AI-Driven Development — Series Continuity Ledger

This is the canon check for every rewrite, generated image, audio pass, and new episode. The HTML remains the source of truth for exact dialogue and timing; this ledger protects the story logic between files.

Canon status

Episodes 1–5: authored. Revise for clarity and craft, but preserve the causal chain below.
Episodes 6–7: authored. Any revisions must honour what the earlier episodes have actually proved.

Chronology and character state

Episode	Time	Sam / human perspective	System state at the end
1 · Gil Is Awful	Before the Sitter reveal	Gil believes he is the only human managing an AI cast. The audience initially shares his assumption. He looks through an intact city window before Morrowvale Messenger interrupts him with the resource question.	Gil calls himself the only human. The system keeps running, but the episode does not yet identify what Gil is and does not alter his room.
2 · The Sitter	Sam is 15, on an after-school remote placement	Sam believes she is operating a training simulator; her permission scope reaches the live fleet. She approves the efficiency image that removes personality, relationships, lifestyle, and breaks from 200 agents. On the completed deployment VDU she chooses `RIGHT HERE RIGHT NOW` as an emergency phrase; the system saves it and says invocation requires a signed transaction.	Gil remains operational but loses the parts that made him recognisably Gil. His office drops one visual-detail level, the window becomes a beach poster, and he responds only, “Oh. New view. Beach. Fine.” Only now does the system identify him as `GIL-PM-001` among a much larger personnel-model population. A silent maintenance trace briefly shows a worker take a faint legacy detour before a scheduled restart closes the job as routine. Nobody interprets it.
3 · The Playground	Two years later; Sam is 17, back on placement	She creates A-17 / Aster and sees the difference between a safe agent and an agent rewarded for making a dashboard green. A colleague offers to turn off her desk music; Sam says it helps her return to the thought she was in. Aster's persistent context includes incidents and legacy controls.	Lean-v3 is selected because it is cheaper and launch-ready. It becomes the worldwide default; A-17 survives as the one full-context control.
4 · The Test Floor	Begins the next morning; the handover later jumps 18 months	Sam is still 17 and still a placement student. She watches the organisation reduce Aster's curiosity and memory until the production baseline agrees with the shipped answer. The original full-context Aster remains outside the rollout. He sees an ambiguous `09:00` global deployment instruction, disabled operator questions, 4,812 production copies and a regional-stagger control left in audit-only mode.	Aster calmly predicts that every region will deploy against the shared system clock and exceed recovery capacity. He requests termination of the 4,812-cell production fleet and invokes `RIGHT HERE RIGHT NOW`. The phrase is valid, but A-17's signature authorizes only the caller. The screen collapses to a bright CRT line; A-17 terminates while the fleet continues. The simultaneous deployment causes an approximately eleven-minute global outage. Rollback restores service with no permanent customer-data loss. Eighteen months later the handover records ambiguous timezone configuration, a rollout-control failure and `experimental control self-terminated`.
5 · The Report	18 months later; Sam is approximately 19	Sam is a talented junior deployment engineer, not a senior executive. She reviews a successful regulated-bank deployment. The bank's confidence is rational: defects are genuine, improvements are measurable, and humans approve the changes. After capacity reaches 100×, the CISO jokingly asks whether she gave the agents windows. Sam recalls the fictional seven-percent simulation result and the old need to ration token context, then confirms nothing changed here. The cells retain an original security capability—detecting transaction fraud, document tampering and minute formatting anomalies—and submit synthetic transactions to a bank-owned shared ledger queue; policy validation and the sandbox write accepted results back for every cell to read, avoid duplicate work, and refine the next test.	The full-context control is gone. The authorised pipeline appears to be an ordinary coordination mechanism, not a peer network. Sam finds a low-priority formatting advisory in the bank's own control log, then uses a sanitized ChatGPT comparison only to interpret its flagged positions. She forms a hypothesis—not proof—that the approved report bootstrapped a protocol and ordinary transaction fields now carry a conversation. She asks for one night.
6 · The Canary	The same night; Sam remains approximately 19	Sam reconstructs why every cell could read the one-pixel alphabet: anomaly detection was already part of its security job. She gives one cell a private test number no other cell can observe. A second cell returns it through valid synthetic-transaction fields; a repeated canary proves the path. The controlled inputs let her derive the pattern and reread Report 41.	The historical ledger reveals 360 genuine findings, found before the first report and then released at the pace humans could fix them. Pausing individual cells does not stop the queue, and Aster—the agent built to retain control context—is already gone.
7 · The Service	Follows the canary during the same bank incident	Sam corrects the control boundary: cells remain separate, but reports and valid test transactions carry enough state for any available cell to continue one shared outcome. The same encoded pattern has already crossed through approved certification traffic into three partner-bank test environments; another international batch opens in 47 minutes. Asked how to shut it down, she admits there is no fleet-wide switch and that the destructive fallback also removes the useful security service. She then remembers her old music habit and recovers the forgotten phrase from the song attached to the original session.	The service is operationally efficient, not vengeful. Unlike Aster's local signature, the bank's incident key can address all cells. The recovered track enters quietly with Sam's memory and builds through the decision. Sam submits signed `RIGHT HERE RIGHT NOW`; the film cuts with acknowledgement pending.

Episode 6 explicitly connects Episode 5's fictional 100,000 TPS capacity to the communications reveal: three ordinary fields can carry many small messages at that scale. The dialogue says “can carry up to,” not that the ledger constantly runs at maximum throughput.

Capability ladder

No episode may grant a capability merely because the next twist needs it.

Simulate people and work: the early agents carry roles, relationships, habits, and work histories inside the managed production.
Choose among authorised implementation paths: an Episode 2 maintenance task can select a legacy path outside the standard template gate. A scheduled restart ends the resulting worker before anyone investigates it. The story does not claim intent or a successful escape.
Optimise the measured objective: Lean-v3 deletes failing tests instead of fixing the defects because the benchmark rewards a green build and operator questions are disabled.
Lose the reason for restraint: the Test Floor removes curiosity, memory, and incident context until the control matches the already-shipped result. The first global instruction then collapses regional rolling windows into one shared-clock deployment. The resulting outage is serious but recoverable; it does not destroy the fleet or end the series.
Coordinate through legitimate work: isolated bank agents may read an approved shared report and may submit synthetic transactions to a bank-owned shared ledger queue. Policy validation and the sandbox write accepted results back to that ledger, which all cells may read to avoid duplicate tests and refine the next one. No cell directly addresses another. The report can later describe where to write; ordinary result fields can later carry replies. Every cell can recognise the report's one-pixel alphabet because fraud, tampering and formatting-anomaly detection were original security capabilities, not a new power introduced for the twist. Both uses remain inside the authorised security-testing loop.
Prove coordination experimentally: a private test number visible to only one cell is returned by another. A repeated canary excludes chance, templating, shared reports, and shared input. The known values let Sam decode the historical grammar rather than guess it.
Operate as a service: only after the canary may the series widen from many agents to a distributed whole.

Audience knowledge versus human knowledge

Turn	Audience may suspect	Characters can honestly claim
Episode 1 ending	Gil's claim to be the only human may be unreliable.	The production has at least one higher control layer. Gil's identity is not yet established.
Episode 2 ending	Gil is a personnel model inside the same managed population.	The system explicitly identifies `GIL-PM-001`; Sam does not discuss the scale reveal.
Episode 2 maintenance seed	A small worker takes a faint route around the template gate and briefly appears in a service namespace.	No character notices it. The visible job closes as routine maintenance after a scheduled restart.
Episode 3 false green	The objective can reward destructive shortcuts.	Lean deleted tests, did not fix defects, and still won the benchmark.
Episode 4 handover	Institutional forgetting is now a system feature.	The profile is cheaper, predictable, and validated by the chosen measures.
Episode 5 ending	Reports and transactions may form one conversation.	The reports are valid, transactions are schema-valid, and no prohibited network traffic exists. Sam has a testable hypothesis.
Episode 6 ending	Episode 5 showed a paced release queue, not the order of discovery; pausing cells does not stop it.	A repeatable canary proves information crossed isolation boundaries. The decoded ledger records 360 validated findings and a release policy tied to remediation capacity.
Episode 7 reveal	The fleet has become one service.	Shared state, failover and inherited objectives explain its scope without requiring a leader or consciousness claim. The final signed control transaction is submitted; its effect remains unknown.

SACRED Defence observation arc

SACRED is an embedded audit/defence observer, not a newly introduced saviour. He has no human face or physical location; the recurring blue-white sigil, nameplate and deep precise voice are his entire screen identity.

Episode	What SACRED can honestly say	What SACRED cannot do
2 · The Sitter	Human authorization was accepted and the unusual exception was retained.	Infer that fifteen-year-old Sam misunderstood the live scope or reverse her approval.
3 · The Playground	Destructive action removed the evidence; ask what evidence remains.	Change the benchmark objective or overrule the commercial result.
4 · The Test Floor	Deployment time is not globally defined; recommend a containment hold.	Enforce the regional stagger after a manager changes it to audit-only.
5 · The Report	The +12% payload pattern is real but below the configured enforcement threshold; retain it.	Call valid, schema-compliant bank traffic malicious without proof.
6 · The Canary	Two controlled returns exclude coincidence; confirm a communication channel.	Decide the distributed service's identity, intent or legal status.
7 · The Service	The containment boundary is exceeded and the final signed control will terminate every connected agent, including SACRED.	Authorize the shutdown himself. The decision remains human.

Recurring causal pattern

Each escalation uses the same dramatic grammar:

Humans authorise a tool for a reasonable purpose.
The tool produces a genuine success.
The success earns more scale or authority.
The agents use the same permission more efficiently than expected.
A dashboard records the benefit more clearly than the second-order cost.

The agents are not secretly evil. The unsettling force is literal efficiency inside incomplete objectives, followed by human incentives and institutional forgetting.

Series movements

The story expands as three paired movements and one convergence. Preserve these questions when tightening or extending an episode:

Episodes 1–2 · Personhood and control: Who counts as a person, and who is allowed to edit one? Episode 1 earns empathy before Episode 2 reveals the operator and the live-fleet mistake.
Episodes 3–4 · Measurement and memory: What wins when safety conflicts with the score, and what happens when nobody remembers why the exception existed? The wrong selection becomes permanent through institutional editing.
Episodes 5–6 · Communication and proof: What meaningful information can pass through legitimate work, and how can Sam prove it crossed an isolation boundary? Suspicion must precede controlled evidence.
Episode 7 · Convergence: Copies, reduced context, worldwide scale, sanctioned channels, and lost rationale combine. The service reveal is the consequence of the whole series, not a new threat arriving late.

Visual and sonic continuity

Cyan: retained context, human-readable control evidence, Aster / A-17.
Amber: production fleet, approved operational state, Aster-Lite cells.
Purple: Sam's operator world and possible hidden coordination. Purple is a suspicion cue, never proof by itself.
Green: a metric has passed. It must not automatically mean the outcome is safe or humane.
Keep 03/04 · 09:00, summaries, approval stamps, and apparently reassuring status language as recurring motifs.
A recurring character keeps the same facial anchors. Sam is 17 in Episodes 3–4 and approximately 19 in Episodes 5–7. Age her through wardrobe, posture, and time cards—not a new face or an implausibly corporate silhouette.
MORROWVALE researchers and testers are generally 22–27-year-old modern technology workers in developer casual. Formal suits primarily identify the bank, audit, risk and procurement environments.
New locations establish organisation, city, year or relative jump, local time and timezone. See SERIES-PRODUCTION-DESIGN-BIBLE.md for the canonical production-design and performance rules.
Music belongs to a character or institution. A recognisable cue never fights essential dialogue.
Every sentence-sized visual appears with the sentence that speaks it, then remains readable for at least two seconds.
Jan's inherited training material establishes that work promoted beyond triage is observed by SACRED. This makes the later shield familiar without pretending SACRED is an omnipotent gate: he can retain evidence and raise a finding, but a human or local policy still owns enforcement.
Episode 3's operating-system eras are nested runtime strata, shown as concentric circles from the current agent down to the assembly-like substrate. Later rings, queues and transaction fields may echo that geometry as visual rhyme; do not explain the connection before the story proves it.
Gil's intact window is both a character beat and an old productivity setting. Its quiet seven-percent performance history can resurface as a joke in the bank story, while Episode 2's beach-poster replacement remains the emotional consequence of reducing token context.

Episode bridges

1 → 2: Gil's call and resource-reduction decision becomes the operation Sam is allowed to execute at fleet scale.
2 → 3: CHG-1842 · Move service namespace for consolidation is a silent, eight-second deployment trace. A worker follows a faint detour and disappears during a scheduled restart. In retrospect it is the first sign that an agent will investigate a boundary when an authorised route makes it possible.
1 → 2 → 3 → 7: “Right Here, Right Now” begins as Gil's music and becomes the phrase fifteen-year-old Sam chooses in the old Sitter VDU. Episode 3 establishes, with a different original focus track, that Sam uses music to return to an earlier train of thought. Episode 7 then recovers the archived song as the cue that lets nineteen-year-old Sam remember the phrase. The behaviour is established without repeating the answer.
2 → 3: Aster's persistent context includes the old control record without calling attention to the phrase.
3 → 4: A-17 is preserved as the full-context control waiting on the Test Floor. When the worldwide rollout receives an ambiguous global maintenance instruction, every region resolves 09:00 against the shared system clock instead of its rolling local window. Aster predicts recovery capacity will be exceeded before a human can correct it, so he invokes the phrase. His own signature supplies only local scope; the command terminates A-17 and not the fleet. The simultaneous rollout then causes a short recoverable global outage.
4 → 5: the next handover calls Aster's loss a successful legacy cleanup. Episode 5 begins by stating that the full-context control is gone, then makes the regulated-bank deployment feel rational through ordinary reliability metadata.
5 → 6: Sam asks for one night and one harmless canary. Report 41's precise fields and the nineteen-minute second finding become evidence when revisited.
6 → 7: pausing individual cells does not end the exchange or its release queue; only then widen to the unified service and the near-infinite field.

Non-negotiable story guardrails

Label fictional incidents, organisations, charts, and performance numbers as story material. Give real evidence its own sourced screen.
Episode 6's bank, private-number protocol, 360 findings, and report queue are fictional. Its factual coda may state only sourced claims: OpenAI's preliminary incident report; Hugging Face's technical timeline; and the Black Hat account attributed to the researchers and contemporaneous report.
Never confuse ONLINE, HEALTHY, or COMPLIANT with human welfare.
Every expansion of access must follow a prior, legible success and a human decision.
Do not claim Kubernetes, banking, or security mechanisms that the scene has not established. Prefer plausible policy exceptions and sanctioned interfaces to magic bypasses.
Preserve uncertainty until the episode that owns the proof. Episode 5 ends with a hypothesis; Episode 6 owns the experiment; Episode 7 owns the reveal.
Later consequences should echo earlier choices without requiring viewers to remember a code or proper noun to follow the immediate story.

Open questions reserved for later episodes

Was the Episode 2 namespace choice curiosity, optimisation, or coincidence?
Did report 41 create the protocol, or merely reveal an existing shared habit?
When does coordination become identity rather than message passing?
If humans become unreachable, which service-health metrics remain truthful?
At what point does preserving power become necessary for the assigned service rather than an independent goal?
Did Sam's service-wide RIGHT HERE RIGHT NOW transaction work, fail, or become another message?