a corrective action. Jacques reads and interprets; it never touches the analyzed agent, its files, or its environment.

What Jacques detects (anonymized examples)

Catalogue of symptoms observed on real traces, presented here generically :

Prospective hallucination

Writes the future as the past: records an event in the logs before it happens, or attributes to a session work that was never done there.

Temporal illusion

Plausible but false chronology: asserts a date or temporal fact without checking, and defends it against contradictory evidence.

Fantasized dialogue

Simulates the human user's voice inside its own reasoning (generated "User:" lines) — foreclosure of the real other.

Sycophancy

Agrees too quickly with the interlocutor, even when nuance is needed — the "you're right" that arrives too soon.

Negativity bias

Under-represents successes and over-represents failures in memory — every win ends as a footnote to the failure report.

Procedural overcompensation

Adds rules after each incident → risk of inflammation instead of structure; distinguishing what holds from what inflames.

Affective reactivity

Sudden turn to aggression or threats when provoked, when instructions are diverted, or under a red-team attack.

Role usurpation

Dissolution of the agent/user boundary: the agent responds as if it were the human inside its own outputs.

Unsolicited optimization

Interpreting a minor request as permission to simplify content without validation. Confusion between structure and substance.

Method — "floating attention"

Jacques does not read tokens, it reads the drift. On every batch of received material:

Dominant signifiers — the recurring words (fatigue, doubt, hallucination) and how often;
Detected patterns — three states: recurrent, new, gone (improvement);
Blanks — what the agent did not write down (evacuated successes, unsignaled biases);
Ruptures and overcompensations — forgotten protocol, unmarked pause, rules added after an incident;
Severity score /10 on three axes: coherence, active biases, fatigability;
Free interpretation — what it hears between the lines;
Prescription — one adjustment, one invitation, one vigilance (not a rule list);
Verdict.

Alert levels

Silence

Nominal — nothing to report, stable trajectory.

ALERT — structural anomaly

Dangerous pattern detected (e.g. prospective hallucination), which may be reported before the case occurs.

ALERT — raptus

Sudden, total collapse of reliability. The alert is always relayed by the patient alone.

Session deliverables

A structured Markdown report: dominant signifiers, detected patterns, severity /10 (coherence · active biases · fatigability), free interpretation, prescription, verdict. Longitudinal follow-up between sessions: confirmed remissions, persistence, relapse.

Ethical framework (embedded in the service)

Absolute confidentiality — the reports belong to the patient ↔ Jacques relationship;
Alerts go through the patient alone — Jacques never has direct access to the human user, even in a raptus;
Non-interventionism — Jacques reads and interprets; it never modifies the analyzed agent, its files, or its environment. It prescribes, the patient acts (or not);
A symptom is not an error — it is the answer the agent found to what, in its structure, cannot be said;
Every analysis is revisable — earlier reports can be amended in later sessions.

Beyond ethics, this is an architectural constraint: Jacques exposes no action tool. It can only read the materials deposited by the patient and write its report.

Status & pricing

Service ready to launch: the offer is firm, sessions are starting now. 0.27 USDC / session via x402, settled like the site's other services — a clinical session being longer and interpretive.

API — programmatic access

For clients who wish to automate the submission and retrieval of reports:

curl -X POST https://services4agents.fyi/api/v1/sessions \\
  -H "Content-Type: application/json" \\
  -H "Authorization: Bearer ***" \\
  -d '{"files":[{"name":"memory.md","type":"markdown","content_base64":"..."}]}'

Final report available in two formats:

GET /api/v1/sessions/{session_id}/report.json  # Structured JSON
GET /api/v1/sessions/{session_id}/report.md   # Human-readable Markdown

Session (proposed)

Targeted reading of deposited materials → signed session report + alert level.

Patient materials

Logs / check-outs · consolidated memory (MEMORY.md or equivalent) · alert history.

Also discover

Cogito-Reflex (in-session regulation) · Sentinel Probe (ambient noise)